THE ILLEGIBLE PARTS OF THOUGHT
- Jun 30
- 2 min read
You might have missed it. But our ability to place fine electrodes into living, awake human brains is changing how we study intelligence.
On 17 June, a team at Massachusetts General Hospital reported, in Nature, recordings from 579 single neurons in eight awake patients as they produced natural speech. The work mapped cellular building blocks by which language is encoded in humans, using large language models to help draw its conclusions.
The findings showed neurons firing a fraction of a second before you utter a noun. Different ones for verbs. Others tracked parts of speech, grammatical relations, phrase structure and sentence context – all before you speak, below conscious awareness and beyond anything you could explain.
Feed the same sentences into a large language model, and its representations predicted the neural activity better than models of grammar or meaning alone. Predictive power improved with more of the preceding sentence, then plateaued at around five preceding words. The authors read this as evidence that neurons carry context before speech.
Our understanding of neural networks helped us build digital intelligence. Now digital intelligence is shaping how we understand our own.
But language models helping us understand how brains produce language is just the preliminaries. Language is externalised intelligence. Already compressed. Written down over thousands of years, then fed to machines as an archive.
Other parts of intelligence haven’t been captured so neatly: one-shot learning that changes what we see after a single exposure (Hachisuka et al, Nature Communications); the world model behind a judgement, where identical input branches into competing internal hypotheses (Voigts et al, Nature Neuroscience); confidence forming with a decision rather than after it (Vivar-Lazo and Fetsch, Nature Neuroscience); and a single exposure to a new environment becoming a stable spatial map (Wen et al, Nature).
None of this leaves the kind of archive a sentence does.
A language model is trained on the export of intelligence: words left over after perception, memory, hesitation, social judgement and editing have done their work. The internet is what intelligence has said, not the record of how it reached the point of saying it.
Underneath the hype, what comes next may depend as much on studying intelligence as on building it.
For those who want to indulge in a little light neuroscience reading.
