@alz_zyd_@gaussianReverie A cool exercise that is super neat: consider the set of all polynomial functions. Encode every polynomial as a vector of its coefficients, then come up with a matrix that computes its derivative. Now find the eigenvector of this matrix, and compare to the Taylor series of exp(x)!
@alexolegimas@alz_zyd_@danielrock Well, it produces a long sequence of tokens, which can be seen as iteration. There is evidence that LLMs “plan ahead” (https://t.co/UXs6vFA22Q) with their tokens. Also, maybe they cannot go back on their produced tokens, but they can reconsider their approach during reasoning.
@itsandrewgao One obvious I could think of is that chess has a very clear reward function (its easy to verify correctness), while 'reasoning' does not really right?