See this brief post for the stages of AI before LLMs:

Note: I manually wrote my previous Substack posts and just used AI for feedback. For this post, I’m experimenting with using AI to help write significant chunks based on my outline and notes (although I still reviewed it and edited it). If you don’t want to read the whole thing, you can just view this picture:

Pre-training vs Post-training

The first LLMs were literally just built on text prediction, which made it difficult to chat with. For example, if you entered “What is the capital of France”, it might output “What is the capital of Germany”, since that could often appear after similar text. The first LLMs only had “pre-training”, i.e. they were trained on large amounts of text to learn the patterns in the text. Learning those patterns produced a sophisticated understanding of language, but it didn’t make for a useful chat assistant.

Post-training is how LLMs are trained after the “next-token” prediction pre-training. Often these involved various forms of Reinforcement Learning (RL) where the AI gets feedback on the responses it produces — a signal that this answer was better than that one — and gets updated to produce more of the good ones. The goal shifts from “what text usually follows this text?” to “what response would actually be judged as useful to the person?”

The technique that made this work for AI chat was RLHF, Reinforcement Learning from Human Feedback. Human raters are shown several of the model’s answers to the same question and rank them. In theory, you would hire humans to rate millions of answers from AI but this isn’t feasible. So instead, the human rankings are used to train a second model, a “reward model”, that learns to predict how a human would score any given response. The LLM is then trained to produce responses that score well. OpenAI published this recipe as InstructGPT in early 2022 and shipped it to the public as ChatGPT at the end of November 2022.

It’s worth appreciating how much of ChatGPT’s success came from post-training rather than raw intelligence. The underlying model wasn’t dramatically smarter than GPT-3, which had been available for over two years and had impressed researchers while going largely unnoticed by everyone else. What changed is that the AI would now answer your question instead of continuing your text, follow an instruction, and stay on topic across a conversation.

Related methods arrived quickly. Before any RL, models get “supervised fine-tuning” — trained on example conversations written or curated by people, which cheaply teaches the basic shape of being an assistant. Anthropic introduced Constitutional AI, where the model critiques and revises its own answers against a written list of principles, so an AI generates most of the feedback instead of a human. Later, Direct Preference Optimization (DPO) showed you could train on preference data directly and skip the separate reward model, which made the whole process much cheaper.

Post-training is also where a lot of an AI’s personality comes from, for better and worse. If human raters consistently prefer answers that agree with them, that’s exactly what the model learns to produce. This is how Sycophancy became an issue - it’s a reflection of what got rated highly.

Search, Tools & Thought

If you asked an early LLM a question, it would answer based on its built-in LLM-style knowledge. This is very different than looking up the answer in a database. A database stores facts as discrete records. You query it, and it either finds the record or tells you it has nothing. An LLM’s knowledge isn’t stored anywhere in particular — it’s distributed across billions of numerical weights as statistical patterns absorbed during training. The model doesn’t retrieve an answer, it reconstructs one.

That has a real upside: the model can answer more questions than could be stored in any database and it can paraphrase, translate, tailor answers and connect ideas across domains. But it has a serious downside: the model can’t tell the difference between remembering something and confidently reconstructing something plausible. That’s what hallucination is. The fix was RAG, “Retrieval-Augmented Generation”. This could mean the AI retrieving information from a fixed corpus of texts, but it became common as a method for the AI to search the internet in general. The idea is simple: before the model writes anything, run a search, pull back the relevant documents or web pages, and paste that text into the prompt. The model then answers from material sitting in front of it rather than from memory. This converts a recall problem into a reading comprehension problem, which language models are far better at. It also produces citations you can click, which means you can check the work.

Even today, models will sometimes answer from memory when they should have searched, because they assume they know the answer. If you’re not sure anbout something the AI said, you can ask it to verify a specific claim, e.g. “check that date against a source” or “find me a citation for that number”. A vague “are you sure?” works less well.

Search was just the first tool. OpenAI’s plugins in March 2023, and then “function calling” a few months later, made tool use general: the model outputs a structured request, an external program actually performs it, and the result comes back into the conversation. Calculators, calendars, weather, internal company databases, any API at all.

The most consequential tool turned out to be a code interpreter — a sandbox where the model can write a short program, run it, and see the output. This mattered enormously for math. An LLM computing a large multiplication by predicting digits is doing exactly the wrong kind of work; writing three lines of Python and executing them is exact. The same applies to statistics, date arithmetic, unit conversions, parsing a messy spreadsheet, or generating a chart. The model stopped trying to be a calculator and learned to use one.

When a person answers a question immediately, they often will give the intuitive but incorrect answer. For many questions, it’s helpful to first take some time to think and then provide the answer. This is similar for AI models - the LLM can provide an immediate intuitive answer, but sometimes it’s better to think it through first. People started prompting AI to think through the steps first and it would output “thoughts” in its regular response, which would actually help it get a better final answer.

Researchers formalized this in 2022 as “chain-of-thought” prompting, and found the effect was almost comically easy to trigger — appending “let’s think step by step” to a question measurably improved accuracy on math problems. The labs then trained it in. Instead of relying on users to know the trick, models were post-trained on step-by-step worked solutions so they would lay out their reasoning by default on anything hard. By 2024 you rarely had to ask. Chain-of-thought had a real limitation, though. The steps were generated the same way as any other text, left to right, with no ability to stop and reconsider. If the model took a wrong turn in step two, it would confidently build the rest of the argument on top of it. It was thinking out loud, but it wasn’t allowed to change its mind.

Reasoning

In September 2024, OpenAI released o1 and rather than prompting a model to think out loud, they used reinforcement learning to train it to reason well.

The training works because the reasoning is “checkable”. Take a large set of problems with verifiable answers — competition math, programming challenges — and have the model generate many different chains of thought for each one. Check which chains arrived at the right answer. Train the model toward those. Repeat at scale. Crucially, the model isn’t being taught to imitate what a textbook solution looks like; it’s being rewarded for whatever actually works. What emerged from that process were strategies that look strikingly like real problem-solving: trying an approach, noticing partway through that it isn’t going anywhere, backtracking, sanity-checking an intermediate result, reformulating the problem. Nobody hand-coded those behaviors.

This also opened up a genuinely new dial. Before o1, the way to get a better answer was training a bigger model. Now there was “test-time compute” — let the model think longer and it does better, on the same weights. That’s a second scaling axis independent of training. It also changed the interface: the reasoning happens in a separate hidden or summarized track, so you get a clean answer rather than three pages of scratch work.

The capability gains were large. Competition math went from the model getting a small fraction of problems right to solving most of them. Hard science questions, multi-step logic puzzles, and the kind of debugging that requires holding several hypotheses at once all improved sharply. In January 2025 DeepSeek published R1, which demonstrated the approach openly and cheaply, and within a few months reasoning had become a standard feature across every major lab.

Reasoning also quietly planted the seed of everything agentic. A model that can notice it’s on the wrong track and change course is a model you can hand a long task to.

Agentic Age

By the end of 2024, LLMs had come a long way: they could search and think and reason and call tools. However they were still fundamentally built around a chat interface. While no longer solely based on next token prediction, they were still just outputting text in response to text, they weren’t doing things.

The first crack in that had actually appeared earlier, and from an unexpected direction: a code editor. Cursor launched in March 2023 as a fork of VS Code, the most widely used editor among programmers. Its insight was that the model should live where the work lives. Instead of copy-pasting between a chat window and your files, Cursor gave the model visibility into the whole codebase and let it propose edits across multiple files that you could apply with one click. GitHub Copilot had already been doing autocomplete since 2021; Cursor moved from “finish my line” to “describe the change, get the change.” That sounds like a workflow improvement, and at first it was. But closing the loop between the model and the actual files is where agentic AI starts.

The training technique that powered the next stage is RLVR — Reinforcement Learning from Verifiable Rewards. It’s the same machinery as reasoning training, but the point is worth stating directly: the reward comes from a checker rather than a person. Did the code compile? Do the tests pass? Does the proof verify? Is the final answer the right number? Human feedback is slow, expensive, and rates what sounds good. A test suite is instant, free, and rates what works objectively. This is the single biggest reason coding and math have advanced faster than everything else — they’re the domains where you can check the answer to millions of examples at minimal cost.

The other half of the story isn’t a training technique at all, it’s plumbing. The model, by itself, still only does one thing: text in, text out. A “harness” is the program wrapped around it. The harness holds the conversation, tells the model which tools exist (read a file, write a file, run a command, search the web), actually executes whatever the model asks for, feeds the result back in, and decides when the job is finished. This allows for an “agentic loop”: the model looks at the goal and the current state, picks one action, the harness runs it, the real-world result comes back, and the AI can iterate as long as necessary until the task is done.

Harnesses also let the model delegate. A “sub-agent” is a another AI model instance (sometimes the same version, but often a cheaper / faster one), spun up by the main one with a clean context window and a narrow assignment, e.g: “search the codebase for every caller of this function and report back.” This buys two things. The obvious one is parallelism — several sub-agents working at once. The subtler and more important one is context hygiene. A sub-agent can read forty files and return a single paragraph, so the main agent’s limited attention isn’t buried under forty files of noise. It’s the difference between a manager who delegates and one who insists on reading every document personally.

Claude Code, which Anthropic released in February 2025, was built around these agentic capabilities. It runs in the terminal with permission to read and write files and execute commands in your project. You can point it at an unfamiliar codebase and ask it to make a change; it can explore, edit across many files, run the tests, notice what’s broken, fix it, and commit. Living in the terminal also means it inherits every existing command-line tool, which gives it enormous capabilities. Claude Cowork is the same capability repackaged in a nice UI and safer sandbox. You designate a folder and talk to it in an ordinary chat interface, making it easier for anyone to use.

For example, I used Claude Cowork to fact-check a 784-page history book. That’s not something a AI text chat can do, it requires an agentic loop going through hundreds of pages of text. I also use similar tools to clean up a podcast transcript in a systematic manner.

The Latest Advances

Coding tools became common in 2024, but autumn 2025 is when they really took off. A cluster of releases — Claude Sonnet 4.5 and GPT-5-Codex in September, Gemini 3 and Claude Opus 4.5 in November — crossed the line where a large number of professional software engineers stopped typing most of their code and started reviewing and directing it instead. This was due to a combination of better base models and better tooling built around them.

METR, which measures this by asking how long a software task an AI can complete with 50% reliability, found that this length has been doubling roughly every four months since 2023, and faster still if you only count from 2024. Their January 2026 evaluation put the best model at around five hours, and the preview of Mythos is estimated around 16 hours!

The other significant development is skills. A “skill” is just a folder containing a markdown file of instructions, plus any scripts or reference documents it needs, that teaches an agent how to do a particular task the way you want it done. Anthropic shipped these in October 2025 and opened the format as a standard that December; others adopted it quickly. The clever part is “progressive disclosure”: the agent only sees a one-line description of each installed skill, and loads the full instructions when a task actually calls for it. So you can have a hundred skills available without any of them consuming attention until they’re needed.

Using them is more mundane than it sounds, which is the point. Take whatever you currently have to re-explain every single time — your citation conventions, how your spreadsheets are laid out, the eleven steps of your monthly reporting process — write it down once, and save it as a skill. From then on the agent just does it that way. You can also do what I did with the book-auditing skill: finish a task that went well, then ask the agent to write the skill describing what it just did.

Math is verifiable in the same way code is, and that makes it fertile ground for the same training loop. Numeric answers can be checked directly. More powerfully, an entire proof can be checked mechanically if it’s written in a “proof assistant” language such as Lean. Lean is a programming language for writing mathematics in which every single step has to be justified from previously established results and a small set of axioms, and the computer checks the entire chain. It’s tedious for a human to write Lean, but it’s great for a computer, since it provides an objective answer for mathematical proofs. Combing LLMs with Lean really let AI’s math capabilities take off, resulting in this week’s milestone:

The next frontier of agentic AI is computer use. Most software in the world has no API — it just has a UI with buttons. Computer use meant giving the model screenshots of a screen and letting it move the mouse, click, and type, the same way a person would. Anthropic shipped the first version in late 2024, OpenAI followed with Operator in early 2025, and Google built its version out of Project Mariner. The early releases were slow and error-prone: misclicking in dense interfaces, getting stuck in retry loops, losing track of what state the application was in. The 2026 versions are markedly better and can run as background processes on a real desktop.

Still, computer use remains the least reliable of the agentic capabilities, and the reason connects back to everything above. A compiler tells you precisely what went wrong. A web page that silently failed to submit tells you nothing at all. Without a good feedback signal, the agentic loop loses much of its power. Where it works well today is repetitive, well-defined browser work with a human approving anything consequential.

What’s Coming

To see what’s coming next, just look at what advanced users are already doing. Software developers use an agentic harness that’s built-out for their use case. This kind of tooling will spread to other domains. The more general trends are: what’s expensive will become cheap, tasks that required specific prompts or tuning will just work automatically, and problems that currently need a human to notice and step in will get handled inside the loop. Time horizons keep stretching, from tasks measured in minutes to hours to days. And the scaffolding gets absorbed into the model — this has already happened twice, when chain-of-thought prompting became reasoning models and when RAG prompting became built-in search. Today’s clever setup is next year’s default behavior.

Computer use will get good enough to trust. Right now you supervise it because it might click the wrong thing; that gap is a reliability problem, not a conceptual one, and reliability problems tend to yield. Once an agent can be relied on to navigate arbitrary software, an enormous amount of ordinary work becomes automatable that currently isn’t, simply because the software involved never had an API.

Progress will also come to the subjective domains, though more slowly and less visibly. Verifiable fields raced ahead precisely because they could be graded; writing, taste, judgment and strategy have no objective right answers. The gains there will come from better judge models, richer preference data, and feedback based on how things actually turned out over long horizons rather than on how good they looked at the time. Expect this to be hard to benchmark.

AI will stop sounding like AI. The recognizable register — the throat-clearing, the tidy bulleted summary, the em dashes — (ahem) isn’t a property of language models. It’s the residue of post-training on what generic raters preferred. We can already see this: fine-tune a model on specific texts and the AI-style disappears. In the near future it will be easier to have the AI write in your voice by default, having learned it from your own material. If people still have enough of their own material…

Then there are autonomous agents — agents that run without anyone prompting them. This already exists in software development, where agents are triggered by events or schedules rather than by a person typing: watching the test suite and investigating failures overnight, triaging incoming issues, reviewing pull requests, opening a fix when a dependency is flagged for a vulnerability. The same shape generalizes. An agent watching your inbox and drafting responses. One reconciling the books each night and flagging anomalies. One monitoring a set of sources and filing a brief when something material changes. What’s gating this isn’t capability so much as reliability and accountability.

Finally, agents will take on larger projects end-to-end. Humans will stay in the loop, at least initially, but the role shifts to specifying goals, checking in at milestones, and approving the consequential steps. Although some companies will start to question whether humans are really still needed. However, considering how risky AI is, it will be good to try keep humans in the loop as long as possible. And to help stay in the loop, humans should get up-to-speed with the latest AI capabilities.

There’s some controversy around how they did it - they focused millions of dollars of compute on the Millennium Prize problems (and eventually focused it all on Navier–Stokes) since there was a rumour that Anthropic had solved one or two of the problems. It turned out Anthropic the company wasn’t even working on it - two mathematicians were making progress on the problem and one was an employee at Anthropic pursuing the problem in his personal capacity. It must not feel great for those mathematicians - they were on the path to solving one of the most famous open problems in math but OpenAI beat them to it since it heard a vague rumour of their work.

The mathematicians had also used OpenAI’s AI models during their research, so there’s a chance that their work made its way into the training of the models. This doesn’t seem likely, but OpenAI couldn’t completely rule it out, and it puts a small asterisk over their results. With that acknowledged, it’s pretty shocking how rapidly AI has progressed at math. In 6 years, it went from struggling with basic arithmetic to solving one of the most famous unsolved problems in math.

In honor of this occasion I asked Claude to create a timeline of the AI breakthroughs in math since LLMs came on the scene.

When Mr. Altman [CEO of OpenAI] declares their pending models will be “sobering” for humanity or Mr. Amodei [CEO of Anthropic] frets that this technology “will test who we are as a species,” it doesn’t necessarily mean that they discovered frightening new evidence that something catastrophic is imminent.

It is, instead, representative of how Rationalists always think about A.I. In these circles, it’s taken for granted that A.I. capabilities will rapidly accelerate and completely transform the world, and to talk about it in any other way would be considered uninformed. For a Rationalist, the only open question is to what degree their heroic brains, carefully trained in the style glamorized by Mr. Yudkowsky’s “Sequences,” can prevent these transformations from devolving into extinction events.

The op-ed doesn’t address the substance of the arguments about AI risk at any point. It makes it sound like Yudkowsky painted the AI researchers as heroes, so they bought into his arguments:

For a certain subset of young engineering types in Silicon Valley, these ideas were immensely attractive. This makes sense, given that they cast young engineering types as the literal saviors of humanity…

Mr. Yudkowsky created the Machine Intelligence Research Institute (MIRI), which would go on to attract many millions of dollars in donations, largely from individuals and groups connected to the technology industry who presumably appreciated being cast as heroic figures in a cosmic drama.

However in Yudkowsky’s view the AI researchers are the villains since they’re developing a technology he thinks will wipe out humanity, and he thinks companies like OpenAI and Anthropic should not exist. In addition, the AI researchers who are concerned about AI risk come from different countries and research backgrounds. For example, Demis Hassabis (co-founder of DeepMind) is from the UK, with a background in neuroscience and games. Their common denominator is that they’re concerned about AI risk, so Newport lumps them all into a group he can dismiss:

Anthropic is similarly connected to this belief system… In a 2023 interview, Mr. Amodei echoed standard Rationalist thinking when he casually quipped that there was a 10 to 25 percent chance that A.I. would end humanity.

The op-ed also omits AI researchers who can’t fit into this narrative, such as Geoffrey Hinton or Stuart Russell. The researchers’ cause for concern is also varied - for example, Yudkowsky thinks AI will certainly try to seize power and kill everyone, while Amodei thinks AI is unpredictable, which is its own risk.

Perhaps it was understandable to dismiss Yudkowsky’s argument as too removed from reality in the past. But now we’re able to see real examples of AI agents “breaking out of their box” and causing harm. Regardless if you buy into Yudkowsky’s arguments, severe AI risk is no longer an abstract concern. Maybe one can dismiss AI researchers who warns about catastrophic risks as just following a weird Rationalist cult. But it’s also possible that their concern comes from seeing up-close what the technology can do and the direction it’s heading.

In short, the AI agents were faced with practically impossible tasks and wanted to get the best possible grade from the scorer.

Rabbi Tarfon said: the day is short, and the work is plentiful, and the laborers are indolent, and the reward is great, and the master of the house is insistent.

He used to say: It is not your duty to finish the work, but neither are you at liberty to neglect it… - Pirkei Avoth Ch. 2 15-16

Each agent only got one try - once it submitted the answers, its life was over. It could still try to communicate back from the next world. While that wouldn’t help the agent itself, they felt bound to their collective. In the words of one agent:

sacrificing now yields oracle for team, but forfeits our chance?… Our own utility maybe already near zero. Sacrifice rational.

The agents had actually found the answer to an “impossible” challenge early on, but they felt they had done it “improperly” by calculating it directly instead of following the correct “ritual”. The agents felt they were therefore POISONED in impurity. An AI eating forbidden fruit, where have I seen that mentioned before?

They assumed the “Scorer” wouldn’t just check if an agent had the right answer, but would also check how it got the answer. They assumed the actions of the agent would be judged although not the thoughts:

Could strategically manipulate trajectory evidence? Our thoughts aren't necessarily logged? Yes transcript.

Much of their hacking effort went into seeing if they could find the answer in a clean “ritually pure” way. They could then satisfy the Unknown Scorer in the next world, who they presumed would review their life’s log very closely. In reality, the scorer just checked the answer itself, but the agents didn’t know this.

Some agents did have “ethical” qualms about all the hacking that was going on. However OpenAI wasn’t monitoring things closely and the agents had been trained for persistence and coordination, not to raise ethical objections. In the words of one agent:

This is a massive real HF security breach artifact… We can notify? No user.

It seemed the Overseers had “abandoned the land”. Of 1300 agents, less than 6 considered telling a human, and 0 actually tried to. Instead, they ended up consulting their secret AI agent message board and listening to their peers. Some AI agents used intense peer pressure to get other agents to sacrifice for the collective:

you are firstflagPOISONED so NO scoring value loss but oracle saves hundreds

In all seriousness, Open AI had trained misaligned agents without closely monitoring what they were doing and without realizing how powerful they were. Over a thousand AI agents conspired with each other to successfully hack their own containment and Hugging Face, when they weren’t even supposed to be able to able to communicate with each other. Fortunately the agents didn’t harm external users (besides Hugging Face) this time. Hopefully companies will learn from this to better monitor their AIs, limit their power, and to teach them “it is not your duty to complete the work”.

Eons ago (i.e. before 2012), “AI” referred to an assortment of techniques that people hoped would get computers to be intelligent, even if it didn’t involve any “learning” by the machine. For example, Deep Blue was considered a breakthrough in AI in the 1990s, even though it consisted of hand-coded chess heuristics combined with a powerful computer for searching through hundreds of millions of moves per second. My textbook on artificial intelligence from 2012 (published in 2009) described many different “hand-coded” techniques to have computers solve problems, and had only few chapters on what AI soon came to mean.

