An AI leaderboard suggests the newest reasoning models used in chatbots are producing less accurate results because of higher hallucination rates. Experts say the problem is bigger than that
I get that. We want them to be creative and make up an eMail for us. Though I don’t think there is any fundamental barrier preventing us from guiding LLMs. Can’t we just make it aware whether the current task is reciting Wikipedia or creative storywriting? Or whether it’s supposed to focus on the input text or its background knowledge? Currently we don’t. But I don’t see how that would be theoretically impossible.
They don’t think. They use statistical models on massive data sets to achieve the statistically average result from the data set.
In order to have increased creativity, you need to increase the likelihood of it randomly inserting things outside that result: hallucinations.
You cannot have a creative “AI” without them with the current fundamental design.
I get that. We want them to be creative and make up an eMail for us. Though I don’t think there is any fundamental barrier preventing us from guiding LLMs. Can’t we just make it aware whether the current task is reciting Wikipedia or creative storywriting? Or whether it’s supposed to focus on the input text or its background knowledge? Currently we don’t. But I don’t see how that would be theoretically impossible.