My intuition is that the interesting similarity isn’t “parameters plus an algorithm,” but the fact that in both cases a large body of observations is replaced by a smaller generative structure capable of making predictions.
We group apples and oranges together not because they are both made of atoms, but because they share enough structure and function that we call them fruits. My intuition is that scientific laws and LLMs share more than just a computational substrate—they both act as predictive compressions of large bodies of observations.
Do you think that similarity is still too generic, or is there a more fundamental difference between scientific laws and LLMs that I’m overlooking?
One quick question: what do you mean by the “[dead]” part? Are you referring to the article link itself or to something in the argument?
That would be a fascinating experiment. If an LLM trained only on pre-Newton or pre-Einstein material independently arrived at something resembling those theories, it would be strong evidence that powerful new representations can emerge from existing information.
That’s an interesting point. Brains and LLMs are certainly very different in their implementation. My argument, though, is less about the implementation itself and more about the fact that both systems are capable of generating and manipulating representations.