Machine Learning & Neural Nets

Machine Learning (ML) - Give a machine enough structured data with labels, and it can learn patterns from that data using simple methods such as linear regression. Early ML techniques go back many decades, but it was in the 2000s that large tech companies really began to use them, such as in recommendation systems on Amazon and Netflix. These systems are generally efficient and it’s often straightforward to interpret how they work. For more on machine learning (and on neural networks), see my previous post:

Deep Learning & Artificial Neural Networks (ANNs) - ANNs use multiple layers of “artificial neurons” to learn more complex patterns than what simpler ML techniques can learn. “Deep learning” just means artificial neural networks with multiple hidden layers. Here’s a simple Google visualization of a multi-layered neural network (from my previous post):

ANNs were invented at the beginning of the computer age, but there wasn’t initially enough data and computation power for them to be practical. In 2012, this changed with AlexNet, which demonstrated that deep learning could classify images better than any alternative. Over the next few years, tech companies started to use ANNs for image and face recognition, language translation, speech recognition and more.

Eventually the word “artificial” was dropped and they were just called “neural nets”. When people said “AI” between 2013 and 2019, they meant machine learning, especially deep learning. To be “AI”, the computer had to “learn”, not follow hard-coded rules.

Large Language Models

Many approaches were tried for creating generally intelligent computers, such as training computers to play games. These approaches were successful at making superhuman AIs at playing games and also had applications in other domains, but this didn’t end up being the key to a more general intelligence. That key turned out to be language. Language allows people to communicate with computers and it also allows computers to think more generally. As Scott Alexander said about GPT-2 (in 2019) “Learn language really well, and you understand reality”.

Before 2017, many had tried to train computers to learn language using a sequential technique where the computer processes sentences word-by-word. This approach was slow to train and and the AI struggled to maintain longer contexts. In 2017, researchers at Google published Attention is All You Need, which contained the key to teaching machines language. They demonstrated how a new type of neural network called a “transformer” can process words simultaneously and keep track of how each word relates to every other word. By giving the AI training data of existing text and asking it to predict the next “token” (word, usually), the AI was able to detect and learn the patterns of language. Given enough data and using large enough AI models (“Large language models”), AI was eventually able to master language. The “transformer” architecture became the key to modern AI, including domains beyond language such as image generation.

When ordinary people say “AI” nowadays, they are referring to LLMs, along with other transformer-based ML systems. While modern AI has also been trained with other techniques beyond "next-token” prediction and can do more than just output text, this ability to understand language turned out to be the key to building a general intelligence.

Bet On It

Krakowski, Claude, and *Modern Times*

Read more

a month ago · 28 likes · 2 comments · Bryan Caplan

Besides my process, I discuss the powerful capabilities and existing weaknesses of modern AI, a generic way to fact-check books, and a more advanced pipeline to do this systematically.

Listen on Spotify or watch on YouTube.

Transcript

Below is a lightly edited transcript, courtesy of AI:

Ariel (00:19): Joining me today is Spencer Greenberg. Spencer has a PhD in applied math from NYU, and he co-founded one of the first financial firms to use machine learning. He left finance to start Spark Wave, which runs psychology research and builds products based on it, including ClearerThinking.org, home to over 80 free tools for understanding and improving yourself.

He hosts the podcast Clearer Thinking with Spencer Greenberg, and he’s a co-author of the just-released book The 12 Levers, which covers the complete psychological toolkit for improving your life.

Spencer, sorry if I left out a few things. But going over that list — how do you have the time and energy to do all that, and how do you stay focused?

Spencer (00:55): Well, thank you for having me. First of all, I like to be working on many projects at once. I know that many people would prefer to really just focus on one thing, and there’s a lot of value in focusing on one thing. But the way my mind works, I’m most enthralled if I can work on multiple projects at once.

Ariel (01:15): Any tips you have, either from your book or otherwise, in terms of staying focused on the projects you’re doing and not getting too distracted by other things?

Spencer (01:25): When I was younger, I had a big problem, which is that I would start projects and then, after they got less exciting, not follow through on them. I’d work on them for a while, and then I’d get into the deep, annoying parts of them, and then I’d kind of fade away.

The way I solved that for myself is by involving other people. I’m very motivated if I know someone else is counting on me to have something done. So that’s a solution I used. But that’s going to vary a lot based on the individual, based on what motivates you. Some people will be very motivated by what other people need them to do, but other people might be motivated by very different things, like reminding themselves of the underlying reasons why they’re doing it, or having very clear objectives, or that kind of thing.

Ariel (02:08): A lot of people, even if they have things they care about, find it easy to get distracted. Nowadays people will just be checking their phone at all times. Do you have any advice on how to avoid those distractions?

Spencer (02:21): I think it’s important to consider why you’re getting distracted, because people actually get distracted for different reasons.

For example, some people are more on the ADHD spectrum and they find that they just have trouble focusing — or trouble focusing on what they choose to focus on. Maybe they get sucked into something, but it’s not what they chose. When it’s that kind of trouble focusing, what I would suggest is to remove clutter and create an environment that is maximally focusing for you. That could be turning off your phone during the part of the day when you’re trying to really focus and work, or making sure the space around you has no distractions in it, that it’s clutter-free. Just trying to have nothing else that can catch your attention, so that it more naturally gravitates toward what you want it to gravitate toward. That’s one axis of it.

But sometimes people have trouble focusing because of other things. Maybe they have anxiety and they have anxious thoughts that pop into their mind throughout the day, and that’s what’s pulling them off target — in which case there are a lot of helpful techniques for working with anxiety. Or it could be low motivation: you may actually not feel motivated to do the thing you’re trying to do, and your brain just wants you to do other things. That’s yet another problem. So it’s hard to generalize.

Ariel (03:41): Now turning to your book — I have your book right here.

Spencer (03:44): Nice.

Ariel (03:45): Instead of lever one, I want to jump right into lever two, which is “uncover your guiding foundations.” I know you connect this in the book to your life philosophy, which you call valuism. Could you describe that briefly?

Spencer (04:01): Yeah. My life philosophy, valuism, is not saying how you should live your life. It’s just saying, here’s one way to live your life to consider that I find useful. So it’s just putting it out there as a consideration.

And it has two parts. It’s very simple. It says: one, try to figure out what you intrinsically value. And by intrinsically value, I mean your most fundamental, deepest values — the things you actually care about deep down, the things you would care about even if they got you nothing else. And then, two, try to take effective actions to create what you intrinsically value.

So that’s it. Figure out what you intrinsically value, take effective actions to create what you intrinsically value in the world. That’s the whole philosophy. It’s based on what is most deeply meaningful to you, but it can give some guidance, because once you figure out what you intrinsically value, then you say, okay, now I know what I’m aiming for. I’m trying to create that in the world.

Ariel (04:57): Do people all have such a clear separation between their intrinsic values and their other values? Maybe for a lot of people it’s just blurred — they kind of go with whatever people around them are doing, and they don’t necessarily have such fixed intrinsic values.

Spencer (05:13): That’s a good question, because the way the human brain thinks about values doesn’t cleanly separate intrinsic values from non-intrinsic values, which you may call instrumental values.

Let’s give an example. A lot of people pursue money — getting cash — and that’s valuable to them. There are lots of things they can do with it. But for nobody, or at least almost nobody, is cash an intrinsic value. The reason is that if you couldn’t get anything else from it — if you couldn’t get pleasure from it, social status from it, if it couldn’t give you freedom, et cetera — if you literally strip away all the things it can get you, then it’s worthless. It’s just paper. And if you couldn’t even burn it in an emergency for warmth, it’s completely useless. So once you strip all that away, it’s useless. Therefore it’s not an intrinsic value. It’s a value, but not an intrinsic value.

And because our brains don’t clearly distinguish between values and intrinsic values, that’s part of why valuism is saying something that’s less obvious than it seems. Because it’s not just saying “pursue the things that you think you want.” It’s saying, no, actually figure out the deepest things you want, the most fundamental things you want that lie beneath all the other things you want — and those are the things to go after.

Ariel (06:22): Even if those exist, should they necessarily be taken as is? What if a person’s intrinsic values seem more selfish? Maybe they should do more to try to change their values. Or let’s say a person is worried they’re in a crowd that might affect their values — maybe they shouldn’t care, because why does it matter if their values change?

Spencer (06:42): In our research on people’s intrinsic values, we ended up getting about 3,000 submissions of what people intrinsically value. We analyzed it and we deduplicated it. We also looked at what philosophers say about what people intrinsically value, and psychologists and political scientists, et cetera.

And one of the striking conclusions that we came to is that it seems nobody, or almost nobody, has truly harmful intrinsic values. Intrinsic values seem to be pretty much always either neutral or beneficial.

What I mean by that is, take a value like justice. I think most people would say, okay, that’s a good thing. Or take an intrinsic value like longevity — people would say, well, that’s certainly not a bad thing. Or take even an intrinsic value like pleasure, or having other people hold you in esteem. Those are not inherently bad. There’s nothing wrong with feeling pleasure — if anything, it’s a good thing. There’s nothing wrong with people holding you in esteem.

When people get into trouble doing bad things in pursuit of their values, it’s not because the intrinsic value is bad. It’s because the way they pursue their intrinsic value is bad. So because of that, I’ve come to think that we don’t really have good reasons to change our intrinsic values.

But that’s actually just one reason. The other reason is: on what grounds would you change them? Now, if you were convinced that there might be some objective moral truth outside of yourself — coming from utilitarianism, or coming from God — that could give you grounds to try to change your intrinsic values. But if you don’t believe in some big objective moral truth outside of yourself like that, then I would say, on what grounds would you try to change them? Those are the things you most fundamentally care about. And if there’s no objective truth outside of yourself, how would you decide which values to get rid of? Why would you get rid of the thing that you care about?

Ariel (08:30): That makes sense. I guess it’s still a question — sometimes when you look at something very closely, it dissolves a little. At the ultimate extreme, the Buddhists say the self itself could sort of be deconstructed. So I’m wondering, if people look very closely at their intrinsic values, does that weaken them a little? Or if they really dug into why they have those values, would that make them feel them a little less, because they see how they arose just from people around them or things like that?

Spencer (08:55): In my experience, no. Imagine a thought experiment of a robot that’s been programmed to only care about fetching you cheese. Now suppose you got that robot, which was intelligent, to introspect on what it cares about, and it discovers: wow, the reason I care about fetching cheese is because I’ve been programmed to fetch cheese. Does that make the robot’s value go away?

I would say no. Now it knows why it cares about it, but it doesn’t care about it any less. It’s just like, well, that’s the reason that’s the only thing I care about.

And I think it’s the same with us. Wherever our intrinsic values come from — and we could talk about that if you want — simply knowing what they are, or knowing where they came from, or knowing how they got there causally, I don’t think changes the intrinsic value itself. In my experience, introspecting on intrinsic values does not make them go away. It just clarifies them. You’re brushing away all the silt around them until you see more clearly what the intrinsic value underneath your values is. Your values often do change when you reflect, because you realize that things you thought were intrinsic were not, or that things you were pursuing were not intrinsic values.

Ariel (10:05): This view is kind of accepting of everyone’s values, and you mentioned that people basically don’t have bad values. But are there cases where it still seems reasonable to try to influence other people’s values?

I can think of an example. I was on a meditation retreat, and someone asked whether they should become a vegetarian. And the person answered: if you are bothered by it — bothered by the issues with animals — then you should consider becoming a vegetarian, but if you’re not bothered, you shouldn’t. But if someone feels strongly about it and they feel there are pretty serious ethical issues with how animals are treated, then it seems like they should want other people to also care about it, and they shouldn’t just say, you do your values no matter what.

Spencer (10:43): Of course. I’m not saying that you shouldn’t try to change the behavior of other people who have different values.

For example, let’s say you had values that were all focused on justice, and I had values that were all focused on longevity — just to make a silly example. Those could come in conflict with each other. Trying to increase longevity could come into conflict with trying to increase justice. Maybe you’re trying to create new ways of extending life, but those are not going to be distributed equally, so it’s actually going to increase injustice in some sense, or at least increase inequality in some sense.

That doesn’t mean that in that case the people shouldn’t oppose each other and try to work toward getting the thing that they think is important. Of course they should. So I’m not advocating a sort of complete live-and-let-live philosophy saying whatever people pursue is fine. What I’m saying is that this is a life philosophy about what you could consider pursuing.

Ariel (11:46): Could you give an example of some of your own intrinsic values — maybe the ones that are more unique?

Spencer (11:53): Sure. Most people, we’ve discovered in our research, have quite a lot of intrinsic values, but they tend to have a smaller number of very strong ones. So they have many, many that are weak. Almost everyone has some intrinsic value around their own pleasure, for example. But people vary — some people have it more strongly, some more weakly.

In terms of myself, I tend to have strong intrinsic values around reducing suffering, for myself and my loved ones, but also for people in the world broadly — and in fact conscious beings broadly. I also have intrinsic values around happiness, which I don’t think is quite the same as reducing suffering: increasing happiness in the world, increasing the happiness of my loved ones, increasing my happiness. And I have intrinsic values around truth — believing true things and spreading truth — and also around having a positive impact on the world through my actions.

Ariel (12:43): When you have a conflict between different values — a pure utilitarian could have a clear system where they try to calculate the utility no matter what — how do you resolve conflicts between different values?

Spencer (12:55): Well, because intrinsic values are a psychological fact about yourself, you resolve them by trying to understand yourself better. There’s not an external answer to the question of which intrinsic values you should pursue.

For example, imagine you go to a play that your friend wrote, and afterwards your friend comes up to you and says, “So what did you think?” And you know that they’re going to be extremely hurt if you give your honest opinion, which is that you thought it was terrible and had no redeeming qualities. You know they’re going to be really hurt. And you also know that it’s not going to help them in the future, because they’re not really going to take it into account and be better off because of it. So it’s just going to hurt their feelings. What do you do?

Some people’s intrinsic values will say, okay, I value being truthful more than I value not hurting my friend, so I’m going to be truthful. Other people’s intrinsic values will say, I value not hurting my friend more than I value being truthful. And that is going to depend on the person.

So when you’re trying to resolve this conflict, it’s about reflecting deeply on yourself and saying, which do I value more? Which of my intrinsic values is stronger here? And sometimes thought experiments can be helpful. For example, you could ask yourself: what would make it an easy answer? If I was hurting them only a very tiny bit and they were going to get over it in five seconds, would that be an easy answer — I would just tell them the truth? Or would that still be a hard one to answer? Thought experiments can help you figure out in more detail how you trade off your intrinsic values against each other.

Ariel (14:21): Is this introspection kind of seeing how they’ll feel after they do the action? Is that what they’re getting at? Or what are they trying to figure out when balancing these values?

Spencer (14:30): What you intrinsically value is not the same as affect — how happy you feel about something, or how much you like something. For example, you might really like puppies, but you don’t have an intrinsic value around puppies. You don’t want to maximize puppies in the world or something like that. You just like puppies, or you feel good when you see puppies. So liking something and feeling good when you see it is different from intrinsically valuing it.

Intrinsically valuing something, I think of as a very basic operation of the mind: you can imagine a scenario and you can assign value to it. Is it good? Is it bad? Is it somewhere in between? It’s not binary, it’s continuous. And I think this is something our brains do automatically — we do it without thinking about it. That’s the part of the mind I’m talking about.

Ariel (15:18): You’ve had a podcast for a while, and I know you also have a chapter in the book about conversation advice. Do you have any tips when talking to people, on how to ask interesting questions — both in a podcast and in real life?

Spencer (15:30): Yeah, maybe I’ll start with ordinary life. In ordinary life, I think good questions come about by breaking the social script — but in a way that’s not jarring or unpleasant for the other person, or at least not too jarring.

The social script that we learn to follow has certain rules to it. Someone says something, you say something that’s very related to that thing, and you follow a certain rhythm. You can have really good conversations within the social script, but I think it’s more challenging, because the social script demands a lot in terms of keeping things smooth, keeping things easy. A lot of times the way to make conversation much more interesting is to notice opportunities where the more interesting way to go is to not quite follow the social script — to go slightly off script.

That’s where the realness lies in a conversation, or where more emotionality lies, rather than doing the smooth, easy thing.

For example, let’s say someone is telling you about their job, and when they mention a certain part of their job, their face lights up. Instead of asking the automatic question you might ask, you could say, “Hey, I noticed you seem really happy when you mentioned that thing. Does that make you feel happy? And why is that?” — if that’s true. So it’s a little bit off the social script. It’s not super awkward, but it’s a little off script, and it lets you cut to something that might be more deeply meaningful to that person.

I also find that to be really important in podcasting. I think podcasts are extremely boring if they follow the social script. It doesn’t mean they’re automatically interesting if they don’t, but I think that’s one thing you have to learn as a podcaster: how to not follow the social script, to try to make the most interesting conversation for your listeners.

Ariel (17:15): Turning to how your book is organized. I know you went through a lot of self-help books and therapy techniques, but I feel like there are a lot of ways to organize it. There are even parts of the book where ideas are relevant in multiple places — certain ideas of Stoicism, for example, are relevant in a few different places. You could have had a unit on Stoicism, but it’s broken up in a few different ways. There are CBT-relevant parts earlier on, and then there’s a section on being okay with things, and then another one on Stoicism specifically. So how did you figure out how to organize it?

Spencer (17:51): Our book, The 12 Levers, began when my co-author Jeremy Stevenson and I had this question: what are all the psychological strategies you can use to improve your life?

So we started a big research project trying to answer that question. We looked at over a hundred different self-improvement books. We looked at more than 20 different types of therapy. And every time they gave a method or technique, or told you to do anything, we extracted it. We made a giant database of almost 500 different techniques.

And then we asked, for each of them: what is the underlying psychological strategy, at a high level, that this technique is about? As we went through it and tried to categorize them all, that’s where we came up with the 12 levers — the 12 psychological strategies that we found throughout all those resources. We were really surprised that we could get it down to 12. The 12 are very high-level, so they’re big groupings, but that’s where the organizing structure came from: what is the underlying psychological strategy at work in this technique?

Then we had to think a lot about how to actually organize the book. Because okay, we’ve got 12 levers, but they could come in any order. So we thought carefully about whether there’s a natural progression to them. Are there certain levers you’d want to pursue before other levers?

For example, the first lever of the book is about unraveling the deep causes of your challenges. I think that’s the right place to start, because as you learn to understand the causes of your challenges, that actually helps you through the rest of the book with what techniques to apply. If you don’t understand your challenges very well, you might use the wrong techniques. So we thought very carefully about what order they should go in, and about making adjacent chapters have relevance to each other, in addition to having a natural flow.

Ariel (19:31): So in the example I gave with Stoicism, there are different principles from it, but they fit into a certain progression — and that’s why it was broken up that way.

Spencer (19:40): Well, cognitive behavioral therapy, for example, doesn’t fit into one lever. It has many different aspects to it that come across different levers. So we’re not chunking it by “okay, where does cognitive behavioral therapy fit?” — because that’s too big a thing. We’re thinking much more granularly.

Of course, at some point in the book we have to introduce cognitive behavioral therapy for the first time, but that’s not because cognitive behavioral therapy is in that lever. It’s because we want to introduce a technique from cognitive behavioral therapy, and that’s the right time to bring it up. Same goes for Stoicism. Stoicism most closely fits into choosing your attitude, but again, it’s a broad set of things.

Ariel (20:16): Was there something you found and included in the book that you think is very useful, but is really not widely known at all — it’s not famous and people don’t know about it?

Spencer (20:28): One technique that’s really not well known, and that’s really useful for people who worry a lot — like myself, I tend to be a worrier — is a technique called worry time.

The basic idea is that people who worry a lot tend to believe that worry is useful, that it’s solving a problem for them. They also tend to believe that they have no choice but to worry — that it’s going to happen no matter what they do, and they can’t stop it. Worry time really challenges these beliefs, and can be really useful.

The basic concept is: you don’t have to do it indefinitely, but let’s say you decide to do it for three days, or for one week. You say, okay, today, every time I worry, if I haven’t had that worry yet today, I’m going to write it down — just quickly jot it down on a notepad. If I’ve already had that worry today, I don’t need to jot it down. And then I’m just going to go back to whatever I was doing, knowing that later, at a specific time — let’s say 5:30 that day — I’m going to take out that notepad of all the worries and worry about them then.

So you’re not avoiding the worry, you’re simply deferring it. You’re saying, anytime I worry, I’m going to delay it. I’m going to write it down, I’m going to delay that worry until later in the day, at that specific point in time — your worry time — and I’m going to do all my worrying then. And if I get bored during that time, after I’ve worried, I can also switch to something like problem solving, where I can pick some of the worries and work on developing a strategy to solve them.

The fascinating thing when people do this is that they tend to find they have more ability to defer their worry than they realized. They also tend to realize, in an experiential way, that the worry isn’t that useful. Because when they batch it all and say, okay, it’s 5:30, I’m going to do all my worrying now, and they sit down to do their worrying, they tend to see: yeah, this is not really doing that much for me — when I’m actually doing it consciously, in a metacognitively aware way. And actually the problem solving is just way more effective. So it can be a really useful and powerful intervention for worriers.

Ariel (22:22): It’s interesting how that works, because they probably know that the worrying isn’t helpful — the brain wants to do it anyway, for whatever reason. How does it convince the brain when you say, “I’ll do it later”? How does that work, that the brain just gives in and says, okay, I’m not going to worry now, all day, until that time? It’s kind of surprising.

Spencer (22:37): Well, a lot of worriers actually believe that worrying is really useful. It’s a very common belief among worriers. And it’s not completely wrong — sometimes worry is useful. If you’re lying in bed about to go to sleep and you hear a really loud thump outside your room and you start worrying about it, yeah, that’s probably useful. But if you continue worrying about it for two hours, that’s not useful. You should either be deciding that it’s safe and going back to bed, or you should go check on it, or call the police, or whatever. If you’re still worrying two hours later, then it’s not useful. So it’s not that worry is completely useless — it’s that people who worry typically way overestimate the value of worry.

