Overtraining as the path to human-like AIThe anonymous blogger Gwern recently completed a thirteen thousand word post called Human-like Neural Nets by Catapulting, in which he offers a theory about why LLMs don’t possess truly flexible human-like intelligence, and how we might train LLMs that do. Theories like this are entirely unremarkable: every crank researcher on the internet has a theory about how to crack AI. But Gwern is remarkable. Outside of OpenAI itself, Gwern is the earliest person to anticipate the potential of large language models, and the scaling arms-race involved in making them larger and more powerful still. I often cite Leopold Aschenbrenner’s Situational Awareness as an example of someone correctly predicting the future of AI. Written in 2024, just after the release of GPT-4, Aschenbrenner gets a lot of things right: the rush to build billion or trillion-dollar GPU clusters, the importance of the code around the LLM (what he calls “unhobbling”)1, and the fact that scaling would continue through the decade. Gwern’s essay The Scaling Hypothesis anticipated the broad strokes in 2020, immediately on the release of GPT-3 (two years before the release of ChatGPT and the beginning of the AI boom).

And yet, as far as I can tell, Human-like Neural Nets by Catapulting hasn’t yet received much public attention: one recent Hacker News thread with twelve comments, all of which are about whether human brains are anything like neural networks. Part of the reason is that (a) it’s such a long post, (b) the potted summary describes Gwern’s claim, but not the reasons for it, and (c) much of the beginning of the post looks like it is indeed arguing from analogy with human brains. However, I don’t think that analogy is load-bearing. Let me try and explain what I think Gwern is saying.

What is grokking?

First, let’s talk about “grokking”. In 2022, OpenAI published a paper showing that if you train a model on a simple dataset (for instance, a simple mathematical operation like division), and keep training it long after the training looks like it’s stalled out, the model will suddenly make a massive jump in capability. Why does this work? The first stage of training is like rote memorization: the model has to compress as much of the training data as possible into its weights. But if you keep going, then regularization techniques (such as the pressure on the model to use smaller weight values) will motivate2 the model to find simpler and simpler ways of compressing the data. This doesn’t look like much at first (the training loss remains at zero), until the model notices that you can express the data via simply performing the underlying mathematical operation, at which point it instantly gets massively smarter. In other words, over-training a model can pressure it into actually understanding its training data. OpenAI named this process “grokking” after Robert Heinlein’s neologism, which for Heinlein means something like “gaining a deep, intuitive and fundamental understanding”3.

Gwern’s argument goes something like this:

Modern LLMs are worse generalizers than humans because they have not grokked their core domains

Grokking requires overtraining an over-parameterized model on a (relatively) small dataset, which is the exact opposite of what frontier labs do

However, (2) is basically how human brains learn

Somebody should spend a a few tens of billions of dollars3.5 on trying it, since it might immediately usher in truly human-like LLMs

I’ll skip (3), since I think the argument is still compelling without the analogy to human brains.

Are LLMs bad because they can’t grok?

I think his first point is hard to dispute. LLMs are very smart in specific areas, but they routinely make errors that humans wouldn’t make. More to the point, they routinely make errors that any human as smart as the LLM would never make. This pretty clearly points to a failure of generalization: LLMs are as strong as smart humans in specific areas, but can’t generalize that intelligence to as many tasks as humans can.

Do LLMs not grok? I read through this paper that argues they do. If you graph “how much data has the LLM memorized” against benchmark performance, you can see a small initial spike in benchmark performance, followed by a big drop, followed finally by a big jump in benchmark performance. This pattern doesn’t track memorization at all: memorization increases smoothly in the background the whole time.

I think this paper highlights the difficulty of distinguishing grokking from generalization. Obviously LLMs learn to generalize during training, and it’s plausible that learning to generalize would require a certain baseline level of memorization (so that the LLM has the raw material to generalize from). So it’s going to look like grokking.

When Gwern (and others) say that LLMs don’t grok, I think what they mean is that there’s at least one more giant generalization leap waiting to be made. Is this plausible? As an...