Whose AI conversation is this?

On hype, doom, and the case for thinking it through yourself. Mental models are the way forward.

Something has been bothering me about the AI discourse, and I want to put it down here.

Look at who is shaping the conversation. The CEOs of the major AI labs have spent the last few years making confident predictions about what their technology will do, when AGI will arrive, and how governments should respond. In January 2025, Sam Altman wrote that “we are now confident we know how to build AGI” and that AI agents would “join the workforce” that same year, materially changing the output of companies. Elon Musk predicted AGI by 2025 in May 2024, then revised to AGI by 2026 in late 2025, and is now talking about end of 2026. Dario Amodei has said up to half of entry-level white-collar jobs may disappear within five years, that everyone should use AI to adapt, and that governments should tax AI revenue to redistribute the resulting wealth.

I don’t think any of this is bad faith. They probably mean what they say. But what they say also happens to be very useful to them. Optimistic AGI timelines bring in investors and talent. Predicting that AI will eliminate jobs makes the technology sound more powerful, and easier to sell. Calls to tax or regulate AI often end up protecting incumbents from smaller competitors. The trouble is that we can’t easily tell where conviction ends and incentive begins, and they may not be able to tell either.

The situation would make us uncomfortable in any other domain. Imagine if pharmaceutical CEOs were the loudest public voices on all health matters, on which professions their drugs would render obsolete, and what tax structure should govern the resulting profits. We would find this strange. In AI we treat it as normal.

The most prominent voices on the other side are no more neutral, though for different reasons. For more than two decades, the most vocal AI doomers, Eliezer Yudkowsky chief among them, have predicted catastrophic outcomes that have not arrived. In a 2023 TIME piece, Yudkowsky wrote that the most likely result of building superhumanly smart AI was that “literally everyone on Earth will die,” and called for international agreements to bomb noncompliant data centres. He has placed the probability of human extinction from AI at over 95%.

I don’t doubt that he believes this. I think he very much does. But as the years have passed without anything resembling these scenarios, the tone has only grown more confident and more vivid. Yudkowsky’s 2025 book, co-authored with Nate Soares, is titled If Anyone Builds It, Everyone Dies. I find this kind of certainty dangerous. If you truly believe something is an existential threat, then almost any action against it starts to look justified (including, apparently, bombing data centres).

In most fields, extraordinary claims that fail to materialise lead to revised estimates and quieter language. In AI doomerism the language has only grown louder. It is hard to take such predictions at face value when those making them do not appear to update on their own evidence. They just keep postponing.

The temptation for anyone is to pick the source you find most credible and outsource your thinking to them. Most people I talk to do something like this. They follow particular journalists, or particular CEOs, or particular skeptics, and they form their views by absorbing that person’s takes over time. It is efficient, but it is also a form of intellectual surrender.

I think the most reliable way out of this is to use AI yourself. Regularly, on real work. As a serious practice over weeks and months, on tasks that actually matter to you.

When you do this, things start to happen that no commentator can tell you. You discover that the model is astonishingly good at certain tasks you assumed it couldn’t handle. You discover that it falls flat on tasks everyone said it had mastered. You understand the shape of what is working and what is not (call it “jagged” if you must). You develop intuitions for when to trust it and when to verify. You stop being impressed by announcements.

Perhaps most importantly, you also stop being moved so easily by other people’s framings. The hype loses its grip because you can see where the demo ends and the hard work begins. The doom loses its threat because you can see what these systems cannot do, and how often. You can also see all the wonderful options that suddenly open up. These things are very powerful, after all! You see the speed at which things are advancing. You may find yourself disagreeing with everyone in particular ways, but this is what having your own view actually feels like.

Clarity comes from doing the thing yourself, repeatedly and attentively. The good old Feynman quote “What I cannot build I do not understand” comes to mind. It is more useful than any clean position on whether AI is good or bad.

Now, I want to be careful here. What I am saying could sound like: ignore everyone, figure it out alone. That is certainly not what I mean! Doing this alone is hard, and most people who try give up within a week or two. Most of us have an actual job, after all. Perhaps you sit down with a chatbot, you don’t quite know what to ask, you get an answer that looks plausible but you can’t tell if it’s any good, and the easy path is to close the tab and go read someone’s thread about AI instead. Or you keep hearing that AI is really good at coding now, but you can’t really test that for yourself because you have never coded, and wouldn’t even know where to start, let alone interpret the results.

So even if you are using AI regularly, you might still not get the full picture. And let’s admit it: this thing is too big, too generalist, for any one of us to evaluate it, especially as a side project in our lives.

Mental models

This is where I believe mental models come in. Mental models are ways of thinking about what is in front of us that hold up across different situations and stay useful, even as the technology, the products, and the services change.

A good mental model is durable in a way that a hot take is not. A hot take about which job will be automated next quarter is probably useless within months; but a clear understanding that a job is best seen as a bundle of tasks - and that technology often changes the tasks without eliminating the job - stays useful for years. A hot take about the latest model release becomes almost irrelevant when the next one arrives, usually within weeks; but a first-principles grasp of what these systems are actually doing when they generate text helps you interpret every release that follows. In other words, the half-life of a take is short; the half-life of a mental model is long.

This is also how we can stop being pushed around by the loudest voices. Once we have a few good models in our head, we do not need to track every new AI announcement or every fresh argument from a CEO or a critic. You can hear a claim and immediately tell whether it fits what you know about how these systems work, or whether it is hand-waving. The doom and the hype both rely on us not having that grip. Models give us the grip.

I have spent the last few years writing pieces like this one, and I have tried, in many of them, to offer something like this. I think some of it has worked. The feedback was generally quite positive, often the kind of “huh; I haven’t thought about it this way”. The truck driver and radiologist piece was an attempt at a mental model for thinking about jobs and automation, and our inability to predict the impact of technology on the labor market. The pieces on AI sovereignty were an attempt at a model for thinking about technological dependence. I want to do more of this work, not less, and I want to do it in a different place than a newsletter. By its very name, a newsletter should be about news. Mental models are not news.

That is why I am stopping the newsletter. Helping people build durable ways of thinking about AI matters more than ever, and I want to put my energy into doing that properly, in a place that doesn’t depend on you sitting in front of a screen, and in a way where the value accumulates rather than scrolls past. More on that very soon. If you are a paid subscriber, that will affect you directly, but I will have a good solution to offer, again very soon.

If you have made it this far, you are probably someone who wants to think carefully and independently about AI from your own experience. So in the meantime, I would encourage you to do two things. Open one of these tools today and use it for an hour on something that actually matters to you, then do it again tomorrow. And start collecting the mental models that help you make sense of what you are seeing. You will notice the difference quickly.