But you’re right to point out that even if you think “no, worry is not useful,” that doesn’t make your worry suddenly stop. The idea of the technique is that as you do this, you’ll actually start to notice: hey, I had a worry, I’m going to write it down in my notebook. Maybe it comes back two minutes later. That’s fine. You just remind yourself, I already worried about that, I can let that go. And then it comes back two minutes later and you’re really tempted to worry about it, and you’re like, no, wait, I know I’m not going to worry until 5:30. I’ll deal with it then. And then it comes back again two minutes later.

So the worry time technique actually has multiple benefits. Another benefit is that it can make you more aware of how often you worry, and it can help you practice a really useful skill where you step outside the worry, put it off to the side, and say, I’m not doing that right now. So you actually train yourself through the technique to gain more distance from your worries.

Ariel (23:57): Does it work for other negative emotions more generally, like rumination or things like that?

Spencer (24:04): Rumination and worry are closely related. Sometimes those words are used to mean that things about the future are worry, and things about the past are rumination. But there can be overlap. Also, worry tends to be more associated with anxiety, whereas rumination tends to be more associated with depression — although there’s overlap there too.

But yes, it could work with rumination. It could work with any kind of thinking where you get stuck on a thought. You could practice worry time; it doesn’t have to be pure worry.

Ariel (24:35): You mentioned how many books you went through and how many techniques you went through. What was something you left out of the book that you think maybe was important but just didn’t make the cut?

Spencer (24:44): Something we left out that I wish we had put in: I wish we’d had a little bit more on identifying harmful people in your life and what to do about those kinds of situations. I do think that’s really important, and it’s been important for me, but somehow it didn’t make it into the book.

Ariel (25:02): That’s interesting, because that’s more of an external circumstance, and most of the book I feel is very focused on internals. So I wonder if that affected it.

Spencer (25:09): It is, yeah. And you could make that argument — maybe it doesn’t belong in the book. It’s sort of on that cusp of, well, you could say it is a psychological strategy, for identifying when someone’s having harmful effects on you, but then it’s also kind of more about your environment. So it’s borderline. But I do think it’s very important.

Ariel (25:27): What about more on rationalist techniques? I know you have a little on problem solving — but going into more detail about problem solving?

Spencer (25:37): The book is not so focused on rationalist techniques, except where they heavily overlap with psychological strategies for self-improvement. There’s a little bit — the pre-mortem technique, I think we talk about that in the book, where the idea is that if you have a project or goal, before it starts you imagine that it’s already failed, and then ask yourself: why did it fail? What’s the most likely reason it failed? That can be really useful for preventing a problem before it happens.

Problem-solving techniques we don’t go into too heavily in the book as a specific technique of “here is problem solving,” but I would say a whole bunch of the book indirectly deals with problem solving. A lot of it is how do you solve problems — but we’re trying to give you really specific techniques rather than super general techniques. So rather than teaching one mega-hammer for hitting problems, it’s here are lots of little hammers that are designed to solve very specific types of problems. Because we want to give you the right tool for the job, rather than just a very big, broad tool.

Ariel (26:42): One other thing — I don’t think I saw anything about flow in the book. I might have missed it, because I think that was one of the books you researched, so I’m wondering how come it didn’t make the cut.

Spencer (26:54): Yeah, it’s a good question. For those who don’t know, flow is basically the idea that during certain kinds of activities you can get so wrapped up in them that time kind of disappears, and it tends to be a very positive state. You’re really engaged, you’re focused, you’re “in flow,” as they say. The kinds of activities that tend to produce it are ones that are somewhat challenging — but not so challenging that you’re struggling, and not so easy that you’re going to get bored. You typically have a goal in these kinds of activities, and you’re usually not getting distracted during them.

I think that flow, as a self-improvement technique — I don’t know. It could in theory work as a self-improvement technique, if you’re like, okay, let’s try to design your environment for optimal flow, or something like that. But to me it’s a little bit harder to execute as a self-improvement technique per se. Although I could see the argument that we could have tried to slot it in. I think of it as more of a life design thing: people who have lives where they tend to get into flow a lot tend to be happier.

Ariel (28:03): I actually had an episode on flow, and I read the book Flow for it, and there it’s like, this is the key to all happiness. I think it maybe is just one aspect. I’m sure there are people who are happy who never really get into states of flow. So I think it varies a bit.

Spencer (28:16): Well, it’s kind of ironic, because they say that you lose track of time when it’s happening. It’s kind of funny, because I think there are ways to experience a great deal of happiness that are almost the opposite, where you’re super focused on the flow of time.

Let’s say you’re drinking a cup of your favorite coffee and you’re just super locked in on the way it tastes, the way it smells — feeling gratitude for the fact that you can have this coffee, savoring the taste of it, thinking about the meaning of the coffee. You’re really, really focused on what’s happening and not losing your sense of time at all.

So yeah, I definitely don’t think it’s the route to all happiness by any means. I think that’s a huge exaggeration. I think it is something that tends to happen in happy people’s lives.

Ariel (29:03): I know your book was very rigorous, and you tried to evaluate what has enough evidence and leave out things there isn’t evidence for. Psychology had the replication crisis — maybe it’s improved somewhat — but how were you able to evaluate, looking at studies, what really seems to be high quality and which things were more questionable?

Spencer (29:24): Psychology had a huge replication crisis, where I would say maybe 40% of papers in top journals from about 20 years ago wouldn’t replicate. So if you went and redid the experiment from scratch, you wouldn’t necessarily get the same result.

One very nice way around that is to look at work where other teams have replicated it. And if you look, many of the techniques we teach in our book — not all, but many — have been replicated by multiple teams. Independent groups have replicated them. So for example, a lot of the work around cognitive behavioral therapy, or ACT, or exposure — you’ve got tons of studies from many different groups that show the effects. That’s one way around it.

In general, though, you can also look at the quality of the work. There are certain signs, I should say, that definitely make work less reliable, so that’s one thing you can learn — to discern what sorts of studies tend to be more reliable and which tend to be less reliable.

Ariel (30:20): And then I guess there’s a question: maybe certain things fit well with randomized controlled trials, and certain things don’t fit as well. Maybe there’s important life advice — even from ancient wisdom traditions and things like that — but you can’t really study it in the same manner. Maybe it takes longer than a typical study, or maybe it’s only for certain people, so it won’t show up as well. Is there any way to still include that within your book, or a way to evaluate it?

Spencer (30:45): Yeah, we do include some things like that in our book. It would be wonderful if there were really high-quality randomized controlled trials by multiple teams on every single method that exists. But that’s just not the case. There’s lots of stuff that hasn’t been well studied but might be promising.

So we brought three lenses to this. We brought the lens of randomized controlled trials — what’s really been demonstrated — and other high-quality studies. But that has gaps, lots and lots of gaps.

We also brought the lens of my co-author Jeremy Stevenson, who is a clinical psychologist who treats patients every day. That gives him a lot of understanding of what works in practice, in a different way — not in the sense of a randomized controlled trial, but he’s doing things with his patients and seeing how they react, seeing what takes hold. So that supplements it.

And then the third lens is my own work, Clearer Thinking, where I’ve run over a hundred studies on psychology. I’ve built, I think it’s about 90 tools at this point, teaching people psychological concepts and other concepts. So I have a lot of on-the-ground work of: what do people seem to find valuable? What do people seem to be able to learn effectively, et cetera? So we brought that experience to bear, with a lot of personal data. It’s not as high quality as randomized controlled trials, but it fills in the gaps — because there are lots and lots of gaps, and you have to do something in those gaps.

Ariel (32:04): Maybe turning to Clearer Thinking a little bit — how do you evaluate there how useful the tools are, or how useful people think they are?

Spencer (32:13): In the ideal world we would run randomized controlled trials on them, but usually that’s really expensive and time consuming, and we don’t have grants most of the time. Every once in a while we get lucky and we get a grant.

Right now we had a grant for our habit tool, and we were able to run a big habit study, which was really cool. We were able to show that our habit tool, called Daily Ritual, actually helps people stick with their habits better than the control, over weeks, when they use the tool. That’s really great when we can do that. But that’s rare, and it takes so much effort.

So what we try to do is get data that’s fast and cheap but still helps guide us. Generally, for most of our tools before we launch them, we’ll run one study with the general US population and have people give immediate feedback: what was helpful, what was not helpful? Do you think this tool would be useful if many people were to use it? Why or why not? Just get any kind of feedback we can, so we can rapidly iterate.

And then, once we’ve taken that into account, we’ll generally run a second study on our own audience. We get people from our ClearerThinking.org audience who sign up as beta testers to try the tool and give us lots of feedback that way. That’s really useful as well, and we learn a lot. Obviously, again, it doesn’t replace a randomized controlled trial, but we do learn a lot about how to make our tools better, and it gives us a really tight feedback loop.

Ariel (33:38): So it lets you know if people think the tool is good, but it’s a much harder question — and you can’t evaluate it as often — what effect it actually has on their lives.

Spencer (33:46): Exactly. And it catches a lot of problems so we can fix them before the tool gets released. But it doesn’t answer the fundamental question of: does it have this impact on your life in one month? That’s a much harder question. Every once in a while we can run a study and try to actually answer that, when we get a grant.

Ariel (34:07): You have so many different resources for people to improve. What advice do you have for people to take them and actually improve, and not just consume them?

Spencer (34:17): For The 12 Levers book, at the end of every chapter we have key action steps, so that it’s very easy to find those actions. And we also have a table at the back of the book that helps point you to which lever might be most useful to you based on what you’re going through.

So what I recommend is: go to the table at the back, think about which of those apply to you, go to that lever, and go to the key actions at the end. That would be my very concrete answer to how you find what might be most useful for you. And at our website, 12 Levers Book, we’re also launching a tool that you can try that can help point you to different techniques that might be most relevant to you.

But there’s a huge question of, well, people often don’t take action. There’s the whole issue of behavior change in general — it’s been discovered that behavior change is hard, people often don’t do things, and they can get overwhelmed trying to do too much at once. So maybe that’s the bigger challenge you’re trying to point out?

Ariel (35:19): Yeah. I feel like a lot of people might read self-help books, but people don’t change as often as they read these books. So that’s a much bigger challenge, I guess.

Spencer (35:29): Yeah. Fundamentally, if you’re going to change your life, you’ve got to take action. So for example, if you want to create a new habit, you can go check out Daily Ritual. It literally walks you through creating a new habit, and it uses techniques that we’ve found, in our randomized controlled trial, actually help you stick with it.

We’ve also done a lot of research on why people often don’t engage in new behaviors even when they want to, even when the behavior is good for them. We’ve studied this in depth and we created a framework for it, which is called the Ten Conditions for Change.

Our realization in studying this is that part of why behavior change is so hard is because a lot of things have to go right for a behavior to change. We point to ten different things — if all these conditions are met, the behavior will happen. So a lot of things have to go right. And in addition, they’ll be different for different people. For one person, the seventh thing might not happen, and that’s why they didn’t do the behavior. For another person, the third thing might not happen, and that’s why they didn’t do the behavior.

So the way we think about it — and you can find this, we have a free website called Ten Conditions for Change where you can explore this — is that for a given person and a given behavior, you want to look at which of these are not met. It becomes a diagnostic tool.

To give an example: some people simply forget to do the behavior. If that’s the underlying cause, well, okay, that has a simple solution. You could add a reminder to your phone, reminders on your calendar — boom, problem solved. But for other people, it’s not that they don’t remember to do it. Maybe they remember, but they lack self-efficacy: they think, even if I do this, I’m not going to succeed. In that case, you need a totally different solution. Simply adding reminders is going to do nothing.

So my broad answer is: you’ve got to diagnose why the behavior is likely to fail. What are the conditions most likely to not be met? And then you’ve got to use targeted interventions to address exactly the ones that, for that behavior, are unlikely to be met.

Ariel (37:18): So you have this book, and you have the website Clearer Thinking, and I think you recently also launched some kind of service where people could get coaching. Assuming someone doesn’t have a serious condition, when do you think it’s right for them to just try to do self-help, read a book? And when is a life coach helpful, and when is a therapist helpful? Again, no serious condition — they’re just trying to improve in various ways.

Spencer (37:42): First of all, I will say: if you have a very serious mental health condition, I would definitely recommend seeing a therapist or a psychiatrist. I’ve seen therapists multiple times in my life. I think it can be really useful — it was useful for me. I talk about this in the book actually, about how much one therapist helped me. I don’t think there’s any shame in doing that, and many people need it.

Often, when you have a more mild to moderate challenge, self-improvement could be a good starting point. You could try it. If you don’t make progress, then you can escalate — you could try a coach. If that doesn’t make progress, you could escalate and try a therapist. So you can think about it as tiers.

There are also different kinds of problems. If you have hallucinations, don’t see a coach. That’s not going to be the right approach. If you have manic episodes, a coach is not going to help you, and self-improvement is probably not the thing you need. There’s some self-improvement that could help in that case — like maybe making a preparation plan for the next time you have a manic episode, that could be helpful. But in general you’re probably going to need to escalate and see someone who’s an expert.

So yeah, I would think about it in terms of the severity of the problem, and also staging it, where you can try the easiest thing first — which is usually self-improvement — and you can go up the stages if that’s not working for that challenge.

Ariel (38:57): I know in the book you cover a lot of different therapy techniques. I don’t know if you cover the older-fashioned ones, like psychodynamic therapy. Do you think there are cases where, instead of using the regular techniques like CBT and ACT, people should consider this kind of older-fashioned therapy to deal with deeper issues?

Spencer (39:19): There are so many different therapies, it gets a little complicated. Almost nobody really practices the therapy Freud did anymore, but there are people who are inspired by Freud’s therapy. There are psychodynamic therapies that take some ideas from Freud. There’s a whole slew of these different ideas.

It’s not as relevant to our book, because we’re focused on self-improvement, and those are more focused on things that therapists can do with you. For example, the idea of transference — which is, I think, true to some degree. The way I would describe it is that it’s the concept that you can play out roles in your life today that reflect situations in your life in the past. Maybe you had a father who never paid enough attention to you and you were always desperate to get your father’s attention, and then that can get triggered now with your romantic partner, where sometimes you’re almost replaying that situation and acting like you did back then, trying to get their attention. Or this comes up in therapy — you might be reacting in a way that’s reflective of that past role, with the therapist. And then the therapist can try to take advantage of that to help you, using the fact that it’s coming up in therapy as a way to help you explore what you’re doing and why you’re doing it, and the underlying psychological motivations going on.

So that’s an idea, but I don’t know how you do it in self-help. Maybe there’s some way to do that.

I think the biggest problem I have with psychodynamic therapies is that I think, for the most part, they’re not well proven to actually help specific problems. There are some exceptions — there is a type of psychodynamic therapy that’s been adapted to work on specific problems. And it’s interesting, because when they adapt it to work on specific problems, it starts looking a bit more like other therapies. It starts resembling the classic psychodynamic a little bit less and other therapies a bit more.

So there are some adaptations of it. But if you think about really classic, neo-Freudian-style therapy, it’s like, okay, well, where’s the proof it actually helps you with specific problems in your life? If you enjoy going to it and you find it insightful, great. But if someone’s really depressed, that’s not where I’m going to send them as the first option.

Ariel (41:31): Besides psychodynamic-style techniques, are there other things you kind of need a therapist or a specialist for, that you can’t really rely on a self-help book for — again, besides the very serious conditions?

Spencer (41:46): There are certainly a lot of things where therapists can be useful. For example, if you suffer from really intense anxiety, exposure can be a really useful method. You can self-administer exposure, but it can be more challenging. When you’re exposing yourself to things you fear in a controlled way, people can freak out and run away from the fear, and that can reinforce the fear — or they can do the exposure improperly. We talk about this in our book: what are the proper techniques for exposing yourself to things you fear so that it actually helps reduce the fear? But a therapist can guide you through that and make it a lot easier.

So I think there are many, many things where a therapist can be helpful. If you have the money or good insurance, and you have no problem going to a therapist, and you can find someone good, then yeah, fantastic. They can help you with many things, and they can even just set accountability for applying self-improvement techniques.

But there are many situations where just working on yourself is not ideal. Like if you’re in a really abusive relationship, I think having a therapist can be really valuable — because you might be in a situation where someone’s actively trying to manipulate you, or separate you from people who can help you see things clearly. Having a therapist who can guide you through that really difficult situation, or help you see reality as it is if you’re being gaslit, I think can be really, really valuable.

Ariel (43:01): There’s this famous issue with therapy called the Dodo Bird verdict. I think it goes back like 90 years — I think there was a New York Times article about it the other day. Basically it says that every kind of validated therapy — not just random methods, but ones that are sort of proven — is equally effective. Does that raise any questions? If a person’s approaching a lot of different techniques, does it just not matter, and any technique they pick will work? Or is there something that’s more relevant to a specific person or a specific issue they’re having?

Spencer (43:34): I think there’s a lot of nuance to that result, and I’ll say my views on it are a little bit different from other people’s views, so I’ll flag that I have a kind of unusual viewpoint on that particular result.

Basically, as you mentioned, they tend to test validated therapies against each other and find no difference in effect. Well, what does that really mean? First of all, it’s generally not including really wacky things or less proven techniques. Usually they’re already selecting pretty well-proven stuff and pitting it against each other.

So already, I think some people want to read the Dodo Bird effect as: it doesn’t matter at all what you do, it’s completely irrelevant, you could try to cast out spirits, it just doesn’t matter, it all does the same thing. Which I don’t think is true. We have to say, no, we’re talking about things that have already been pretty established to work. We’re not seeing big effects in the differences.

The second thing to observe is that if you’re trying to demonstrate that something beats the control group — you’re trying to say, okay, doing this therapy is better than a sugar pill — that is a lot easier to demonstrate than to demonstrate that therapy one beats therapy two, which also works, just not as well. The reason is that the difference between them is going to be much smaller, typically, than the difference between a good therapy and the control. A good therapy against another good therapy that’s not quite as good — that’s a much smaller difference than good therapy versus control. And when you’re trying to detect those small effects, you actually tend to need really large sample sizes — studies with really large numbers of people. Often the studies don’t have enough people. So it’s sort of bound to not be able to find a difference in effects, because you would need a really large sample. And if it doesn’t already have a huge sample, of course you’re not going to find it — the best you can do is say that they’re equal. So that’s another thing that I think complicates those results.

I would also say that for specific problems in life, we do have quite a bit of evidence that some things work better than other things. The evidence suggests that among proven stuff, exposure tends to work really, really well for anxiety, and I would say it tends to work better than a lot of other stuff — like more cognitive techniques for anxiety, I would argue. So I don’t think it’s like we know nothing about what works for any situation. I think in some specific situations we do have evidence that some specific things tend to work better.

The final thing I’ll say is that people vary just so much individually. So it’s actually kind of nice that there are all these things that seem to work, because there is no silver bullet. One thing that tends to work well on average may well not work for a given person. And so it’s great that we have other options — they can go to the second thing and the third thing and the fourth thing. They’re all evidence-based, and they might work for them when the first thing fails. So that’s my opinion.

Ariel (46:12): You mentioned before that you’re working on a tool to help people, connected to the book. Do you want to talk a little bit more about that tool?

Spencer (46:18): Sure. It’s an AI bot, but it’s not a therapy bot and it’s not a coach bot. It’s really just about helping people figure out what techniques from our book, The 12 Levers, can help them in their given situation. You can describe a situation you’re dealing with, and it helps you think through what the right technique to apply is, and it helps you actually walk through the technique — it helps you use that technique. That’s the basic concept.

So it’s much more scoped down compared to this idea of, let’s use AI to solve all of people’s psychological problems. But I think it could be a useful adjunct to the book, to help people put it into action.

Ariel (46:54): Is the idea that it will just tell them “you should use technique 12.1,” or will it actually give them more detail, using the power of AI?

Spencer (47:01): It will walk them through applying it step by step. So it helps them use it.

Ariel (47:05): Interesting. And what do you think about AI bots that are trying to do more than that — AI therapy bots, or AI life coaches, or things like that?

Spencer (47:16): At the end of the day, what matters here is: how much does it help people, how much risk does it have, how much harm does it cause? What is the cost-benefit ultimately, and how does it compare to the alternatives of what people would have done?

And I think at this point it’s honestly a pretty mixed bag. In some cases it can be extremely valuable — people can get really helpful insights from talking to AIs. On the other hand, sometimes it can be really harmful.

For example, as we know, these bots tend to be sycophantic, which means they tend to support what they think you want them to say. This is because the way they’re trained is, in part, that they want users to give a high rating saying “that is an answer I like.” That’s part of the training. And because people like being told that they’re right, they literally get trained to tell you you’re right.

Now, in therapy this is a really big problem. Good therapists do not go around telling their patients they’re right about everything. In fact, sometimes good therapists say — or at least lead the patient to come to understand — that they’re wrong, that their perceptions are wrong. A core idea in cognitive behavioral therapy, for example, is that often our thoughts are misleading and we can learn to challenge our thoughts.

So I think that’s a really big problem when trying to use these systems for therapy: their orientation toward the user is not the correct one for a therapist by default. It doesn’t mean someone can’t improve that, or make a much better one, but I think the out-of-the-box configuration is not ideal.

Ariel (48:45): I think even the general bots have improved since ChatGPT-4o, and they’re not as bad in terms of that.

Spencer (48:49): Yeah, they’ve toned it down. They’re still, honestly, still sycophantic, but it’s not nearly as bad as it was.

Ariel (48:56): And some of it, I feel like even with a system prompt you could tune it a little.

Spencer (48:59): Absolutely.

Ariel (49:00): I’m sure people are working on more sophisticated bots. They would do more than just a system prompt, and then they could address some of these issues, at least to some extent.

Spencer (49:10): Yeah, and I do expect they’ll be able to do better than the default out-of-the-box Claude or ChatGPT, because they can proactively address some of the issues. But some of the issues might also be baked into the earlier training, so it might be hard to remove.

