in more niche problems, it will often hallucinate that there’s a magic parameter that does exactly what you want. It will never tell you “Nope, can’t be done, you have to restructure around doing it this other way”
This was why, in spite of it all, I had a brief glimmer of hope for DeepSeek – it’s designed to reveal both its sources and the process by which it reaches its regurgitated conclusions, since it was meant to be an open-source research aid rather than a proprietary black-box chatbot.
Anthropic’s latest research shows the chain of thought reasoning shown isn’t trustworthy anyway. It’s for our benefit, and doesn’t match the actual reasoning used internally 1:1.
As you say, hallucinating can be solved by adding meta-awareness. Seems likely to me we’ll be able to patch the problem eventually. We’re just starting to understand why these models hallucinate in the first place.
deleted by creator
This was why, in spite of it all, I had a brief glimmer of hope for DeepSeek – it’s designed to reveal both its sources and the process by which it reaches its regurgitated conclusions, since it was meant to be an open-source research aid rather than a proprietary black-box chatbot.
deleted by creator
deleted by creator
Anthropic’s latest research shows the chain of thought reasoning shown isn’t trustworthy anyway. It’s for our benefit, and doesn’t match the actual reasoning used internally 1:1.
deleted by creator
As you say, hallucinating can be solved by adding meta-awareness. Seems likely to me we’ll be able to patch the problem eventually. We’re just starting to understand why these models hallucinate in the first place.
deleted by creator