This is probably because of the autoregressive nature of LLMs, and is why “step by step” and “chain of thought” prompting work so well. GPT4 can only “see” up to the next token, and doesn’t know how its own entire answer upfront.
If my guess is correct, GPT4 knew the probabilities of “Yes” or “No” were highest amongst possible tokens as it started generating the answer, but, it didn’t really know the right answer until it got to the arithmetic calculation tokens (the 0.9 * 500 part). In this case it probably had a lot of training data to confirm the right value for 0.9 * 500.
I’m actually impressed it managed to correct course instead of confabulating!
![[Chatlog] Wait, actually, yes](https://i.redd.it/2g2p54zrziob1.jpg){kind=link}
This is probably because of the autoregressive nature of LLMs, and is why “step by step” and “chain of thought” prompting work so well. GPT4 can only “see” up to the next token, and doesn’t know how its own entire answer upfront.
If my guess is correct, GPT4 knew the probabilities of “Yes” or “No” were highest amongst possible tokens as it started generating the answer, but, it didn’t really know the right answer until it got to the arithmetic calculation tokens (the 0.9 * 500 part). In this case it probably had a lot of training data to confirm the right value for 0.9 * 500.
I’m actually impressed it managed to correct course instead of confabulating!
deleted by creator