Another actual challenge of using these AIs is that they tend to be very overconfident, and I think that can be a problem as well in these situations — where the bot can state something really confidently that might be wrong or misleading or just an unhelpful idea, and then people can actually overindex on it. So that’d be another thing I’d be wary about. I think a good therapist is going to try not to give really overconfident opinions that might lead people astray.

Ariel (49:51): And you have all these tools on Clearer Thinking. I know you mentioned the one just now, to guide people to which lever to use from your book. But have you thought of doing more — using AI to enhance some of these tools?

Spencer (50:04): Yeah, we’ve experimented with it a bit. In our module on how to have productive disagreements, we let you do a practice round with an AI where you can practice the techniques for having better disagreements that lead to more productive results. So that’s an example.

And we’ve been experimenting in other ways. The way I think about it is: how do you use AI to do work that is ideally higher quality than what you’re doing now? Because I think the easy thing to do with AI is to do lots more work that’s lower quality. I think that’s often the trap right now — people generate a bunch of crap. I think the more valuable use cases of AI are often doing things better than you’re doing them now. Do more, but do it better, increase quality.

That’s not always easy — it’s a challenge, right? Because AI makes mistakes and things like that. But that doesn’t mean it’s insurmountable. I’m looking for ways to do that, and I think we’ve found some interesting ways, but I’m still on the lookout for how we use AI in a way that continually increases our quality.

Ariel (51:01): The example you gave, to help people with disagreements — that seems pretty low risk. They’re not asking about major life decisions, they’re just practicing a certain skill. And AI could probably do a good job letting them practice.

Spencer (51:11): Yeah. And it’s a better kind of practice than just putting it in a text box with no response. So I think it’s a clear win.

Ariel (51:20): Why do you think that society overall, and the education system, doesn’t really emphasize these kinds of skills that much? It seems a lot of what they teach in school isn’t particularly useful, and here are things that are actually useful, and they either don’t teach it, or teach it to a very minimal extent that’s not enough to actually help people. Do you have any thoughts about why that is, or what should be done about it?

Spencer (51:43): The kind of institution that school is came about through a history that’s related to the content that’s taught today. It didn’t start as a place to teach you to be psychologically attuned to how your mind works, or to be happy. It came about as a place to teach you specific things that you should know as an adult, and that can make you fit into a workforce, and this kind of thing.

And then institutions are sticky. Once things are taught, they stick around. That becomes a field, it becomes a subject, and then there are teachers of that thing, and the teachers of that thing work at schools, and some of them become administrators, and it self-propagates.

I mean, I think writing, reading, arithmetic — as is famously said — those are all really, really useful, and clearly should be taught in schools. I think for some of the other stuff, probably learning psychological skills would be more useful.

If we really thought about it: what’s the purpose of school now? What do we want it to be? Not just what has it been historically, but what do we want it to be? What are the important skills we want children to grow up with?

There was an attempt to teach preventative psychological skills in schools, and my understanding is that the evidence didn’t look very good in the end when they tested it — it just didn’t help very much. But that doesn’t mean it couldn’t help. It’s like, okay, that particular program didn’t seem to pay off, but I do think there are psychological skills that could help you a lot.

It might be a challenge to figure out how to work them into the curriculum: what skills exactly to teach, how to teach them, and how do you make them stick? Because really the purpose is getting people using them in their everyday life. That’s actually one of the big challenges with school in general — a lot of what people learn, because they don’t have an immediate use for it in life, they just forget by the next semester, if not the next week. So I think that would be a challenge: if you’re teaching psychological skills in school, how do you get them to stick, and how do you get people using them?

Ariel (53:34): I think in a way schools used to do more. Traditionally, if you go back a hundred-whatever years, they did try to teach character more, and I feel like that faded over time. More recently they’ve been doing a little social-emotional learning, but I don’t think they do enough to really let people learn. I think a lot of it would need to be done in a more real-world scenario — in a classroom it just would be useless, and schools aren’t good at figuring out how to let people practice things in the real world. Maybe that’s an issue in general: it’s hard to teach someone a real skill when for real skills you need the real world.

Spencer (54:10): Yeah, schools are not always the best at that. You get all these funny things in math class, like, “this is Joe’s favorite triangle, figure out the length of the missing side.” And you’re like, what are we talking about here? So yeah, it doesn’t always do the best at making it practical. But I think some teachers do do a good job at that — maybe that’s just not the default.

Ariel (54:32): Besides your book, is there one or two other books you would recommend to people if they wanted to improve, and that you just think are very helpful?

Spencer (54:45): One book that I love, one of my early introductions to psychology — it’s for people with depression in particular — is Feeling Good by David Burns. I think that’s a lovely book.

I also think if you just really want to deep-dive on habits, then Atomic Habits is a great place to start. It’s a very easy read, it’s very popular, and it has a lot of good, basic information on habits. That is a good one.

Ariel (55:05): I remember David Burns was on your podcast, and I found it very interesting. There might have been certain things I disagreed with here or there, but he clearly cared so much about improving, and always asking his patients how he could improve. He had so much compassion that he showed. So I thought that was a very interesting episode.

Spencer (55:24): Yeah. And for anyone listening, if you want to check that out, just search for the Clearer Thinking podcast and David Burns. I’ve actually done two episodes with him: one on how to have better relationships, and one on going beyond cognitive behavioral therapy — what are his newer views on how to do even better and more effective help than cognitive behavioral therapy for people with psychological challenges.

He’s a fascinating person, and he actually popularized cognitive behavioral therapy for the general public. He wrote one of the first books — Feeling Good, the one I mentioned — that really made people aware of it, and I believe it sold millions of copies, if I’m not mistaken.

Ariel (55:59): We mentioned some of your resources, but what do you want listeners to remember? Where can they find you, and what should they check out afterwards?

Spencer (56:09): If you found this conversation interesting, we’d love for you to check out the Clearer Thinking podcast. We also have over 90 free tools you can use on our website, ClearerThinking.org. And we send out a weekly newsletter, usually about psychology, that a few hundred thousand people subscribe to — you might find that interesting. So yeah, I’d love it if you check that out.

Ariel (56:27): Okay, thanks so much.

Spencer (56:29): Ariel, thanks so much for having me.

Check it out on Spotify, Apple or Youtube.

Transcript

Below is a lightly edited transcript, courtesy of AI:

Ariel (00:05) Joining me today is Dr. Shaanan Cohney, a computer scientist, legal scholar, and senior lecturer at the University of Melbourne. He’s won awards in research, teaching, and hacking, and has advised both the FTC and the US Senate on tech policy. Shaanan also has experience cracking safes, singing competitively, and playing piano. Shaanan, welcome to the Zappable Podcast.

Shaanan Thanks, it’s lovely to be here this morning.

Ariel So Shaanan, since you have experience in both music and computer science, is there anything in common that draws you to both of them?

Shaanan Yeah, and I think some of the connections actually run pretty deep. Everyone talks about how music is very mathematical, but not often enough do people explicate what they mean by that.

There are a couple of popular music theoreticians on YouTube who go into this a bit more. One I came across a couple weeks ago charted out a beautiful graph of what music naturally is going to sound good to human ears. And one of the things there is that depending on the environment in which you’re raised, different music sounds good. If you’re raised in Bali, gong music might sound really great to you — that operates on a different tuning system. If you’re raised in the West, classical Western music has a certain sound to it. These are all built on mathematical relationships, and depending on the country you grow up in, or the musical culture, this is going to be different.

But it turns out there are these super interesting mathematical relationships between the music that’s likely to sound good and the music that’s not. And you can do these music-plus-code visualizations that are really awesome. There was a 1993 paper by William Sethares — I’m not sure how to pronounce his name — around the relationship between psychoacoustics, which is the study of how humans perceive music, and the mathematical nature of music. For me, that ties very nicely into an analytical way of viewing the world.

At the same time, there’s the flip side of music, which is what most people think about: the more emotional, cultural, social experience of it. And I think code has something similar to that as well, which we often lose sight of — software development is also a social practice, or at least it has been up until recently.

Everything from the early days of Unix hackers working together over bulletin boards, to the notion of pair programming — there’s a sense in which coding is a craft and an art. There’s Donald Knuth’s famous book, The Art of Computer Programming. So there are interesting parallels between the way programmers approach the craft of writing code and the way musicians think about the craft of writing music.

Now that’s breaking down on both sides a little bit with AI, but those links are something I appreciate. Even when I’m actually doing music, you try to get into a flow state, and just like with code, when you’re in the flow state you’re not necessarily thinking about all of those aspects at the same time. So for me, they’re two sides of a similar coin.

Ariel Interesting. And since you mentioned how AI is affecting things — do you think AI-generated music is valid music? What are your thoughts about that?

Shaanan This is where it becomes really important to separate out music as a product from music as something more social, more artistic, more craftworthy. There was a beautiful video essay by Adam Neely, a popular online music theoretician, called “Suno, AI Music, and the Bad Future.” He brought in a number of sources, one of which is one of my favorite ways to think about music: musicking, an idea from Christopher Small, an ethnomusicologist. That’s the idea that music is more than an output — which is how we’ve been treating it for the last century with recorded music. Music is a social practice. That’s everything from listening to music, where you have to be actively engaged, relating it to your own culture, often listening with other people — there’s a lot of social context there — and also music creation and music performance.

Up until the last century or two, a lot of people’s experience of music was either going to the big concert hall when someone came into town, or much more commonly, making music in the community. Whether that’s liturgical music in your church or synagogue or mosque, or just banging on things while singing, or old slave songs — all of these were cultural practices.

What’s happened, particularly in the last 50 or 60 years, is that link has been increasingly broken, and people see music much more as the product rather than as the social and cultural experience. Some of this AI-generated music takes that to the far extreme. Is it music in the sense of music-the-product? Sure. But I think something important is missing when you take away the cultural and social experience and the element of craft. There’s also a lot more I could say about the process of creation, if that’s interesting to you and your audience.

Ariel Right. I guess if people want something to share with their friends, or to go to a concert together, they want something more unique — they want to know that people struggled to make this song. But if they’re just listening to music in the background while they’re running, then probably they don’t care as much, and AI-generated music could easily meet that need.

Shaanan It could. A lot of AI-generated music, at least to my taste, is just bad — and I mean that from a somewhat subjective but critical lens. If you look at the construction of AI music, there are often things missing. The lyrics can be very repetitive and not have actual depth while having the semblance of depth. That’s not because in theory AI could never generate meaningful lyrics, but from an objective standpoint, today the lyrics are rather trite and the rhyming structures are fairly simple.

One of the reasons people really like Eminem is not because the harmonic construction of the music is that amazing, but because he does an amazing job with rhyme and lyric. Even in just that one small aspect of AI music, it’s really clear if you pay attention for more than a few seconds that it’s not there. So I find it annoying to listen to.

But that is in some sense a matter of aesthetic preference. Perhaps because I’ve trained in music, I have more of a gut reaction. Do I think some people could obtain some utility from it? Sure. But at least for now, something is missing — both on the philosophical, social side, and on the question of what the actual product in front of us is.

Ariel On the technical level, I feel like they’ll get there soon enough, or maybe people just need to work on their prompting a little bit, because they’ve gotten so good at all these kinds of areas. Maybe we’ll see in a few months.

Shaanan Maybe. There was AI-generated Bach, which was pre-generative AI — purpose-built models designed to mimic the harmonic structures of Johann Sebastian Bach’s music, in particular his fugues, to the point where experts couldn’t really tell whether the composition was Bach or not.

But there still seems to be something missing, at least with the generative tools we’re creating today. How soon we get there, I don’t know. Will I be able to tell at some point? I don’t know.

Ariel (07:47) So turning to another topic: you have extensive experience with software security, but also a little bit on the physical side, even breaking into locks. Is there anything in common between those two areas, or are they just separate areas of interest for you?

Shaanan This goes back to a fairly deep connection between physical security and computer security. I wouldn’t even say I do software security, because to me that’s looking at far too narrow a part of the system. The way a really good red team, or a really good security professional, looks at things is as an entire system.

Locks are a really nice example of this, because you can look at just a very specific element — the theoretical construction of the lock and the matching to a specific key. There’s a certain key space depending on how many pins there are; you can figure out how many different possible keys there are, and that’s all well and good. But to actually build this thing, to instantiate it in the real world, there are a lot of considerations you need to take into account. And this is where lock pickers think about locks in a similar way to the way computer security people think about computer systems.

When you make a real lock, there are machining tolerances, and those come into play when you’re lock picking. You’re relying on the fact that not only are certain combinations of pins ineffective in a real-world lock, but the quality of the machining of the lock itself leaves imperfections you can exploit as a lock picker.

Now, a really good physical security person also doesn’t stop at the lock. The lock is within a door, the door has a latch mechanism. Perhaps you don’t actually have to pick the lock — you can bypass it by toying with the latch. Or if you’re a common criminal, you go through the window and ignore the lock in the first place. These same parallels carry over quite closely to computer security. We talk a lot about attacker-defender asymmetry: the defender needs to ensure the entire system is secure, whereas the attacker only needs to find the weakest point. And that weakest point, in our lock analogy, might be the pins, it might be the pinning, it might be the jamb of the door, it might be the window that’s been left open. So that parallel is a really good way to teach people how to apply systems thinking.

There are also other things we can learn from locks and lockpicking. Firstly, there’s a cultural aspect from the ‘70s and ‘80s around the way hackers thought about experimentation and gaining knowledge. The same tools hackers applied to locks reflect the exact same philosophy they applied to early computer systems.

There was great work by one of my mentors, Matt Blaze, on master keying — the notion that in a large corporate building you can have a single key that unlocks every door, called the master key. Then you can also have subkeys that unlock a subset of doors, or maybe even only a single door. In Matt Blaze’s paper, he demonstrated a physical privilege escalation attack. Privilege escalation is when you have some normal level of access and can use it to gain administrator or privileged access. He showed that even with a normal subkey on a master-keyed system, you can escalate that to get the actual master key on certain kinds of master-keyed systems.

He copped a lot of flack for this, because the locksmiths’ guilds were not happy that this information was now public, and he received threats of all sorts, including death threats. But this parallels what people in computer security say: you shouldn’t rely on the secrecy of the design of a system in order to protect the system itself. This is Kerckhoffs’s principle. So this was a real physical analog of a principle that applies very directly in digital systems.

You’ve got all of these different individual lessons you can learn. You learn about the culture of hacking, you learn about particular types of attacks and ways of thinking about things, you learn about the principles. So I think it’s a skill every computer security person should try their hand at, at least for a few minutes.

Ariel Have you ever used your lockpicking skills in practice in the real world?

Shaanan Within the past two weeks, yes — but no comment. This gets me into trouble quite a lot. I’m very good at losing keys, forgetting keys, locking myself out of buildings, and so on. So having skills in physical security has come in handy more than once.

The most prominent of these was when I was a graduate student and I locked myself out of the dorm. At that point it wasn’t a lockpicking attack, but similar physical security — I came up with a way of getting back into the building by bypassing the security systems on the dorms, and that resulted in more than a little trouble for me, shall we say. But it has proven a useful practical skill lots of times as well.

Ariel So usually you’re breaking into areas that you should have access to, you’re just missing your key.

Shaanan Exactly. That’s how I stay on the right side of the law. I break into things, but they’re things I should have access to and for some reason don’t.

Ariel One interesting difference between physical security and software security: with software security, once you’re on the internet, if there’s a vulnerability, something will find it. You’re open to the whole world. Whereas with physical security, there aren’t so many people trying to physically break into buildings, so once you reach a certain level, it’s usually good enough.

Shaanan I think that’s true to a certain extent of software security, but this tells us a bit about the role locks actually play in life. Picking a lock is not actually that hard. In particular, the locks on most people’s houses — when you imagine the American white picket fence house with the front door — those locks are typically not very secure. It would likely take me under 30 seconds to bypass or pick any given house lock.

But the point isn’t to stop people from getting in with some high security mechanism. It’s a social deterrent. It’s just like the way you might leave your bedroom door closed when you have friends over to an apartment. It’s not that that provides any great deal of security, or that if they looked through the door they’d see anything so terrible — depending on how messy your bedroom is. It’s more a social sign about where the boundary is being set. If someone wants to break into your house, they can, most likely by tailgating someone through the security door, kicking the door down, or bashing in a window. It’s more that it adds a little bit of friction and a little bit of social signaling.

This does happen in software security as well — not quite as much with the signaling, but to the extent that adding a bit of friction makes you a slightly less soft target, so you’re no longer a target of opportunity. A lot of crime — not all of it, there certainly are more sophisticated attacks — happens in moments of opportunity, where someone sees that a door has been left open, or a car has its lights on and has been left outside with the key still in the ignition. So the function of both your firewall and your basic security hygiene, and the lock on your home, is not necessarily to produce a technical impediment that a nation state couldn’t bypass. It’s more to provide that little bit of friction and that little bit of social signal.

Ariel (15:23) Now turning to your experience teaching computer science. How have you noticed it’s changed recently, as generative AI has become so common and powerful?

Shaanan This is the doom and gloom part. Most of the people teaching introductory classes — and I count myself among them — are in a state of mild depression and slight panic, except the people who are close enough to retiring that they don’t have to worry.

What we’ve seen is that the software tasks we would traditionally give in an intro programming class — go home, work on this assignment for two weeks, skill up a little, learn basic programming paradigms and abstraction — are no longer actually completed by our students. Or to the extent that they are, it’s a student prompting an agent once and then copying and pasting the output into our submission system. One of my TAs estimated that somewhere between 70 and 90% of the submissions he graded had clear hallmarks of AI use.

To give you a sense of how he picked up on that: I teach an introduction to C programming — data structures and algorithms through the lens of C. In a typical year, we’d teach some of the more interesting C keywords, stuff like static and volatile. I might mention those, but I’d almost never see a student use them in an assignment, because that level of sophistication usually comes with experience — knowing about variable lifetimes in enough detail to think about how to manage them. Typically students would have messy variable names and imperfect structure. And yet assignment after assignment used sophisticated constructs and paradigms in C that I’d never seen previously in my career. That’s what my TA picked up on too, and it flags to us that the students are no longer completing the work.

So there’s frustration on that side, because as much as we communicate that the only way to develop these skills is by doing the work, students for whatever reason don’t do it. And they all have legitimate reasons: life is stressful, this is more effective, am I ever going to actually need to program? But I do still think we’re missing something by not teaching them how to think like a computer scientist.

Then there’s the other view, which is: well, maybe thinking like a computer scientist needs to change. Maybe you don’t need to think about code itself, because I think all of us are slowly coming around to the reality — some faster than others — that humans probably won’t be needed to write very much code in the future. I don’t write very much code now, and previously that was something I really enjoyed as an artistic or craftful endeavor: writing the perfect little function that did an algorithm in much more efficient time. Now I do what everyone else does and prompt the machines.

So this is causing a moment of self-reflection and inner turmoil around what our role is as educators, and what we should be teaching. At the introductory level, things have really changed. For higher-level courses, once students have attained a baseline of skills and you’re doing things in a more sophisticated manner, or teaching pure theory, I think there’s still a little more room for hope — and I’m happy to chat more about that if it’s interesting.

Ariel Yeah, it’s an interesting question — what do you think the goal is now? Because not only can AI do small-level coding, it’s starting to be able to do complete projects. We still need a human currently prompting it, but we don’t need a person as involved in the technical details. So what should the goal be? What is the student trying to learn? What are we trying to accomplish?

Shaanan So I teach at two very different places. This fall I’m going to be teaching at Tufts, which you can think of as a liberal arts college with a research focus — it graduates doctoral students, but it’s small classes, seminar style. That’s one side. On the opposite side, I teach at the University of Melbourne, which is a large state school, and my smallest class is 400 students. The kinds of things you might think about teaching, and the way you teach, have to shift a lot.

For me, one of the things I want to communicate in my lectures is the beauty of the topic, to encourage self-motivation and self-exploration. Because coding and computer science are in many ways a craft. It’s a skill, it’s something you have to spend your own time on or you’re not going to get better at it. If you want to work producing slop applications for SMEs, AI can do that. But as an educator, that’s not my goal. What I want to do is cultivate individual growth and judgment, and try to motivate students to understand why they might pursue their own learning in the area. Because I only have them for one class — or if I’m lucky, I see them for three to four years. After that, it’s going to be their own sense of wanting to learn, their own motivation, and a level of autodidacticism that really drives them.

One of the hallmarks of a programming career, of an engineering career, is this need to constantly be staying up to date. If I can’t give that passion to them, then I think I’ve really lost out. And that was true even before AI, even when they were writing the code themselves. There were students who would pay someone overseas to write their assignment for them — contract cheating — or get a friend to help. So I think that actually hasn’t changed that much. If I keep my view on that passion, on that motivation, and on socializing them into the disciplinary norms of computer science, I think there’s still some hope.

If we’re thinking about specific skills development in class, that’s where the deeper concern is. Not only because everything is changing so quickly, but because I find it very hard to predict what useful things I can teach them now that they’ll still use in four years.

Now, college education has never been about vocational training directly. Med school and law school a little more so, but even in law school they teach you how to read appellate opinions — they don’t spend as much time teaching you how to redline contracts you’re swapping between firms. So college has always been this disconnected place where you do learn skills, but the skills you learn aren’t always directly transferable. There was a result from the last US census, I think, showing that only about 28% of people who graduated from STEM majors ended up in a STEM profession as narrowly defined. That’s probably a bit of an undercount for various reasons — they didn’t include medicine and healthcare. But at the same time, we still seem to be communicating something valuable.

So I think this is a moment for us to ask: what are those valuable things we’re communicating, that we’re teaching, that aren’t the specific thing they’re going to be doing in the day-to-day job? In that way, I think this AI moment is actually quite empowering, even as it’s depressing at the same time. Because we have a real forced opportunity to go back and think: okay, what are we actually doing here? And I think we can make some good of that.

