Why I don't think AI is a bubble · honnibal.devMost of the people I like think AI is a bubble. This is a tricky topic to discuss, because<br>the “bubble” framing couples financial and technical issues. It’s like a sports fan debating<br>“Is this player overrated?”. The answer depends on how good you think that player is, and<br>how good you think other people think they are.

I don’t have anything much to add to the financial part of the “AI bubble” conversation. Various<br>equity prices are based on very optimistic estimates about how AI will progress. This<br>post is about the technological question. I’ll leave it to you to judge what sort of<br>forecast any given asset price actually represents.

The main case I want to make is that performance probably won’t plateau — or at least, the<br>common arguments for why it will plateau don’t add up. These arguments have been mostly<br>the same for the last few years, and they’ve been consistently inconsistent with what’s<br>actually happened.

People mostly say that AI performance will plateau because they believe that performance<br>increases have been driven by brute-force scale. The argument goes that the industry has<br>been spending more and more money on training and operating larger and larger models. The<br>basic approach has diminishing returns, but there’s been exponential spending behind it,<br>producing what looks like linear progress — but this trend can’t last. Either some unknown<br>breakthrough bails the industry out, or advancements will halt at a point while the cost of<br>providing the service is still much higher than the value people would actually be willing<br>to pay for it — especially since from the consumer side, people are mostly buying in because<br>they expect the capabilities to improve, not because they see them as currently market-ready.

It’s not autocomplete

I also used to expect a plateau in capabilities, especially around GPT-1 and GPT-2. If I<br>only had a few sentences to say to myself then, here’s one way to put it. You know how there<br>were two big splashy lines of deep learning research: generative AI based on massive pretraining<br>(GPT etc), and deep reinforcement learning (AlphaZero etc)? Well, what people did is put those<br>together. It’s like the least surprising thing ever when you put it like that. But that’s actually<br>something importantly true about problem solving, which explains how the “putting them together”<br>works. For any answer you want, there’s a way to pose the question that makes the last leap really<br>obvious. How do you find that question? Well, there’s going to be some intermediate step that makes<br>the second last leap obvious. Each new text generated is like a move in a game, and the AI<br>“wins” the game if it gets to the right answer.

If you asked GPT-1 a question like “Do Berlin, London and Mumbai together have a greater population than<br>Australia?”, the model would start generating a reply that’s most similar to the sort of thing people would<br>say to that in the web text it was trained on. But what you could do is ask it separately questions like,<br>“What’s the population of Mumbai?”, “What’s the population of Australia?” etc. Assume for the sake of<br>argument it knows correct answers to all those questions. If you give it the output to all that, now it<br>will get the original question right.

By GPT-3, the behaviour is much harder to characterise, but we can say roughly it would generate a likely<br>useful reply from the web, and that might happen to be a plan that breaks it down step by step, which would<br>guide it to the right answer. But this would only work if it had seen questions asking about sets of numbers,<br>and seen the sorts of replies people give to that.

For Claude Opus and GPT-5, they’ve been trained to generate the intermediate questions themselves. In the same<br>way that AlphaZero learns to play chess, when the model succeeds in generating a sequence of steps that<br>leads it to the right answer, that gets reinforced. What it learns from this is the general process of breaking<br>down problems. What sort of steps can I try that might lead me in the right direction? If I’m trying to do an<br>operation like x + y + z > a, maybe I should fetch all the values, and then I can sum up x + y + z, and<br>then when I’m looking at the two numbers together right in front of me I can generate the answer.

Describing models like GPT-1 or GPT-2 as “fancy autocomplete” was<br>reductive but reasonable. For GPT-3 I’d call it directionally incorrect. For today’s models it’s<br>a serious misunderstanding that will lead you to make wrong decisions that have a concrete<br>impact on your life. I hope this post can help.

Credentials

I’ll explain my credentials on the topic briefly before I go into more detail. I published my<br>first paper on natural language processing in 2004, and did my PhD 2005-2009. I stayed in<br>academia 2009-2014, then left to write a popular open-source NLP library and found a small<br>startup. The stuff I worked on wasn’t in the main line of Large Language Model<br>(LLM) research, but...