The Helmholtz Machine

Let's go back to the University of Heidelberg in the 1860s; to Hermann von Helmholtz, who arrived there in 1858. By that time Helmholtz was already a prolific polymath: a physicist, physician and mathematician who had become famous for measuring the speed of nerve impulses, something that had previously been thought impossible to do. For almost a decade he had been publishing in parts Das Handbuch der physiologischen Optik a treatise on how he believed human vision actually functioned. But this empirical work led to deeper obsession: how these experiments and investigations related to perception.

Helmholtz kept finding a gap: what people saw couldn't be explained by the raw sensory data hitting the retina. The signals were too degraded, too ambiguous, too incomplete. The brain, he said, had to be involved in the construction of perception. What is incomplete in what we see, Helmholtz suggested, was filled in by what he called 'unconscious inference'; the process by which prior experience is used to build a coherent picture of reality. What we see, Helmholtz suggested, is less a direct reading of the world than a reconstruction; what we might now call, in modern terms, a world model.

Helmholtz was arguing that perception is inferential: the brain reconstructs the likely causes of its sensory input, rather than passively registering them.

A hundred and thirty years later, Peter Dayan, Geoffrey Hinton and colleagues at Toronto decided to test this idea in a neural network. If we could build a system that implemented the German physicist's ideas, would it learn to perceive? In 1995 they released a paper in Neural Computation describing what he'd built: the Helmholtz Machine.

The paper's opening suggested this implementation would give us a framework for the brain itself: "Following Helmholtz, we view the human perceptual system as a statistical inference engine whose function is to infer the probable causes of sensory input."

The success of Hinton's work, and that of the field over the next thirty years, developed on this idea: that neural networks were a model of cognition. While Hinton's original Helmholtz Machine was explicit in its relationship to the brain (it had, for example, a 'sleep state') what followed stripped out the biological parallels in favour of engineering performance. The step change that birthed the AI scaling era was the advent of transformers, an innovation that improved the ability of neural networks to learn contextually. The next-token prediction systems millions of people use every day are not direct architectural descendants of the Helmholtz Machine, but they inherit its founding premise: that intelligence can be built out of a system that infers the causes of its inputs. And if these AI systems are modelling cognition, then there's no reason they might not model a form of cognition 'better' than ours; that they might not at some point surpass the limits of intelligence imposed by biology.

Hinton himself has come to this conclusion; that these cognitive architectures are close to matching, and will soon be better than, what humans do: "I have suddenly switched my views on whether these things are going to be more intelligent than us. I think they're very close to it now and they will be much more intelligent than us in the future."

So when researchers began to look inside the models they had built, those descendents of the Helmholtz Machine, they did so with an expectation: that a sufficiently complex predictive system comes to represent a model of reality: a world model. All they needed to do was find it.

To find it they went back to board games. Othello is like Go: the pieces are black and white, but the rules are a little different.

And it's where, four years ago, a group of researchers found compelling early signs of an 'emergent world model'

In late 2022 Kenneth Li, a researcher at Harvard, began working with his colleagues to train a variant of GPT to play Othello. The model was 8 layers and 20 million parameters, much, much smaller than the AI systems most people use every day. To teach Othello-GPT the game they did not show it the board, or explain the rules, they just fed it a sequence of tokens that represented legal moves in the game.

The results were impressive: Othello-GPT chose legal moves 99.99% of time.

And when Li and his colleagues decided to investigate what was going on inside Othello-GPT they discovered something amazing. Using nonlinear probes (small neural networks trained specifically to see what an AI model's 'internals' have inside them) they found that Othello GPT had a 'board state'. The model, it seemed, had developed a spatial representation of the board inside its weights. A world model.

And this world model was not just observable, it was causal. If the researchers modified the internal representation of an Othello square, Othello GPT's predictions changed with the new board state.

Neel Nanda, perhaps the most famous figure in Mechanistic Interpretability, followed up on the results, finding that the evidence for a world model was even cleaner still. When he used a linear probe trained on a simpler description of who had what pieces on the Othello board - mine, yours, empty - the decoding of the world model was even more elegant. Nanda framed this result succinctly: it was, he said, "pretty obvious that transformers can do something more than statistical correlations and pattern matching", it was "great to have clearer evidence of fully-fledged world models."

This gave rise to a whole new area of inquiry. Researchers reported analogous findings across other games: chess, poker, mazes, arithmetic, and then extrapolated those ideas out into frontier models like ChatGPT, Claude and Gemini.

These systems had shown, their authors argued, that even on simple tasks AI models develop internal representations of the structure of the task. To the researchers and commentators who followed the results, the Helmholtz Machine's promise was being fulfilled. The models were displaying a kind of 'intelligence'; not just the statistical correlation and pattern matching, but something more. Something like what we do, like what Helmholtz predicted; or perhaps, something even better.

"I'll see before I doubt; when I doubt, prove"
Othello, Act 3 Scene 3

So if, like me, you believe that the intelligence difference between humans and machines can be blurred in quantity but not in kind, you have to contend with these results.

You have to ask yourself: what is this world model, what characterises its behaviour, and is there another explanation for how we found it?