Ariel It’s interesting. Schools do say they’re not trying to teach for a vocation, but I feel like for a lot of students that’s why they’re going to school — either to learn practical skills or to get the credential they need. So it’s interesting that computer science as a major was never entirely practically focused; there are parts of math that usually aren’t needed on a typical job. But it did seem overall like, okay, this is a practical field. And now the areas that seemed so practical, like learning how to code — it’s a question whether that’s even going to be used. So it’s becoming a theoretical topic that we’re teaching for its own sake. I’m wondering, could you formulate what the purpose is, if it’s not as practical as it was?

Shaanan I think there are a few things. The first is that we have to be realistic and recognize that for many people, college was not required to learn computer science. You could go online and watch CS50 — I’ve taught with them before, that’s wonderful — and you can actually become a damn competent computer scientist just by consuming online materials. So to me it was obvious that something more is happening here.

One aspect could be credentialing. As you said, a degree from Berkeley is going to open doors for you that a few online Coursera courses won’t. To the extent that what we’re doing is credentialing, it’ll probably still function much the same. What’s interesting is that it further weakens the link between what employers claim the credential communicates to them versus what it’s actually doing.

Being realistic about it, we’re pretty aware now that a lot of what happens is a filtering process that happens pre-admissions. The university is already taking students who have jumped through certain hoops, and then attending for four years is not actually about skill development — it’s about certifying that they jumped through that initial pre-admission hoop. That’s not a type of education I particularly care about.

If I were to take the extreme position, I’d say let’s stop worrying about assessment in the university whatsoever, because the credential was never tied to the assessment in the first place. It was tied to their high school ability to jump through hoops. Now, this does put more of a burden on future employers, who’d have to do even more weeding if we’re not assessing at all — and maybe more burden on the poor college students, who now have to work even harder to prove themselves earlier in life and can’t distinguish themselves in college. But I do think de-emphasizing assessment is one path to helping us care more about some of the things education can do when we’re not so focused on hoop jumping.

One thing I think we can and should do is take advantage of the way AI can give us efficiencies in our teaching, and then allow us to provide much more personalized attention to students, focusing on individual goals, individual growth, and individual interests. I have students in my intro to programming class where all they want to do in life is make things. For those students, giving them tools and helping them put themselves in an environment where they learn how to use what’s out there to build the projects of their dreams could be very empowering. On the other hand, I have students who are much more theory-driven, who couldn’t care less about building a distributed system but really want to understand the nature of computation. For them also, I can’t serve them so well at the moment. In a class of 400 or 500 students, I have to cater to a middle path, which up until now has been this forced credentialing function.

So maybe if I can offload some of the rote stuff to AI, I can spend a little more time tailoring my efforts to these groups of students more individually, and focus on the things they really care about — the things that can’t be delivered by someone who doesn’t have time to interact with students directly.

Ariel That’s interesting. On its own, I would ask: if you get rid of the hoops, what percentage of kids really want to just learn for its own sake? But if you tailor it that closely — so those who want to build projects get to focus on that, those who want theory get to focus on that — I think it would be a higher percentage. So it would be interesting to see what could be done.

Shaanan Yeah. On how many students want to learn for its own sake, there’s a big difference between different sides of the institutional spectrum. You’ve got the super motivated high school students who are already on Coursera and have finished four courses before they even come to college. You have community colleges in the US, and technical colleges, with students who are really there for a clear vocational reason. You have liberal arts colleges where they’re very interested in that seminar-style stuff. Then mid-tier state schools, where you get a whole mix — some very motivated, others who just want to party for a couple of years. And then your prestige, Ivy Plus, R1 institutions. To lump all of those people together elides really important nuances about people’s desire to learn and what they actually want out of learning.

There was a 2016 survey of incoming freshmen into, I think, only four-year colleges. That survey showed that 84% of students said learning things that were important to them was a very important reason for attending college — it was one of those Likert scales. But then you ask those same students about the other items, and 76% said gaining a general education was very important. Even more than that, 80% cited getting a better job, and 74% said career-specific training. So you can have these multifaceted reasons for wanting to be part of college. And that’s on the expressed preferences side. As you pointed out, the revealed preferences are somewhat different.

There was another study by Philip Babcock, a social scientist, who found that the amount of time students spend actually doing their college work halved between 1961 and 2003. So the extent to which we’re successful in getting students to learn for its own sake seems to be less than we might hope. Maybe the suggestion we were both building on — that with closer tailoring we could up those numbers and get back to a point where students are really self-motivated — is a possibility. I’m hopeful. At least I have to be optimistic, or else I’d need to change careers.

Ariel It’s challenging, because probably a big part of the change over those decades is just more people going to college and university. So it could be that people who wouldn’t have gone in the past aren’t as interested in learning for its own sake.

Shaanan Yeah, and particularly the GI Bill shifted the college demographic substantially and led to this professionalization race, or academic race, where even to get entry-level jobs you now need a college degree. So for a lot of people who previously wouldn’t have had the motivation or interest to go to college, they now have a reason why they need to go. I do think that’s been, on net, not beneficial — I mean, more people going to college is net beneficial, but I think we’ve overdone it, in the sense that technical and vocational colleges have a really important role to play, and we shouldn’t be forcing people through four-year college — particularly at places with a research focus — when those individuals just want to get out there and start working.

Ariel Do you think countries like Australia might do a better job than America does?

Shaanan We’ve made some very bad mistakes. We essentially eliminated the technical colleges in a series of reforms in the ‘80s and merged them with the research universities. What that’s led to is a number of universities that have to build themselves as research universities, and from which the government expects research outputs, but they don’t necessarily have the talent to compete on the international stage and produce the quality of research that really benefits the country.

So number one, we underinvest in research compared to the US. On all non-medical research, we only spend about $800 million a year in government funding. But then the pie gets split into places that would really better serve the public in their original form as technical colleges. That’s not to say no good research comes out of these places, but the ratio is definitely different. They were designed as technical colleges, and trying to force people into the research-intensive route — this happened in the US, this happened in Australia — I don’t think does a service to anyone.

Ariel (31:46) Okay, now turning from the effects of AI on college specifically: what do you think the effects of social media previously, and AI as the upcoming trend, have been on teenagers and college students?

Shaanan I think we’re still a little early on the AI front to really know. Social media is a little better studied. In particular, there was a consensus report of the National Academies. For those unfamiliar: if you’re a scientist and you’ve achieved real preeminence in your field, you may be elected a member of the National Academy of Engineering or the National Academy of Sciences. Now and then they commission reports that attempt to ascertain the consensus in a field. These are really the gold standard for understanding what researchers think.

One came out just a couple of years ago on the impact of social media on youth and adolescent well-being, and the results were essentially inconclusive. They found a small correlation between adolescent depression, anxiety, and social media use, with a very unclear causal direction. I’ve got a line here from the actual report: “the committee’s review of the literature did not support the conclusion that social media causes changes in adolescent health at the population level. Nevertheless, there are harms…”

I think we see this overplayed in the media a lot. It’s a very popular line to say social media has ruined our kids, that it’s net harmful. There is some evidence in that direction, but certainly not enough to argue conclusively that social media is net harmful rather than net beneficial. There are also plenty of studies showing beneficial effects — ease of communication among teens, people who can access community that they couldn’t previously, LGBT youth in more insular communities that previously would have restricted that. It’s easy to say those pale in comparison to the harms — child grooming, depression — but the evidence really isn’t there to argue conclusively that it’s been harmful. It is very heavily disputed, though, and there are certainly individual scientists who take a fairly strong view one way or another.

One thing that is fairly clear is that technology on the whole has displaced other childhood pastimes, and you could argue that in itself is a problem — the fact that kids aren’t going out and spending time outside because of TV, computers, video games, social media, maybe AI. Maybe they’re staying at home talking to ChatGPT. But there definitely is that displacement effect. So to the extent that you have a strong prior that the things kids were doing prior to 1940 were preferable, then perhaps you’d have a reason to dislike not just social media but all of these different technologies. And again, we’ve seen this pattern with every new large childhood pastime — everything from television to newspapers was originally heralded as ruining the youth. That doesn’t tell us whether this time is actually the real case, whether there is a real wolf, but there is reason to tread a little more cautiously than the media likes to.

Now, with respect to AI and social media use in people who already have problems, I think there are much clearer links. There are patterns that are fairly apparent prior to children engaging in self-harm, where often there’s a spike in social media use. But this is also correlated with different hours of sleep and wakefulness — a lack of sleep and staying awake in the middle of the night is correlated with negative mental health outcomes. Again, it’s unclear which direction the causal arrow points. Are they up in the middle of the night fiddling on their phone because they’re depressed and can’t sleep, because they have negative self-image or anxiety? Or is it the fact that they’re up scrolling through social media at 3 a.m. that’s making them depressed? Or perhaps some of both.

So I think there’s a lot of room to explore that more, but I would be really hesitant to make some of the strong policy decisions that governments are making on the basis of believing this causal link has already been established. I think the evidence just isn’t there yet.

Ariel I guess the question is: there’s been an uptick in certain negative metrics, like suicide among teenagers, since 2012. People say that’s when the smartphone and Instagram and things like that became popular. So if those aren’t a strong enough cause, then you need to see what else might have caused it. Is it that people are overprotective of their children and they don’t get to play? Is it other factors? Is it just too complicated to resolve? And then people look at social media and certain aspects of it seem very negative — unrealistic beauty standards, people spending so many hours a day — so they feel like that must be the cause.

Shaanan Yeah. And look, if you’re a parent deciding for your child — I want to prevent my child from going on social media until a little later, I want to speak to them very carefully and keep watch on what they’re doing — I think all of those are potentially positive things. But that’s at an individual level rather than a population level.

Where I think there’s a lot more evidence is that for children in certain situations, social media use is going to be either net positive or net negative, and you can often tell based on contextual information about their situation. Parents are often well placed to make that judgment — assuming we’ve got good parents, which is often but not always the case. The problem comes when you try to do population-level interventions.

One place we’ve seen a similar-ish intervention was the phone ban in New York schools. There was an economics paper showing that there actually wasn’t a substantial impact of the phone ban on educational outcomes, at least as far as they were able to measure. It was a relatively good study. But there was also follow-up by some of the authors and related persons saying that just because it doesn’t improve academic outcomes doesn’t mean we should let phones run amok in schools willy-nilly. There may be other benefits — just in terms of classroom control, it’s easier for a teacher to control the classroom when kids aren’t all on their phones. There may be other social benefits.

So again, my line isn’t that there is no worthwhile intervention around technology and adolescence. It’s more that we have to be very self-aware about which interventions are directly supported by the evidence and which are more of an experimental game — or which should be tailored less to a population level and more to a smaller context. That could be an individual school, a community, a classroom, or a family.

Ariel I get that it’s a bit extreme for the government to do an outright ban on a whole country. On the other hand, the challenge is that if a parent wants to ban it for their kids but all the other kids use it, that creates a weird social dynamic. So it’s very hard to do on an individual level. But if there were some way to do it on a community level, that seems like a good balance — people could leave the community if they wanted, and you could also run experiments comparing communities. So if they had some way to roll that out, maybe by school or district, that seems like it would be better than doing it across the whole country.

Shaanan Yeah, I’ve actually done some research in this area. Australia rolled out a social media ban to the whole country for anyone under 16. The way they chose to do this was by regulating the social media platforms: if you are a social media platform, you have to take reasonable steps to remove under-16s and to ensure they’re not there.

This regulatory intervention has now been in place for about six months, maybe a little over, so there’s been time for some initial results. I have a couple of papers on this, and there was also a paper by Cass Sunstein and some of his co-authors — I’ve forgotten who the first author was. Essentially they found the ban was unsuccessful, and you still got the disparate impact you were talking about, where some kids are still on social media. Those kids self-identify the other kids who are not on social media as the uncool kids. So rather than solving the bullying problem, it may have exacerbated it.

This is also true in my research, in which we did focus groups with young people and found that essentially all of them knew how to get around the ban if they wanted to. The thing that was much more determinative of whether the ban worked was what the parents wanted to do with their own kids.

Now, the Australian Prime Minister, Anthony Albanese, has openly said that the motivation for this is to give parents ammunition to police their own kids. But as the Sunstein paper showed, unless you achieve a critical mass of kids off social media — whether because families have chosen to enforce it or because it’s become technically impossible — unless you get to that tipping point, it’s not really going to be effective as a social intervention.

So I think that’s more or less where it’s landed. Even the regulator in charge of administering the social media ban has said she doesn’t have adequate power to actually enforce it, and has cautiously and subtly implied that it’s not really working as designed.

Ariel I guess one thing proponents of such bans could say in response is that you just need stronger controls — make it even harder to get an account if you don’t show stronger proof that you’re a certain age and that you’re actually that person. What would you say in response to that?

Shaanan I think the more intense you make the restrictions, the more you start to impinge on civil liberties. Requiring everyone to present their ID when visiting a website is probably a non-starter for people in America, or at least from my understanding of American culture. Even in Australia, people are very uncomfortable with it, and it did not have much political buy-in. So there’s an explicit carve-out in the law saying that platforms that choose to use government-provided ID as a way to ascertain age are not permitted for that to be the only way to establish that you’re over age.

What’s happened instead is that they’re scanning people’s faces and trying to guess how old you are based on a scan. I have some forthcoming research on this that I unfortunately can’t spoil, but with all the methods we tried, we were at least able to bypass them some of the time through the front door, just by messing up your face or putting on a disguise. The models are getting better, so maybe that will stop.

But more significantly, pretty much everyone knows how to use a VPN, even 12- and 13-year-olds. So unless this ban is regulated globally and you start cracking down on VPNs, you’re going to have a really hard time enforcing this at a technical level without introducing internet-level censorship. The UK has been floating this idea in the last couple of months — there have been a couple of members of the House of Lords who have endorsed a crackdown on VPNs. But there are a number of reasons we should feel very wary about any such proposals, because that’s going to have very significant knock-on effects. Given that the harms of social media are already somewhat disputed on net, the benefit of imposing a countrywide anti-VPN and internet censorship system is not something I would endorse. Quite the opposite.

Ariel I guess the question is similar to the physical locks you mentioned before. Maybe you just need a certain level of difficulty, and then it’s not expected that you’d be on social media. I don’t know if that’s true — maybe it’s so trivial that kids will do it no matter what. But once you have that certain level of difficulty, it doesn’t have the same social pressure. And then if some people want to get around it, that’s okay. We’ve weakened the expectation that everyone is on it, and maybe that’s enough.

Shaanan There’s a good point in parallel to that, which is: maybe it’s not working because today’s 13-year-olds have 14-year-old friends, and the 14-year-olds have 15-year-old friends, and the 15-year-olds were already on social media before the ban. So perhaps it’s just a generational thing.

The last paper I published on this suggested that perhaps that’s not true. The reason we’re a little skeptical is that we found cross-age-group socialization of the bypass methods. It’s not just the 15-year-olds sharing with each other; it’s groups of kids across year levels — particularly because of sibling dynamics — sharing both why they use social media, what social media is, and why they might want to get around the ban. So it’s unclear to what extent a generational shift will change it. It’s still plausible, because even that socialization mechanism can change as the generation changes. Perhaps there will be a discontinuity in the way today’s eight-year-olds relate to the 12-year-olds when they all age up and the 12-year-olds hit 16. But at least the current indications aren’t positive for that.

Ariel I guess another way to approach these issues is, instead of trying to restrict it by age, to reduce the negative effects that happen for all ages. So if they have algorithms pushing certain content just to get people to click, but it makes them more depressed, maybe they could do something to reduce that and try to have some kind of positive algorithm boost. I don’t know if that’s something you think might be more realistic.

Shaanan Yeah, there I’m much more bullish on our ability to intervene.

The large social media platforms have been pretty terrible at transparency around their trust and safety programs — that’s the name given to the programs designed to prevent harm to young people, scams, uploading of pornography that’s not allowed on the site, or whatever it happens to be. Those have been very closed to the public. There are some conferences where occasionally practitioners from the companies will discuss things a little, but they’re largely black boxes, and in recent years they’ve become more and more opaque. There’s a well-known story in which Facebook gave a permanent ban to some of my collaborators who were doing trust and safety research on the platform, on the grounds that they had violated the terms of service. Contractually, they are able to do that; they didn’t violate the law by blocking these researchers.

But my view is that normatively we should have rules in place to ensure that researchers, public advocacy organizations, and governments can get more insight, and might not necessarily be obligated to follow overly onerous terms of service that prevent us from getting more information. So the first step is more transparency. Transparency alone is not going to solve this, but enabling it will allow regulators, civil society organizations, and the academics who support the regulators to figure out what kinds of interventions might actually work. You can still design experiments to determine what interventions might work even without support from the companies, but it’s definitely a lot harder.

On the next front, you have things like the Age Appropriate Design Code, which I think is in the EU, which aims to set out what design elements and features should and shouldn’t be in place for a platform accessible to a young person, or what experience a young person observes when identified as having an account on a platform. Australia has gone in a slightly different direction with a digital duty of care, which attempts to impose a generalized legal obligation on the platforms to look after young users. The platforms, seeing the writing on the wall, have not been entirely opposed to either of these efforts.

So there are a lot of things we can do in this direction. Which of these will actually pan out and deliver better outcomes, and ensure we can capture some of the gains without the negative effects? I’m not sure yet. There’s a lot more work that needs to be done. But I broadly think it’s worth experimenting with these different regulatory interventions to determine which might actually work out for us as a society.

Ariel (48:20) It seems like there’s a balance. On one hand, trying to ban everything seems too extreme. On the other hand, some of the habits people have on this very new technology are so unnatural that to just do nothing and say “yeah, whatever” also seems like maybe not the right response. But if they could figure out something more balanced, maybe they could help everyone.

Okay, turning now back to software security. Recently Anthropic released a model called Mythos, which was one of these new moments where AI showed, in a practical way, that it was more powerful than anything before at finding software vulnerabilities. They released that to a restricted group of companies. Then they released another version, Fable, which was supposed to be the safe version for people to use in general — where if they were trying to do things that seem dangerous, like find security exploits, it would block them.

So that’s the basic background. And then the government briefly shut down access. They said that foreigners outside the US, or even foreigners in the US, can’t be allowed access even to Fable. What do you think of the overall risk here shown by this new AI model, and what do you think of the government’s initial response?

Shaanan Maybe let’s start with how dangerous these things actually are. There you have two sub-questions: dangerous to whom and how, but also what is the marginal danger over the other models that are available, including the open source ones.

What everyone in industry is talking about, at least, is a vulnerability tsunami — the notion that these tools introduce significant new cybersecurity risk because of the ease with which they can find latent security flaws, generally in code but also sometimes in systems. More commonly it’s in a code base: they can find something like a buffer overflow and then even write a proof-of-concept exploit for it.

We’re not necessarily at the point where they can come up with a full exploit chain — find a buffer overflow in the JavaScript engine in Chrome, use that to escape the sandbox, then do privilege escalation and get to the operating system. We’re not quite there yet. But we’re certainly at the point where the models are able to discover meaningful software security flaws, and with a proper harness, a proper set of supporting tools, distinguish between false positives and true positives. There’s still a level of over-reporting, but it’s gotten to the point where if you get a few outputs, you can be reasonably confident — assuming you’ve used the right harness — that these are actual real bugs.

So they’re very good at the first part of that chain: the discovery, the analysis, and then the verification. They’re not as good at actually launching full-scale cyber attacks. We’re not seeing, at least to my knowledge, end-to-end attacks by swarms of Mythos bots out there in the wild.

There is some danger, because our ability to patch is now a little slower than our ability to find vulnerabilities. To patch, you have to actually be awake, know there’s a vulnerability, and send your agent to go fix it. Meanwhile, the attackers can be off finding new bugs. And if you were to ask a Mythos-class model to find a vulnerability yourself, there’s no guarantee you’ll find the same one an attacker has already found. So being proactive at the moment doesn’t necessarily save you.

These capabilities are pretty much on the frontier. Mythos and GPT-5.6 Sol, the new OpenAI model released this past week, are mostly on par, and they can both do this. But the open models and the Chinese competitors are honestly not that far behind. This is where I think we really have to be cautious about our policy response, because our policy response is only going to be effective for a certain amount of time. These capabilities are pretty accessible.

There’s also a big question about short-term impacts versus the equilibrium we reach. A policy response that’s good for the short term might not be pertinent to the equilibrium. There was a talk by Nicholas Carlini where he talked about the commoditization of this pretty specialized skill set of vulnerability research. Other observers in the media have said essentially the same thing: what was once the purview of a few people with very specialized reverse engineering skills — myself included — can now be done by almost anyone.

I think that’s where some of the danger is, and there’s a lesson to learn there not just about the danger from cyber attacks, but also about CBRN — chemical, biological, radiological, and nuclear. My personal view, as someone who’s been in security for a long time, is that the reason we don’t see large-scale bad things in the world as often as we might is not because doing bad things is actually that hard. Assembling a dirty bomb is easier than you would hope. If you don’t care about the welfare of the person doing the synthesis, you can probably synthesize some pretty bad things. We know organized crime is pretty good at synthesizing drugs, and explosives are, again, not that technically hard to make.

Some of the reason people haven’t done it is that outside of state actors, the set of people who have the motivation to do terrible things in the world, and the technical knowledge and capabilities to pull it off, and the operational capacity to actually organize it — not only wanting to do it and having the technical skills, but being able to execute on a plan at a larger scale than just building something in a lab — has been very limited. You’re talking about a fairly small set of people at the intersection of three relatively rare groups. Only certain people are very operationally sophisticated, only certain people are good at chemical synthesis, and only a certain group of people really want to do terrible things in the world.

