One thing which came to mind is that perhaps "safety protocols" stop chatgpt from mentioning Fischer because of his controversial opinions. If you have ever tried to steer a llm to discuss controversial topics it gets very squirmy and micro-hallucinations start to occur pretty densely.
So on the topic of Chess960, it might make stuff up to avoid mentioning fischer, then it doubles down going forward, probably because LLM has programing to trust previous answer in order to keep responses coherent.
> This is all a nonsense. I suspect that Cowen saw how well o3 performed in his field of economics, and wrongly generalised to assuming that it is reliable in other areas.
The thing is that Tyler Cowen knows a ton about chess. He was New Jersey State chess champion. If he can be fooled by AI on this topic then almost anyone would be.
The fact that GPT/o3 then cited cowen for the claim ..... fantastic in every sense
He said that he writes for LLMs, it's really true.
One thing which came to mind is that perhaps "safety protocols" stop chatgpt from mentioning Fischer because of his controversial opinions. If you have ever tried to steer a llm to discuss controversial topics it gets very squirmy and micro-hallucinations start to occur pretty densely.
So on the topic of Chess960, it might make stuff up to avoid mentioning fischer, then it doubles down going forward, probably because LLM has programing to trust previous answer in order to keep responses coherent.
Just some guesses based off of observations.
Smart people can be the most foolish.
> This is all a nonsense. I suspect that Cowen saw how well o3 performed in his field of economics, and wrongly generalised to assuming that it is reliable in other areas.
The thing is that Tyler Cowen knows a ton about chess. He was New Jersey State chess champion. If he can be fooled by AI on this topic then almost anyone would be.
Despite my scepticism around LLMs, I must confess that I'm impressed by their ability to produce fabrications.