I’ve noticed an uptick in writing about human thought processes using the language of machine learning and neural networks. Jargon-y phrases like “improve your mental search space” or “if at first you don’t succeed, improve your gradient descent”. (I only made up one of these sentences!)
Normally, I’m all about this kind of metaphorical bridge from one domain to another. But in a lot of these cases, I feel like more direct language would benefit both the discourse on language models and using them to approximate human intelligence.
The people who make LLMs have little discernible interest in cognition itself. Some of them may believe that they’re interested in cognition, but what they’re really focused on is product — that is, output, what gets spat out in words or images or sounds at the conclusion of an episode of thinking.
While it’s spectacular that science has found success in modeling human intelligence with matrix multiplication, I think the humans engaged in science would benefit by avoidance of reducing human thought (intelligence, emotion, art, the whole gamut) down to GPU-friendly arithmetic.
A lot of critics think I’m stupid because my sentences are so simple and my method is so direct: they think these are defects. No. The point is to write as much as you know as quickly as possible.
— Kurt Vonnegut