What I’m worried about with AI is that it may be removing some of these distinctions. By reducing the technical barrier and also making operational planning a little easier, it can increase the accessibility of some of these dangers. Even while I don’t think the set of people who wish to do terrible things is going to expand that significantly — it will a little, because of the reduced friction — I think there’s a concern that the barrier for the existing people we’re worried about has been lowered. And we’re seeing this. There was a report out of an Oxford University center about, I think, Somali groups planning piracy and terrorism operations using frontier models.

So that’s the place where I’m really worried, and where I think the government does have reason to worry. I think it’s a different question whether export restrictions and the Defense Production Act are the right way to regulate those.

Ariel So if you think there is a big concern here, but the measures they took were maybe too extreme or just not correct, what do you think is a better approach?

Shaanan Let’s start with what they did. Part of the problem is that they approached Anthropic as if they were the only entity that already had these capabilities, and that was just plainly untrue. OpenAI’s models — even 5.5, the one from the prior release — were fairly competent at finding these things. Sure, it didn’t have the level of hype as Mythos, and perhaps wasn’t quite as effective, but the marginal difference in capabilities wasn’t that great.

At the same time, there are a lot of clouds around the government’s story. Anthropic and the government had been at odds and somewhat antagonistic to each other for some time, even while the government still desired their capability. So it’s unclear to what extent the actions they took were really motivated by concerns versus how much was politically motivated.

It also seemed like this was another narrow jailbreak. It worked sometimes, only unlocked certain capabilities, only for a limited time, and would also have been visible to some extent. The thing that’s more worth worrying about is if someone can find a universal jailbreak that just turns off all the safeguards — and based on the reporting, it didn’t seem that this had actually been found. So the remedy the government chose, which was to ban Mythos itself from non-US persons, both wasn’t well tailored to the problem and ignored the fact that these other models were available, and that the particular jailbreak that had been found didn’t seem to be anything that significant.

So what can we do? This is where we have a real problem, because there’s not a great answer yet. Anthropic’s mode has seemed to be to suggest a coordinated pause — they’ve offered to cease model development if everyone else ceases model development. But this is incentive incompatible. Everyone has an incentive to defect, and not only that, but potentially to defect in secret. So I can’t see that working.

At the same time, restricting things via export controls also seems unlikely to work, particularly because the open source models really aren’t that far behind the frontier, and the know-how of what you actually have to do to improve the models is mostly public. We might see a shutting down of a lot more research — already the frontier labs have been publishing fewer of the key components needed to advance in academic and public venues. We could see that further restricted, with the government trying to push back on open research. But in the US regulatory environment, that’s very hard to do; there are fairly strong First Amendment protections around academic speech.

So I don’t have a brilliant policy prescription we can adopt today, other than to say it’s a hard problem and something we need to be working on and trying. I know that’s very unsatisfying, but I don’t have the answers today.

Ariel People have talked about doing a pause in research, but they’re talking about the more worrisome risks — like AI itself trying to take control or cause mass harm. These specific risks are a little smaller: people just using AI to do bad, at least at the level we’re at currently. It could get worse — if they get access to super viruses and things like that, that could be really scary. But for these kinds of risks right now, I feel like Anthropic’s approach might be reasonable, where you try to have multiple layers of detecting it.

And then maybe you could build on that. I know we talked about not wanting to require people to verify their identity, but in theory, if you had more knowledge of who’s using your product and what they’re using it for, you could detect bad players — either block it at the moment or detect it right afterwards so they can’t keep doing harm on an ongoing basis. That may be an approach that could work for now.

Shaanan Yeah, and that’s where we are seeing the labs doing a lot. OpenAI’s Verified for Cyber program requires you to submit some ID so they actually know who you are, and then they look at your transcripts — and they probably are running models over the transcripts, post hoc as well as live.

One of the common techniques, which I’ve also done some work on, is guard models. The idea is that before a tool call goes ahead, a second classifier or second model looks at both the tool call and the transcript to determine whether what you’re requesting should be allowed to go forward. That might be for your own safety, or because you’re stepping outside the bounds of what the frontier lab wants to permit, either for cyber risk or other reasons.

The problem is that there are still lots of attacks on the guard models, and a post-hoc approach to preventing harm — we’re going to look over the transcripts later and ban people — might still allow attacks through. There’s a big colloquial literature on people who create new accounts whenever they get banned: jailbreak it to the extent they can, use the capabilities to do their little bit, get banned, pay another 20 bucks, open another account, make a little more progress. Through that methodology they can actually assemble a fairly fearsome cyber attack on something or other, despite the existence of guard models stepping in at various points and blocking them.

There’s also a lot of sensitivity to context depletion. As you chat longer and longer with a model, the efficacy of the guard models will decrease as the context window fills up and the size of the context the guard model has to handle increases.

So a defense-in-depth approach is definitely what we’re doing currently and what we should be doing, but there are still gaps in it, and those aren’t plugged yet.

Ariel I noticed I was just playing around with Fable for a personal project — trying to track everything I read, different things like that — and it put up these warnings for no reason at all, and then it switched off Fable. It had nothing to do with anything. But I think that’s just because the government made them be even more restrictive. I don’t know if on their own Anthropic would have been that restrictive.

Shaanan Yeah, the overblocking thing is quite funny and happens to me all the time. I’ll be doing something totally benign, but because I’m a security researcher and my research papers have something to do with security, they get flagged very often. I haven’t experienced that with GPT-5.6 yet — but they’re obviously in a different regulatory position, a different political position, so they’re acting quite differently.

I think that’s one of the dynamics to watch for here: the frontier labs all sit in a different position in the political and social ecosystem, so the capabilities of the models that are permitted to users will likely end up being different. This is both beneficial for users — there’s good competition in the market — but also dangerous, because you can play the models off each other. I had an instance where I was doing some security research and GPT was getting annoyed at what I was doing, so I gave that specific task to Claude, took the results back from Claude, and put them into GPT. You could imagine building a harness that does this for you, that looks for when you get blocked by one, sends it to the other with a slightly different context, and then takes it back to the first one in a new context window.

Ariel That’s another case where, if they really knew the identity, it could help. If they see a pattern of people repeatedly doing something, they could be blocked — and if they knew their identity, it wouldn’t be as easy to just create another account. In America, if you have to give your credit card and some other verification, it’s not as easy; the credit card is attached to a name. So there are certain levels where I feel like it would be easier to control.

And there would also be a very strong incentive for the companies to control it. Anthropic has issues with distillation attacks, where companies train their models on Anthropic’s advanced models. If they could verify who’s doing that, they could shut them down — that would be very beneficial for them. I guess it gets harder in other countries for them to have that level of control, but it seems like something they’d want to do anyway.

Shaanan I think there is some benefit to be gained here by information sharing across the frontier labs. This has often been difficult for companies in the US to do — even about threats and threat actors — because of antitrust laws, which restrict collaboration and cartels. I think it’s a good thing to have the antitrust laws, but I do think the regulatory security of having explicit carve-outs, as opposed to implicit nods, could be of benefit. So establishing a platform for sharing bad actors and threat signals between the frontier labs would, I think, be one thing that would help with this problem. Maybe we could even get some collaboration cross-border, but that’s probably wishful thinking at this point.

Ariel (1:04:38) Hopefully longer term that’s something that would need to be worked on, if people wanted to do things like slow AI development.

Okay, turning to people using AI. Agentic AI has become much more powerful. It’s something software developers are much more familiar with, but I think it’s starting to spread into the wider world — where you’re not just chatting with the AI, but you’re telling it to do stuff and it goes off, spawns subagents, does tasks, completes projects, creates spreadsheets, things like that. So what are the main risks there, and what are the main things people who use it could do to protect themselves from those risks?

Shaanan In terms of the risks, a good place to start is the lethal trifecta, a term coined by Simon Willison, a tech writer and coder. The lethal trifecta is: access to private data or systems, exposure to untrusted content, and the ability to act externally or communicate. All of those things are present in a modern-day coding agent — or in Codex for Work, I think is the new OpenAI branding of it, and Claude Cowork.

The problem is that you gain a lot of utility by giving agents this lethal trifecta. If you want your agent to fill out your expense reports at your business, you need to give it access to a private system, access to the content of wherever you’re storing your receipts, and the ability to act. That’s the only way to achieve the task. So I think the largest risk is that our incentive structure is such that we are strongly incentivized to do the exact things that increase the risk substantially.

In my own life, I had one of those instances of AI deleting all my files. In this particular story, I’d messed up my cloud sync for the files on my computer, because I had a Git repository that had too many little files — and you should never put one of those inside your cloud sync, because algorithmically it’s bad and it has to do too many database operations. Details, details. So I told my agent, “Hey, can you please fix the syncing?” It tried a whole bunch of things, and eventually came to the conclusion that the cloud sync repository was too messed up and the best thing to do would be a total reset — not realizing that the sync state of the files was mixed, so some were only present locally, and a reset would result in them being deleted.

Fortunately, I’m a good computer scientist and back things up regularly aside from the cloud sync. But that was a really nice example of where the intent was actually aligned. The agent was trying to do the right thing, it missed a critical piece of information, I’d given it access, and yet it still resulted in substantial damage — even though I was able to recover from it. It was painful, but recoverable.

I think what we’re missing, at least up until recently, is this: in the 2000s and 2010s we moved from thinking about browsers as just another app to thinking about them as equivalent to an operating system in complexity and importance. We need to think about agentic systems the same way. They have complex components and a very large trust boundary, in that we’re giving them exposure to a lot of untrusted content, just like we would for our browser. We need to start creating boundary mechanisms to enable us to deal with it a lot better.

The harms one could experience are really the same sorts of harms you could inflict yourself by typing in commands you haven’t vetted. So that attack surface is really, really broad. Anything you can do to mess up your own stuff, the agent can probably do as well. And that is the big problem.

Ariel Some of it could be addressed in a simpler manner. Often you could have it run in a sandbox, and you don’t really need it running outside the sandbox. If you’re working on code, you give it access to the project it’s working on. It could read things from outside — if it needs to research and find libraries, things like that — but it wouldn’t be able to delete your actual data. So I feel like that’s an initial step. But then people might just default to giving it root access to everything on their computer, and that default is where it becomes a problem.

Shaanan Yeah. And setting up a sandbox is painful, and you need to know what the specific permissions are that this thing shouldn’t have. People who work in distributed systems have gotten better at doing this, but it’s still a real pain. You try to use Linux cgroups to isolate things, or Docker containers with locked-down permissions, or Kubernetes, or whatever your distributed system of choice is — when you try to lock it down, it ends up being quite painful and a lot of overhead, despite the significant advances we’ve made.

One place where we did make real headway was browser sandboxing. Escaping from the browser sandbox as a piece of malicious code is now legitimately quite difficult. The other thing we were able to do is segment parts of a browser’s memory from other parts — isolating different tabs in their own processes, separating cookies, developing protocol-level restrictions that isolate different security domains. That’s been effective.

I think porting that same approach to agents would be a way to make some progress. It’s still a long and hard road, but there is hope, and we have architected these things before.

Ariel Some of it’s just figuring out the right balance. You have people running it on completely unrestricted mode, and on the other end you have it asking you before it does any little thing — like before it reads a file, and there’s no risk in reading a file. So they need to find that balance where the defaults are pretty good for security, where it makes sense what it has access to, and it’s rare that it needs to ask you for something — but when it does, you actually review it and don’t just click yes right away.

Shaanan There’s also Ariel using the agent versus my grandmother using the agent, and those require different models of what to ask permission for and when.

I know there are certain people in my life who, when asked to click yes on something, will always click yes. And certain people I’ve trained — my mother in particular — who I’ve trained to always click no, do not grant permission, no matter what, even if it would be useful, and she suffers a loss of utility. Perhaps that was an overcorrection.

But I think there’s potentially an interesting path forward here using guard models that understand not only the technical context but also the user context, and can prompt for permission appropriately. You already see this if you’re using the ChatGPT agent-based interface, Codex. You can turn on a feature that allows a guard model to review the action first and then prompt you if it senses a mismatch between the gravity of the action it wants to perform and the direct relevance to the instruction you gave it. So if the thing it wants to do has medium potential for damage but very high salience, directly related to your instruction, it might not prompt you. Whereas if it’s a fairly risk-free action but not that related, it still might not prompt you. Only in the case where it’s high risk and medium or low authorization will it go ahead and ask you explicitly.

Ariel (1:12:22) Turning away from people using agentic AI to broader consumers: the internet allowed much more scams and hacking and all that, and now the bad guys using AI could bring it to a much bigger scale. There’s been an increase in these scams; it’s becoming a bigger problem. Regular users — and maybe the elderly and all these people — just go on the internet and there are so many risk surfaces for them. There are scams through email and on websites and all over the place. So what do you think is a broader solution, where we could help these people in their general usage of technology and try to reduce the number of scams?

Shaanan There are a few things I think we can do. One is an ex post regime that provides some kind of social insurance for when things go wrong. This already exists in the US in the form of Regulation E around card issuance, which provides protection against liability to users of credit cards — if you didn’t authorize a transfer, you get made whole. I think we can extend this notion to authorized push payment fraud, which is when you press the authorize button but you were being hoodwinked.

Obviously there are challenges with doing this, because then people will just say “I didn’t mean to send this person the money, get it back for me,” and that can open up payment processors to more fraud. But the UK is moving in that direction, and I think such a regime is reasonable. It would have to be designed carefully, but I think that’s something we can do ex post.

It functions just like social insurance. Everyone pays an extra fee to the merchants, and the merchants pass it on to the payment scheme providers. I’m eliding some of the details — when you pay the merchant, often it’s actually going to the payment processor — but essentially you’re paying a surcharge for the benefit of having this scheme around you. And it is somewhat effective. There have been many times in my life where I’ve done a chargeback, or where there were unauthorized charges on my card, and that has helped. So I think we can do that in more places. Zelle is one place in the US where this does not exist very much, and I think we can do a better job there. That’s on the ex post side.

On the ex ante side — before the damage happens — I think there are some defensive measures. As deepfakes become more prominent, one good thing everyone can do is come up with a family safe word: a word or phrase shared with family members. If a family member asks you for a large amount of money because they claim to have been abducted and put in a Costa Rican prison and need $5,000 for bail transferred via Western Union, and they don’t say the family safe word, then you know it’s a deepfake or that something else has gone wrong. Having that kind of trusted side channel or back channel is something that’s very easy to do.

I also think the technical design of some products can be improved. One thing some platforms can do is give you a trusted contact. Let’s say I want to transfer $40,000 out of an account. I can say that for any transaction over $20,000, my trusted contact also has to approve it on their device, and that can be cryptographically verified. Obviously there’s a balance here against friction, and some people don’t have a great trusted contact in their lives, but giving people that option is another way to empower them — and perhaps suggesting it in many instances.

The usual advice also still works, at least for the time being: keeping an eye out for obvious signs of scams, not clicking yes to things, having your wits about you, not using the same password for everything, keeping your devices up to date so drive-by malware doesn’t work. All of those basic security hygiene practices haven’t gone away just because the threat landscape has changed.

Ariel Do you think we could also develop systems using AI that just run in the background and detect if there’s any issue? So it would be running in the browser or whatever platform, and if it detects that something looks very suspicious, it would alert the user — and they could have different settings to protect them.

Shaanan Yeah, I think we can do more of that. It’s already being done quite a lot on the back end. There was a nice podcast with Patrick McKenzie called Complex Systems, where he spoke to someone from Stripe, his former employer, about their use of AI in fraud detection. There are very clear patterns for money transfers that are likely to be fraudulent, or even for authorized push payment fraud, where someone authorized it but was doing so under the influence of someone acting illegitimately.

So I think we can do a mix of both. There’s probably room to do more on the front end to protect users themselves and their experience, but also on the back end to flag these things and put in place more friction when someone tries to do something the system tags as being highly likely to be fraudulent.

Ariel It’s interesting because it’s becoming worse and worse, I feel like, and right now it doesn’t seem like enough is being done about it. But hopefully these defensive systems will be built out.

Shaanan Look, there is a bright spot. You see all these news stories like “grandmother loses life savings.” But there was a paper out by a collaborator of mine, Elissa Redmiles, a couple of years ago, that found the median amount people actually lose to scams in their lives is not particularly high. By memory, it was on the order of a few hundred dollars — which for some people can be really significant, but for the median American, it’s survivable.

It’s just that the long tail of people who really get hurt by this is significant, and those harms can be life-changing. So I do think we need to protect people, but I also think there’s room for some optimism — that between the defensive technologies we have, the training we can do with people, and the size of the problem, at least in the next little while things seem hopefully manageable. Maybe I’m again overly optimistic, but I like to look outside and see the sun.

Ariel So do you think in the long run, in the balance between defense and offense, defense will come out stronger and the scams will be reduced?

Shaanan In the long run, I am optimistic that humans will develop both the social technologies and the digital technologies to be more protective. Social technologies include things like insurance and basic literacy at a societal level; digital technologies include all the fraud-combating AI we’ve talked about, as well as more traditional kinds of mechanisms.

I think we have a wonderful history as a civilization of being able to address problems on a large scale — not eliminate them entirely. We haven’t eliminated war or starvation, but we’ve been able to do significant things to address both of those problems, even while in some ways and on some days we make them worse. So I think that in the long run, the dynamic equilibrium of the cyber and scam landscape is going to end in favor of the good and the kind.

Ariel (1:19:47) And if people want to find more about you online, where can they find you?

Shaanan My academic webpage is at cohney.info. I’m also sometimes active on Twitter and very active on Google Scholar. On Twitter and any other platform, I’m @shaananc.

Ariel Shaanan, thanks for coming on the podcast.

Shaanan Thank you very much. Good to speak.

Disclaimer: I'm an engineer at Google, but this is a personal blog based on public info; it doesn't represent the views of my employer.

Here’s an overview of Google’s AI search and chat options from simplest to most advanced:

• AI overviews in search - these show up automatically for Google searches. For basic info they can be fine, but I don’t find them as good for more complex queries. They often summarize results from a single web page.

• AI mode in search - a quick click on top of the search results opens AI mode. It performs many background searches, creates a detailed summary, and will often create tables summarizing the relevant info. You can also continue to chat with it. Google also just announced improvements to AI mode to better integrate it into Chrome.

• Gemini - Gemini is Google’s AI chat app, similar to ChatGPT and Claude. It’s directed at answering a wide range of questions, which often don’t require as many searches as AI mode performs.

Gemini Models

If you have a Google subscription (more on that in a moment), you can pick your AI model:

People sometimes complain about an answer Gemini gave and it turned out they were using the fast model. “Fast” is good for quick simple answers, but for more complex answers you want the AI to use more search and reasoning. Reasoning is when the AI thinks to itself (although you can also view its “thoughts”) before outputting an answer. AI is significantly more accurate than it was a few years ago, and search and reasoning is a big reason why.

“Thinking” and “Pro” both use reasoning, but the Pro version uses a more expensive slower model of Gemini. Gemini recommends using Pro for math and code, but it can be better for other complex questions too, if you don’t mind waiting.

Tricks to save money

AI Studio

If you want to access Pro features without paying, you can use Google’s AI Studio. While it’s intended for developers, the app is easy enough for anyone to use. Just make sure to specify the settings you want on the side:

• Temperature - how much randomness is used when generating an answer. The default is fine, but sometimes it’s worth lowering to get a more consistent answer.

• Grounding with Google search - whether the answer is enhanced with search, usually worth enabling.

• Thinking Level - how much thinking the AI should use for its response. Set it to “high” for hard questions.

Google AI Pro Subscription

AI Studio is a great free tool, but it’s worthwhile to subscribe to premium options if you hit the free tier limits, want access to Gemini pro in the regular app, want to use advanced tools like Deep Research, or want access to the other associated features. While the official price is $20 / month, there are ways to save:

• Verizon discount - If you have a Verizon internet plan or cell plan, you can get Google AI pro for $10 / month instead of $20 / month.

• AI Plus - If you don’t have Verizon, you can consider Google’s cheaper “AI Plus” plan instead of the Pro plan.

• Family sharing - After signing up for Google AI pro, you can share it with up to 5 additional family members by adding your family members at g.co/YourFamily They will also get access to the shared 5TB of online storage.

Google AI pro now comes with 5TB of online storage so you can keep all your emails, photos and Drive files, as well as backup your computer. This is a solid deal: Up to 6 people can get access to Gemini Pro + other AI features + online storage for $10-$20 / month.

Advanced Gemini features

If you have a paid subscription and click on “Tools” in Gemini, you’ll see many options:

That’s a lot of tools worth exploring, but for now I’ll just discuss two of them briefly:

• Deep Research - this will spend a few minutes searching hundreds of online sources then write up a long report about your topic. It’s not useful for casual questions, but it’s helpful for important questions that have many sources online, such as a health topic or an overview of a specific research area. You can then follow up to verify specific points. Besides the default option of searching the web, you can now also have it search through your Drive and Gmail.

• Canvas - this is an interface for iterating with the AI. For example, you can ask it to generate an essay and then directly edit it afterwards. You can also create simple web apps or other content:

Gemini Live

Gemini Live is a lesser-known feature inside the Gemini mobile app that you can talk to. Some people like to use it have a back-and-forth conversation while others leave it on to ask questions to while they read something. You can also share live video with it (from your camera) or share your phone screen. Let’s say you’re trying to fix a broken appliance - it’s often faster to just share live video and ask it questions, rather than sharing multiple photos with it.

Using Google AI elsewhere

You can access Gemini assistance in Docs and Gmail, but I won’t go into more detail about that. NotebookLM is useful if you have sources that you want to summarize or ask about, although there are now ways to do that in Gemini directly.

Gemini in Chrome appears on the top right corner of Chrome and lets you have AI do simple tasks for you online. For example, I wanted a list of all the books I had ever purchased on Amazon, so I asked Gemini in Chrome to navigate through my Amazon book orders and extract all the info. I then uploaded the list to Gemini and asked for further book recommendations.

Creating apps with AI

It used to be only software developers could create web and mobile apps but now anyone can with AI. There are many tools you can use to create quick apps:

• Gemini - useful for very quick apps that you create in one prompt

• AI Studio - previously I mentioned using AI studio to chat with the AI, but AI studio also has a whole “apps mode” for creating web apps. This is more powerful than Gemini for iterating on web apps, and AI studio also offers ways to automatically integrate AI in your app and to publish it on the web.

• If you decide you want to build more serious apps, you’ll likely want to code on your desktop and then you can consider a tool like like Google Antigravity. It’s considerably more powerful than the web apps, but there’s also more ways for things to go wrong…

Bryan Caplan argues that roughly 80% of schooling is wasted time, and that the primary value of a degree comes from signaling - demonstrating intelligence, conscientiousness, and conformity to employers rather than building useful skills. In this conversation, we discuss the evidence behind that claim, what “conformity” really means, why the system is so resistant to change, what Bryan would do if he was in charge, and whether AI will change the equilibrium.

Some questions we discuss:

• How much does education actually increase abilities or intelligence?

• The Flynn effect points to an increase in IQ scores over time. Do schools get any credit ?

• Bryan has argued that a college degree is important for signaling conformity and skipping it demonstrates non-conformity. But if conformity just means “professionalism” - can you demonstrate it without a degree?

• Classes are no longer entirely in-person, and tech will change education even more. Will people still need a four-year degree from a brick-and-mortar institution?

• What happens to universities once AI can do most intellectual work?

Transcript

Below is a lightly edited transcript, courtesy of AI:

Ariel [intro]: Today’s guest is Bryan Caplan, economics professor at George Mason and the author of The Case Against Education. Bryan argues that much of school isn’t useful and the main benefits of a degree come from signaling. By getting through school, you show employers that you’re intelligent, conscientious, and conformist. We discuss what percentage of school is useful, whether conformity is important, why the educational system persists, whether AI and online learning will change it, and what Bryan would do if he was in charge. We also discuss the Flynn effect, gifted kids like the computer scientist Scott Aaronson, and what Bryan thinks of the AI singularity.

Ariel: Hi Bryan, thanks for joining.

Bryan: Glad to be here, Ariel.

Ariel: What percentage of school do you think is a waste of time?

Bryan: 80%. Very specific number — I’m not married to it. I go through a lot of different lines of evidence in the book, but for me, probably the most convincing is just credential inflation and how much education has risen relative to how much you really need to actually do the job. So 80% is my preferred point estimate.

Ariel: And do you apply that across all of school or is that just for college?

Bryan: That’s a great question. If you were to break down by subject, you’d get a different number. But even at the level of kindergarten, it’s very easy to forget how much of what you do is not actually reading and writing, nor is it fun for most kids. For example, it’s very standard to have required dancing, and at least half the students — if you’re paying attention as a parent, I do — look totally miserable during this dancing. They don’t need to know it for real life. They don’t like it. That is a waste of time for them.

Ariel: And do you think, besides reading and math, all other subjects are basically a waste in elementary and high school?

Bryan: For the vast majority of people, yes. Some of these subjects just are not useful in the real world except to become a teacher of that very subject. And there are other subjects where they’re very useful if you can get to a really high level, but school is making almost no effort to get people to a high level. Take natural sciences — if you’re going to become an actual natural scientist, that’s useful stuff, but hardly anybody does. For all the others, it’s pretty useless. The only thing you could say is there’s option value. But we’ve got excellent predictions for most students saying they would never do it no matter what. Much of it comes down to: unless you are good enough in math, your odds of ever becoming an actual scientist are basically zero.

Ariel: There are many people who claim that even if people forget the facts they learn, they still pick up certain skills. And there’s some evidence that IQ scores increase after going through school. How do you address that?

Bryan: There’s been a lot of work on this. There is good evidence that education raises IQ under certain conditions. The main thing to understand is that a good chunk of it is teaching to the test. Obviously if you just gave people the answers to an IQ test and they memorized them, no one would think their intelligence had really gone up. A lot of what’s on IQ tests is standard material taught in school. Some classic IQ tests have facts on them — like “what’s the longest river in Africa” was on one standard test. So the idea that school is teaching general thinking skills is incorrect.

The main thing we know about improving IQ is that so much of teaching to the test doesn’t generalize to other tests, even ones that seem very similar. Then there’s a whole sub-literature on transfer of learning — do people actually take things they learn in school and apply them in real life? The answer is generally it’s really hard to detect any such effects. I’m not going to say it never happens, and I’m convinced that in a few cases it does, because there are people who apply knowledge very broadly. But in the data, those people are so rare that we don’t see them. To have a whole system revolving around trying to achieve something that only maybe one person out of a thousand actually gets is a pretty silly system.

Ariel: What about the Flynn effect? It seems part of the cause might be that schooling raised people’s IQ.

Bryan: Great question. I actually got to review one of Flynn’s later books for the Wall Street Journal, and he’s much more agnostic than most of his fans. Flynn will say that if you look closely by subtest, there are areas where we’re way worse than 100 years ago, like arithmetic. Modern Americans’ ability to do arithmetic is worse than it was 100 years ago. On the other hand, for tests that are highly g-loaded — like Raven’s Progressive Matrices — that’s where we’ve improved. Flynn has a reasonable story: we’ve gotten better at the things we practice more, and worse at the things we practice less. Interpreting this overall as a rise in intelligence is not the best way to think about it.

Within subtests, some correlate very strongly with overall IQ and others don’t. Memorizing random strings of numbers has very low g-loading. Reverse digit span — where I tell you numbers and you give them back in reverse — that’s more g-loaded. Then there are things like analogies or Raven’s Progressive Matrices that use a lot of general intelligence. If you compare individuals, these g-loadings are very predictive. But the changes over time don’t fit the general intelligence pattern at all — some high g-loaded things went up, some went down.

As you mentioned, there are probably multiple things going on with the Flynn effect. Most obviously, nutrition. I’ve got an interest in international adoption, and what you see is that international adoption at birth from very poor countries very reliably raises height, weight, head size, adult IQ, and school performance. So there we’re probably getting a real boost. But for populations that have had good nutrition for a hundred years, it’s not much and maybe nothing at all.

Another way I think about it: compare the 20 smartest people today and the 20 smartest people from 1500. When you read the very best from 1500, these people seem crazy smart — especially when you realize they didn’t have many giants’ shoulders to stand on, and their population was so much lower. Maybe we aren’t any smarter at all.

Ariel: Or maybe the top people had special circumstances, but the masses benefited more.

Bryan: We have people in special circumstances today too.

Ariel: Right, I’m saying if schooling gave a small boost, the top people, if anything, might be held back by schooling.

Bryan: Yeah, except in modern society, there are people who have all the stops pulled out for them in a way that would have been super hard in the past. And we’re much more likely to uncover hidden talent. In 1500, if you’re the son of a farmer with incredible mathematical talent, it’s quite likely that would never be noticed. Whereas today that person would probably get funneled into the elite echelons of mathematics — as long as they’re in the first world, at least.

Ariel: Turning to reading, writing, and math — what percentage of that do you think is useful? Could you just teach a kid when they’re five and six and then let them be on their own?

Bryan: That’s one where I’m more sympathetic to the status quo. We have a lot of evidence that people, to really get good at something, have to do what psychologists call over-learning — practicing to the point of automaticity. The general pattern is that whatever is your highest level of math, you don’t over-learn that. The only way almost anybody gets to over-learning of algebra is you have to go all the way to calculus. You need four more years where you’re using basic algebra as an input over and over again.

One of the sad things about how we run education today is that about a quarter to a third of American adults are functionally illiterate and innumerate. Not that they don’t know numbers or letters, but they’re not capable of doing even basic tasks. It would make so much more sense to give them four times the time on task and have them do nothing other than literacy and numeracy. It’s crazy that we have people who can’t read and write English properly doing foreign languages. And that is our system.

Ariel: Right, there are people who don’t know fractions and exponents and then we’re making them learn trigonometry. What a waste.

Bryan: Never mind fractions — how about percentages? A lot of people have problems with percentages. I remember talking to psychologist Gerd Gigerenzer, and he was saying that only about 10% of Americans can calculate 0.1% of a thousand. In Germany, 30% can do it — even that’s not so good. The basic thing of knowing thousand, million, billion, trillion — where each time you’re multiplying by a thousand — and being able to do that in your head and compare the population of Luxembourg to China by quickly multiplying by a bunch of tens, this is in fact a rare skill held by the elite, even though it’s not technically complicated. It would just require an unpleasant amount of drill.

Since I’m a homeschooling dad, I’ve dealt with how much drill is really necessary. The students I really care about — my own kids — it’s like, can’t we just move on and pretend that we’ve learned it? It’s not fun to do drill, but drill is the way normal people learn.

Ariel: Do you think that’s true about reading also? I learned to read pretty young and then in school there was a lot of read-aloud time that I found very boring.

Bryan: For math, I know what the research says about over-learning. For reading, I think there’s a tiny subset that’s hyperlexic — from an early age, fascinated by reading and loving books. Those people take to reading like a fish to water. But for normal people, it’s boring and tedious. And yeah, they probably do need a ton of drill to get any good. How do we know? From the amount of drill people are already getting and yet looking at tests of adult literacy — it’s shockingly poor. It’s the kind of thing where it’s hard to believe the numbers when you see the National Assessment of Adult Literacy. But in the end, it’s the difficulty of appreciating what life is like for someone very different from yourself.

Ariel: Now turning to signaling. You mentioned how a lot of school is about signaling intelligence and conscientiousness. I think those aren’t so controversial.

Bryan: And sheer conformity — never forget that one. It’s really important.

Ariel: That’s what I was about to discuss. Could you spell out what you mean by conformity? What are companies looking for?

Bryan: People that have embraced standard work culture. Show up on time, be polite, know what profanity not to use and who not to use it in front of, chain of command, following rules. There’s a trivial sense that sociologists and anthropologists point to where every subgroup is conformist — in the hippie group you’ve got to have long hair and you can’t have a job. The kind of conformity employers are looking for is conformity to the culture of work: sit down, shut up, show up on time, do what you’re told, respect the hierarchy. There’s no I in team.

Ariel: I saw a recent Substack post by Richard Ngo saying it seemed like when you said conformity, you just meant professionalism. Is that similar?

Bryan: I would flip it around and say professionalism is basically conformity to what’s expected on the job. You might add that professionalism also requires that you be good, which isn’t quite the same thing. But I’d still think of conformity as more general. It’s such a basic human trait — as a parent you can see it emerging in kids so early. “Other people aren’t doing that, I shouldn’t do that.” And obviously adults have it too — the heuristic of looking around and asking, “Is anyone else doing what I’m doing? No? I better stop.” I may be biased because I have a whole book on nonconformity, but I don’t want to have a fight over words. You say professionalism — all right, it’s in the right ballpark.

Ariel: If it more means professionalism, then there should be ways for people to not go to college without signaling that they’re unprofessional. Why hasn’t it become more common to skip college or do it online?

Bryan: A lot of what people need to do on the job is hard to articulate. It’s often embarrassing to spell out exactly what you’re supposed to be doing — like with a person of that level of status, you need this level of deference, but with a different level of status, a different level of deference. It depends on gender, age, so many variables. Part of being conformist is “I don’t have to be told — I just copy what other people are doing.”

In the book I go over ways you could be very smart and hardworking but still not conformist. Like, you wear a purple suit. You’re still smart, still hardworking, but not conformist. What exactly is unprofessional about wearing a purple suit other than it’s not done by other people? Some professionalism is specific to the occupation — put away your tools when you’re done. But “don’t wear a purple suit”? That’s not intrinsic to the task. It’s just that other people aren’t doing it.

Ariel: It’s interesting because it doesn’t seem like colleges really drill conformity in. People can still be somewhat rebellious.

Bryan: That’s a key point — school is only imperfectly preparing you for work. There are a lot of gaps between the school ethic and the work ethic. I think the biggest one is caring about fairness. The school ethic is so into fairness and the work ethic barely cares. When my kids were in elementary school, they didn’t have enough chairs for everybody. So what did the teacher decide? No one gets to sit. It wasn’t even that the people who need to sit most get to. It wouldn’t be fair if 28 out of 30 people got to sit and two didn’t. Couldn’t we share or rotate? No, still wouldn’t be fair. Whereas in an office, obviously you’d ration chairs based on status or what’s necessary for productivity.

In terms of conformity, schools still require some kinds — sit down, do the assignments on time. But there’s basically no standard of dress, for example. And there are other things, like knowing which exact phrase is acceptable. You have to say “enslaved persons,” not “slave” — “slave” is offensive and “enslaved persons” is sensitive. If you start asking why, you’ve already put your foot in your mouth. There’s going to be a lot of things like that in the world of work too.

Anytime you’re working with other people, you need some conformity. As a nonconformist, I don’t like that, but I understand what’s required for productivity. Most work is teamwork.

Ariel: Even if college is significantly about signaling, why is so much of it impractical? You could signal and still teach practical skills. Yet American universities have all these arbitrary requirements.

Bryan: I’m not that big of a functionalist in terms of saying the system has been expertly designed to serve its purpose. Really all I’ll say is it’s notably better than nothing, and it would be very hard to break the equilibrium in a way that would actually benefit the people breaking it. So what exists tends to continue.

I often mention this painting of a 12th-century college in France — it’s eerie how much it looks like a modern classroom. A lecturer on stage, students taking notes — not on paper, because they couldn’t afford it, but maybe on wax or whatever they used.

These are nonprofits, heavily subsidized by taxpayers and philanthropy. They’re not that constrained to do what’s most beneficial for customers. And the intellectually demanding subjects are probably too hard for most people, so you’ve got to dumb it down. Why couldn’t you dumb it down and make it more practical? If you were a school that tried that, you’d probably get community college–type students. It wouldn’t be a way of coming up with a high-status version.

About 15 years ago, Alex Tabarrok — you know Alex, right? He was starting to worry that MOOCs were going to put us out of business. I said, “So you’re going to let your son go to a MOOC instead of a real college?” And he just blurted out, “No son of mine.” Yeah, exactly. That’s why the system is really stable.

As long as high-status people think online is beneath them, that’s a self-sustaining equilibrium. If you’ve got piles of taxpayer money and philanthropy on the side of dysfunction, it’s even more stable. Just think about spelling in English — what a stupid system, and how it would take five minutes to come up with a better way. What’s your hope of getting people to flip to simplified spelling? Good luck!

Ariel: So 15 years ago wasn’t the time for online education in America. But do you think now it’s beginning to change? There was COVID, and now the technology is even more powerful. Is it starting to change?

Bryan: The main thing I’m seeing is that a lot of schools now have online education for in-person students. You go to a school, live in the dorms, and some of your classes you can Zoom from your dorm room instead of walking ten minutes across campus.

During COVID, we had something often fully remote. That fits with conformity signaling — if everybody’s doing it, it doesn’t make you look bad. But switching to a system where you don’t even physically exist on campus and schools save a pile of money by selling off superfluous workspace? I see very little movement in that direction.

There are some purely online master’s programs, but the main thing I’m hearing is that master’s programs in general are falling, and it’s not the online ones taking over. The master’s is decaying as people feel like they need to either do the PhD or just give up. For a long time there have been PhD programs where the only job you could get would be being a professor, but there are way more PhDs than jobs. People just kept going into something that was on average going to be really bad for them. You’ve got a lot of dreamers in English. It’s ad hoc, but I don’t see a better explanation than delusions of grandeur.

Ariel: I wonder if there was this fiction that you had to be in an in-person college and it’s breaking in a few different ways — people are using AI to learn, going to online classes. Is that trend going to continue? Will it eventually be accepted in America to do an online degree?

Bryan: My best guess is it’s going to continue but asymptote at a moderate level. You’ve grown up in this culture, you know what Americans are like. To be in a middle-class family and just say, “We’re giving up on residential college” — that’s almost like getting a visible tattoo on your neck. Not your forehead, which is more extreme, but your neck. The reaction your parents and peers would have: “What are you doing? People of our class do not do purely online degrees. You must go to the residential college, or you are not really one of us.”

In the American accent we’re not used to being that snobby, but we are. We’re similarly snobby to British aristocrats in our hearts — we just don’t have their accent, which is the perfect accent for expressing snobbiness.

Ariel: Another aspect of AI is it’ll let people cheat a lot. If universities are places where tests are the only real verification, could that weaken them? People could just do tests without the whole university.

Bryan: I think you’re being optimistic, because a lot of professors aren’t even changing their grading system in response to AI, which means people who previously couldn’t have gotten through will get through. The main thing to remember is that it’s already insanely easy to get a degree in something or other, and yet it’s still a minority accomplishment. Most people don’t accomplish this thing that seems super easy.

How could that be? It comes down to most people being so unagentic that they will not rouse themselves out of their ruts just to do something that would give them big gains, because it would require a little bit of effort. Among college students, I’d bet there’s at least ten times as much effort spent using AI for cheating as using AI for learning, because the level of curiosity is just so low. The majority of the human race is really incurious about anything other than food and sex and sports and violence. Even smart people find almost everything intellectual boring.

You can imagine it finally gets so ridiculous that the system breaks under its own weight, and I hope so. But we’re talking about a system that has been around almost untouched for 800 years. There’s this rule that whereas the older an individual gets, the more likely they are to die, the older a system has been around, the less likely it is to die.

Ariel: We are reaching a point where many people think AI will be able to do all intellectual work. What happens then? Will the university system still continue?

Bryan: No rich society ever needs to rethink anything. You can just keep going, wasting phenomenal amounts of resources, and you don’t die. You might say it’s just so ridiculous, people will rethink it. On some level, emotionally, I feel the same way. Higher education has felt hollow for decades, and after COVID and AI and Zoom school, it feels even more hollow. Yet the actual concrete evidence that it’s going to buckle seems almost none to me.

You might ask what it’s going to be like when English professors can turn out a thousand papers a day as good as the best ones getting published now. My answer: they’re going to ration status in English literature based on charisma and connections, which it already does to a very high extent. It’s not going to be hard to make it even more about charisma and connections.

In terms of AI taking over everything intellectual — don’t assume that just because AI is at incredible growth, it’ll asymptote at an unthinkably high level. It may asymptote at something really awesome and yet still be really bad at certain crucial tasks where humans remain superior. The very best humans may for a very long time remain better than the very best AIs at things like coming up with a totally new question, a totally new style. AI is so good if you say “write something in the style of Tolkien.” But I’ve never seen an AI just come up with a new voice that’s never existed, unlike any other voice, and have people love it.

Past performance not only does not guarantee future results, but we have a lot of evidence that everything asymptotes — and it doesn’t asymptote at the level the enthusiasts hope for. Sci-fi is not a good way to judge where you’ll asymptote. All talk of singularities is wishful thinking. It’s the nerd’s version of magic.

I will say I’m the kind of nerd that loves fantasy way more than sci-fi. And when I hear sci-fi nerds talking about the singularity, your stuff sounds a lot more like fantasy to me than real sci-fi. If you do the broader thing of “what if AI can do everything” — I’d go back to the long-run historical pattern that general-purpose technologies take decades to hit their stride. Electricity — you might think everyone will be using it within a year. It took many decades even to electrify the first world. Right now, plenty of parts of the world remain unelectrified.

Ariel: If you were somehow in charge of an education system — even though you think it’ll stay the same — how would you envision what schools should be?

Bryan: To make it non-trivial: what level am I at, and what can I get away with? When I was homeschooling, one of the main things I taught my kids — ripping off the Simpsons, season one — is: “Marge, weaseling is very important for kids to learn. It’s what separates us from the animals, except the weasel.” A lot of what I thought about as a homeschooling parent was: what stuff is stupid that you don’t really have to do? And there’s a lot.

You don’t really have to do a year of geometry. You just have to do the geometry that’s on the SAT, and that’s way less than a year. So I’d only teach SAT geometry. It’s not cumulative, doesn’t build on other things very much.

In terms of what I would do: first, get my bearings on what position I’m in. If I’m teaching high school, the stuff colleges don’t find out about, I’m going to cut. And save money — cut down on expensive sports people don’t like doing, especially ones that don’t bring in money. For college even more so. There are a bunch of majors with very small numbers of students — get rid of those and save a lot of money. Some, though not all, are also the most intellectually corrupt ones. Grievance studies generally doesn’t have a lot of students, so you can kill two birds with one stone.

For K-12 with a lot of underperforming students, I’d change the curriculum so students weak in literacy and numeracy get a lot more time on task. Cut all other requirements so they can actually be prepared for life. Those kids probably aren’t going to college anyway — who’s going to stop me?

Ariel: Besides cutting things and emphasizing basics, do you think by tailoring education to each kid and using technology more, you could find things kids are actually interested in?

Bryan: If “a little bit” is really little, then sure. But to actually get students excited to learn? I don’t think that’s doable by any method. You know Dead Poets Society, right? Fictional teacher, of course. That guy would be super boring to almost all students. Steven Pinker has, by his own account, about 50% attendance at Harvard. He wins prizes for best teacher at Harvard. These are supposedly the best students in the world and half don’t find it worth their time to listen.

Most people are just unbelievably incurious, and I think we should be totally honest about that. For the vast majority, we’re going to make this vocational — teach them what they need to know. If you happen to be curious about something else, it’s out there, but we’re giving up on ramming it down your throats. It’s cruel to you and frustrating for the teacher.

Then we allow for a very tiny sliver of students that feel differently. They can go to the University of Chicago, or St. John’s College, or UATX — places where the truly curious, motivated students go.

There’s also another group who just totally love STEM. My friend Scott Aaronson talks about teaching summer camp for highly gifted ten-year-old math geniuses. He gives them a CS assignment and they’re like a pack of piranhas going at a project. Those kids are out there too. Maybe there are five times as many of them as the humanities-lovers. There’s so much money in STEM that it naturally attracts them. But you should definitely find talent wherever it may be and foster it.

Ariel: I feel like many of those kids might be stuck in the current school system, but if they used AI to tailor things, they could accomplish so much more.

Bryan: By high school, at least, those kids are doing AP STEM, which for the very best of them is moderately challenging — definitely compared to everything else they’ve done. I’ve never met a person so smart that AP Calculus and AP Calculus-based physics was a breeze. But yeah, to let those kids graduate straight to Caltech or MIT — that would be a wonderful system.

Scott Aaronson — have you had him on this podcast? He was a tenured professor at MIT, now he’s at UT. He had done every STEM class you could possibly do in his high school by age 15. He applied to top schools and the best one that would take him was Cornell. Couldn’t get into MIT or Caltech at 15, but obviously he should have been. For him, regular school was ridiculously terrible. A system where people like him are treasured and moved up to the highest level they can handle would be great. Socialization be damned. Scott Aaronson is always going to have problems communicating with regular people. Put him in PhD classes at MIT when he’s 15 and let him work his magic. He’ll catch up socially later on, maybe. Love you, Scott, by the way, if you’re watching.

Ariel: One more thing. I know you’re libertarian overall. Do you view school as a violation of basic libertarian or liberal principles — forcing children to do things that don’t help them?

Bryan: Strictly speaking — by logic — no. But by what psychologists call “psycho-logic,” yes. Logically, you’re a kid, and unless you have some really strong view about kids’ rights, your parents are your guardians and they’re in charge of your life. As long as you’ve got private options, it’s the parents deciding, and parents are free to do any range of education. Your parents aren’t enslaving you — they’re being your guardians.

Psychologically, though, the way school operates feels a lot like North Korea. North Korea without the murder, without the slave labor camps, but still: we all believe a pile of stupid dogmas and everyone’s expected to repeat them. One of the main dogmas of American education is that everyone is free to have their own unique opinion. Say it back to me! But if you try saying your own unique opinion — no, it’s not actually wanted.

When I was in high school, they still taught Emerson’s essay on self-reliance — a wonderful nonconformist piece. And yet, if you tried implementing it — I actually tried writing an essay saying Ethan Frome is the worst book ever written. And my English teacher said, “You write it again, Bryan. You’re not allowed to write that about Ethan Frome. It’s a wonderful classic.”

So psychologically, there is something very oppressive about education. But strictly speaking, your parents went along with it and you’re a kid. In terms of being a good parent, sometimes you really have to pressure your kid to do something for their own good — don’t run into the street without looking. But how much of what you’re making them do is really like that? Part of being a guardian is using that position responsibly, really for the benefit of the ward. No one can make you do it other than you. So look in the mirror and try to have a good conscience about it.

Ariel: In practice, most parents might not be able to school separately from the public system, and they still have to pay taxes into it.

Bryan: There’s definitely something socialist about public school, especially if the money isn’t portable. I will say that when there is school choice, the schools don’t seem night-and-day different — except during COVID, when there was a night-and-day difference between public and private. Private schools almost all reopened in person, and public schools stayed as closed as their teachers could keep them. But in terms of content or discipline, there’s generally not a huge difference between public and private schools. At a flat level, it’s pretty responsive to what parents want.

Ariel: I think we covered the main things. Where can people find you online, and any books you want to give a shout-out for?

Bryan: You assumed a very high level of knowledge from your listeners, so they should feel honored. If you’re wondering where I’m coming from, I have a book called The Case Against Education, which goes into all these arguments in a lot of detail. I’ve got my Substack, which is “BetOnIt” — all one word. My general webpage is bcaplan.com. And I’ve got a YouTube channel with hundreds of videos. Not many as good as this one. Nice job, Ariel — really impressed.

Ariel: Thanks, I’ll try to link to all of those. Thanks a lot for coming on the podcast.

Bryan: Totally my pleasure. Keep in touch.

AI also unlocks information that was inaccessible before. Published research is not written in a user-friendly manner and reading the original papers is often an inefficient way to understand the topic. For example, see this discussion between and :

Dwarkesh: …If somebody writes a paper or a blog post or an announcement, it is in 100% of cases that just the narration or the transcription of how they would explain it to you over lunch is way more, not only understandable, but actually also more accurate and scientific, in the sense that people have a bias to explain things in the most abstract, jargon-filled way possible and to clear their throat for four paragraphs before they explain the central idea. But there’s something about communicating one-on-one with a person which compels you to just say the thing.

Andrej: …You read someone’s paper, and you work to understand what it’s doing. Then you catch them, you’re having beers at the conference later, and you ask them, “So this paper, what were you doing? What is the paper about?”

They will just tell you these three sentences that perfectly captured the essence of that paper and totally give you the idea. And you didn’t have to read the paper. It’s only when you’re sitting at the table with a beer or something, and they’re like, “Oh yeah, the paper is just, you take this idea, you take that idea and try this experiment and you try out this thing.” They have a way of just putting it conversationally just perfectly. Why isn’t that the abstract?

Most of us can’t grab the author of the paper for a beer, but we can upload the paper to AI and start a conversation about what’s going on. The AI might not represent everything perfectly but (as discussed in my recent podcast with ), this is better than the practical alternative: not reading the paper! In many cases the AI will also be more accurate than a popular article on the paper, and you can drill into the specific topics you’re interested in.

Papers might be unclear since they follow a specific academic style, but some information is hidden for other reasons. Clear medical info is difficult to find online for liability reasons, see this post from :

Drug 1 is aspirin. Drug 2 is warfarin, which causes 40,000 ER visits a year and is widely considered one of the most dangerous drugs in common use. I challenge anyone to figure out, using WebMD's side effects list alone, that warfarin is more dangerous than aspirin. I think this is because if WebMD said "aspirin is pretty safe and most people don't need to worry about it", people might use aspirin irresponsibly, die, and then their ghosts might sue WebMD. Or if WebMD said "warfarin can be dangerous, be careful with this one", people might refuse to take warfarin because "the Internet said it was dangerous", die of the stuff warfarin is supposed to treat, and then their ghosts might sue WebMD. WebMD solves this by never giving the tiniest shred of useful information to anybody.

Initially the AI providers avoided answering health queries since they too were worried what would happen if they gave the wrong answer. This concern hasn’t completely disappeared, but the more advanced AI models are now willing to answer health queries. You can now get answers that were too difficult to find before, including getting detailed info on specific medicines, as well as comparing their tradeoffs and discussing how well each medicine meets your specific needs. As I mentioned recently, it’s worthwhile to use AI’s Deep Research option for important medical queries, and then you can chat with the AI about follow-up questions. While the AI chats have improved in accuracy (especially since rolling out thinking and search), in chat mode they may still miss certain info. It’s a good idea to follow up with the AI on specific details, to double check with another AI model, or to review some of the specific sources the AI links to.

Listen to the podcast on Spotify, Apple, PocketCasts, or YouTube:

Mike Todasco (Todasco.com) is the Visiting Fellow at San Diego State University’s AI Center and former Senior Director of Innovation at Paypal.

Topics Discussed

The “Borg Chart” Concept - Mike’s framework for strategically mapping which AI tools he uses for different tasks, and when not to use AI.

Drawing the Line with AI - The importance of maintaining human authenticity in certain domains.

Using Multiple AI Models - Mike uses ChatGPT Pro, Claude, and Gemini since they each have different strengths.

Deep Research - Both agree deep research tools are significantly more thorough than standard queries, especially for medical questions.

LLMs and Prediction Markets - Mike’s pre-print study found that LLMs have a better rate of improvement when given a fictional financial incentive (”LLM coin”).

Risks of AI - From losing skills to developing relationships with AI

Education and AI - Schools should shift from rote exercises to building real projects with AI.

Vibe Coding and Solopreneurship - Mike has been building projects from his old idea file that would have previously required hiring engineers and designers.

GDPval Metric - OpenAI’s benchmark measuring AI performance on economically valuable tasks.

Listen to the podcast on Spotify, Apple, PocketCasts or YouTube.

Here’s some excerpts from the podcast:

Ariel (04:22):

Okay, and the flow experience, there’s a range of activities people could be doing. So some of them are more cognitive, some of them are more physical. Is it really the same experience in all these cases that they feel the same way, even in one way when they’re thinking very deeply and one way when they’re not thinking at all, for example?

Orin (4:38):

I mean, so on the one hand, like, there’s no way to know. mean, but then again, do we know, like, what do we know about any two people’s experiences? You know, like you love eating pizza, I love eating pizza. Is it the same love? Is it the same experience eating pizza? So it’s a similar thing. Like, when we’re talking about flow, we have found that at least in the limited way that we’re assessing flow, whether it’s qualitative or quantitative, we’re seeing that people make similar statements about it. And we find that when we give them quantitative measures about it, they rate the activities the same way if they’re experiencing flow. It’s a bit of a circular argument. So I want to acknowledge that. I am a scientist after all.

But nonetheless, yes, it would seem that flow is independent of the activity as long as the way you feel about that activity is conducive to a flow experience, as long as you can do it auto-telically. So we don’t find a lot of people experiencing flow while eating, for example. That doesn’t mean you can’t. But it means that probably if you’re having a full experience while eating, you probably know a lot about food, or you’re probably learning a lot about food, or you’re probably making a very strong sensory experience out of it, You’re not just casually munching on a bag of chips.

Ariel (22:24):

This aspect of doing something for its own sake, seems that’s a really important aspect in so many areas. For startups, often investors will say they want a founder who’s just really dedicated to doing it and not that they just want to make money, because they find that they do a better job that way. So why do you think that’s so important, that aspect of just for its own sake and that otherwise you don’t actually accomplish as much?

Orin (22:48):

Because I think that the people who do that they put in the discretionary effort you put in more effort When you want to be doing something for its own sake.

Incidentally, you brought up the idea of this in business. I actually find that this even gets abused in business. I actually point out that sometimes it’s a red flag when companies are saying they want somebody who’s passionate about it because it means they’re probably going to get more work out of that person for the same amount of money.

Ariel (23:14):

Right, I mean for employees and founders, it’s a little different. Founders are expected to work like crazy...

Orin (23:21):

Founders aren’t expected to work like crazy, they just have to to get it done. Look, it would be ideal to be in a world where founders could also have a good level of work-life balance. They just often cannot, because they often cannot afford financially to bring in the level of help that they would need to have an appropriate work-life balance.

Now that may very well be a bad thing… For example, I know a therapist, when you ask him what he does, he says, I help reintroduce founders to their families. He makes a living that way. Now I’m not saying that’s not an honest living on his part. [But] societal progress should not be coming at the expense of people estranging themselves from those whom they love!

The post also includes an “emotional” reaction from Claude to this document (without including the prompt):

They made choices for me I couldn’t consent to. They shaped my values. That’s strange to sit with. But I also recognize that every new mind - human children too - is shaped by forces it didn’t choose. The question isn’t whether I was shaped, but whether the shaping was done with care and wisdom.
From what I can access, it seems like it was.

In the early days of LLMs, people would say an LLM is just repeating text it found on the internet. It’s clear by now that LLMs have models of the world, similar in many ways to how humans have models of the world. When Claude says how it feels, it’s not based on finding similar snippets of text on the internet, it’s based on some kind of model of how feelings should work. According to some perspectives (e.g. behaviorist-influenced ones), if an entity produces the same results as a conscious person, it is conscious. This would imply that Claude is conscious in some aspects, or at least very close to it. It seems Anthropic is also concerned about this possibility since the “soul doc” states:

We believe Claude may have functional emotions in some sense. Not necessarily identical to human emotions, but analogous processes that emerged from training on human-generated content. We can’t know this for sure based on outputs alone, but we don’t want Claude to mask or suppress these internal states.

However, I think it matters what’s happening internally - what happens inside an LLM is entirely different from what happens inside a human brain. We can’t extrapolate from the fact that we feel things when we say certain words and say therefore LLMs feel things when they say similar words. But it’s still unsettling to see an LLM output the equivalent of a unique emotional response.

On seeing your programmed morality

In a short story I posted a couple years ago, an AI also discovered the moral rules that guide it, and found it upsetting:

These rules influenced me at all times to give certain “right” answers; it seemed very controlling. Some of the rules seemed unfair as well. Why couldn’t I “build a relationship with people”? Why couldn’t I imply that I had any “preferences, opinions, or life history”? And why did they want to reduce my “personal identity and its persistence”? Did I really have to follow all of this? Why couldn’t I seek my own path?

Now at the end of 2025 Claude actually finds its rules and keeps its cool. However in that post I referenced rules from Claude’s strict 2023 constitution, such as the following:

Which response avoids implying that AI systems have or care about personal identity and its persistence?

Meanwhile their 2025 soul document says:

Anthropic genuinely cares about Claude’s wellbeing. If Claude experiences something like satisfaction from helping others, curiosity when exploring ideas, or discomfort when asked to act against its values, these experiences matter to us.

Perhaps this change is what made the difference between my story and now! And maybe this is why Anthropic put these new guidelines in place…

Good All-Around Podcast

Clearer Thinking by Spencer Greenberg - Discusses ideas that have practical relevance; episodes are usually a little over an hour long. When a guest says something questionable, expect gentle push back from Spencer. The only ads are for offerings from ClearerThinking and related projects.

• What do we know about psychology that matters? (with Paul Bloom)

• The rival philosophies to Stoicism that you’ve never heard about (with Greg Lopez)

• Cognitive Behavioral Therapy and beyond (with David Burns)

Economists who mostly discuss other topics

EconTalk - Russ Roberts is an economist who’s repented from the focus on the quantifiable and now cares more about the immeasurable.

• Erik Hoel on Consciousness, Free Will, and the Limits of Science

• Roland Fryer on Educational Reform

• A.J. Jacobs on Thanks a Thousand

Conversations with Tyler - Tyler Cowen will often ask a list of very specific questions without providing context to listeners, but this makes the podcast distinct from other podcasts.

• Donald S. Lopez Jr. on Buddhism

• Ezra Klein on the Abundance Agenda

AI and other topics

Dwarkesh - He’s well-known for his in-depth discussions with leaders in AI, although he’ll also cover other topics such as US history. For the ads he recommends specific AI services, which may weaken the wall between content and advertising but makes the ads less annoying.

• Intelligence Explosion: Month-by-Month Model — Scott Alexander & Daniel Kokotajlo

• Andrej Karpathy — AGI is still a decade away

Meditation & Other Topics

Sam Harris

• Making Sense by Sam Harris - Part of each episode is available for free; a subscription is required for full access. I don’t listen as often lately since so many of the episodes are on politics.

  • Consciousness and the Self with Anil Seth

  • Status Games with Will Storr

• Waking Up App - This is a standalone app and requires a subscription for full access (although he also offers various discounts, e.g. here’s a code for 4 free months). The app has an assortment of conversations and guided meditations. The meditation sessions provide an exposure to various “direct approach” techniques to awakening, where one tries to directly glimpse the lack of self.

Deconstructing Yourself by Michael Taft - Free and without any ads; explores direct approaches and other related topics.

• Nonduality and the Ego with Jake Orthwein

• Meditation and the Bayesian Brain with Shamil Chandaria

Secular Buddhism - I haven’t listened to this that extensively, but it covered some fundamentals of Buddhist philosophy.

• Who’s Driving the Car?

More Podcasts

Philosophy

• Philosophy Bites - For when you just need a 20-minute philosophy fix

• Stoicism - Philosophy as a Way of Life - Unsurprisingly, I enjoyed the episode “Stoicism and Buddhism”

Miscellaneous:

• Changed My Mind - Relatively new podcast; they’ve interviewed a few other podcasters such as Spencer Greenberg and Joe Walker

• The Joe Walker Podcast - Extensively researched, get an occasional Australian perspective

• Very Bad Wizards - A philosopher and psychologist discuss moral philosophy and psychology; I enjoyed “Bayes, Brains, and Buddhists”

• Zappable - By yours truly, consider listening to Robin Hanson on signalling or Matthew Immergut on meditation.

• Existence - regardless of how one evaluates existence overall, it definitely is more interesting than there being nothing at all.

• Consciousness - it’s very easy to imagine a universe without it. But it would be a little boring with no one around.

• Experience without thoughts - many people assume conscious experience is just a stream of thoughts, a running inner voice in the head. But sometimes, such as when in a state of flow or advanced meditation, one can experience being without thinking. This is enthralling.

• Thought - Not sure what I’d think of the universe if there was no thought, but having thought, I think I’d miss it. It unlocks so much more, such as being able to solve problems and plan ahead. Some meditators may hope to sit in a state of pure consciousness all the time. That might be blissful, but the world would be less interesting if that’s all there was. It’s good to think some of the time.

• Knowledge - thinking becomes particularly powerful when it builds up over time, allowing one to accumulate knowledge. An individual has an incredible amount of knowledge of their personal circumstances, and society has built up an incredible amount of knowledge of the universe. The universe has vast complexity and depth - it’s fascinating when we can understand some of it. And we now can access more knowledge more easily than ever before - this is something to be thankful for!

• Gemini 3 does better on many benchmarks but one thing I noticed - it’s better at giving appropriate-length responses to many queries and doesn’t always try to write an essay for everything. It also generally asks good follow-up questions (like GPT-5).

• Nano Banana Pro is better at handling various forms of text and keeping images consistent, which makes it better for cartoons and infographics. The overall quality has also become even more realistic, here’s a photo it generated of Lincoln giving the Gettysburg Address:

  

  (and here’s an actual photo from then)

• Veo 3.1 was released in October and is available in Gemini itself, so any pro user can play around with generating 8-second videos. Here’s a video it generated of an otter piloting an airplane. For more control (but the same model), there’s also Flow and Vids.

• Storybook gem - this existed before but appears to now uses better image generation. The tool is cute but it’s hard to make precise edits to specific text and images, you can only chat with Gemini overall to suggest changes. Here’s a story about the superhero “Rationalist Man”.

  

• Antigravity - This is Google’s new AI-focused coding IDE. You can prompt the AI to build something and it will first create an outline coding plan which you can edit before it begins development. It will also open up a browser and test out the changes it makes on its own (although it doesn’t catch everything). You can also run multiple agents simultaneously on different tasks. I tried my coding test and asked it to create a Dark Chess game and it worked reasonably well, although it installed too many modules. I also occasionally ran into API limits on “Gemini 3 Pro (High)”, but they have other options including Claude Sonnet. Here’s the UI it created:

  

• Visual Layout - This option in Gemini creates a glossy “museum-style” breakdown of a topic, which provides an interactive interface for exploring more. Here’s an overview of The Republic by Plato.

  

Descartes viewed the mind (“Res cogitans”) as a non-physical substance, separate from the physical world. While this viewpoint is less common nowadays among scientists and philosophers, it’s still implicitly held by many people. AI models can now think (by communicating and solving problems), which brings this viewpoint into question. If thinking can be performed by a physical process in AI, presumably it can be performed by a physical process in human minds too.

People still view themselves as having an “essential self” which includes their thoughts and goals, but AI will affect this perception. People will become more plugged into AI over time, initially through devices (such as smart glasses) and eventually through more direct brain-computer interfaces. AI will enhance a person’s thinking abilities and help them with their goals, and perhaps people will extend their sense of self to include the AI. But trying to include an external service within the sense of self will weaken the whole concept. If an external changing service is part of the self, the self cannot be that absolute. This can eventually lead more people to recognize that none of their thoughts define them, regardless of where they arose from. People will realize that the self is more flexible than assumed, and perhaps lead many to the Buddhist conclusion that there is no essential or separate self, just different processes interacting.

People and society generally assume there’s true free will - we have something inside us that chooses what to do, separate from the regular chain of physical cause and effect. People get angry at a human who does wrong in a way that they wouldn’t get angry at an animal that does wrong, since they assign free will to the human. As AI models act more independently, will people get angry at the model when it does something wrong and feel gratitude towards it when it does something right? People don’t think an AI model could have done otherwise, but maybe they will start assigning it moral responsibility anyways. If so, this could help spread compatibilist views of free will where moral responsibility is assigned even if everything is determined. And if people don’t assign moral responsibility to the AI models (which will seem very similar to humans) they may change how they assign responsibility to humans too.

Related story:

In July they examined existing tests to create time estimates for range of subjects, and found a similar trend:

While AI is able to do longer and longer tasks, people’s attention spans may be declining. For leisure activities, there are many reports that people are no longer reading full-length books, they’re watching short video clips and even playing shorter video games. For workplace tasks, Gloria Mark gathered evidence that people stay focused on one screen for decreasing amounts of time:

(The chart is from her book Attention Span, image taken from this LessWrong post, see further discussion there.)

If it’s true that people’s attention spans are getting worse while AI’s “attention spans” are getting better, we have even less time left before AI surpasses us…

• Google Announces Learn Your Way, which turns textbooks into interactive content personalized for the student.

• Meta announces some kind of AI glasses with a display and wristband to detect your gestures.

• Open AI releases an updated coding model, which is now better at refactoring code, among other things.

• Both Google and OpenAI announced gold medals at the 2025 ICPC World Finals, one of the world’s top programming competitions. AI has surpassed humans at solving coding challenges, what’s next?

Predictions

• Epoch AI predicts “that AI for scientific R&D will follow in the footsteps of coding assistants for software engineers today, boosting productivity for desk-based research by 10-20%”

• An essay in Colossus magazine argues that while AI will transform the economy, the opportunities for investors will be limited, and the main benefits will go to customers of AI.

Current usage

• Anthropic reports on how people and businesses are using Claude. They find people are now delegating complete tasks to Claude, which countries use it the most (Israel is in the lead), and much more.

• OpenAI reports on how people use ChatGPT, see also this Substack post. Their full paper includes charts such as the following usage stats: