3 ms·
How is an AI supposed to work that out? Given the talk about how these models have a tendency to confabulate, this would seem problematic. Idk though, maybe thi
by greenbit 2y ago
How is an AI supposed to work that out? Given the talk about how these models have a tendency to confabulate, this would seem problematic. Idk though, maybe this is now a thing?
- gruez 2y agoNot prompt engineer, but can't you ask it to answer yes or no to "Are these two comments vaguely about the same thing"? I'm not sure where the confabulation can happen, aside from misclassification which can be solved by getting a human and/or more LLMs to review the decision.
- greenbit 2y agoInteresting, did not know these things could be used to make assessments, judgments, categorizations, etc. Had been under the impression they just riff out text based on suggestions.
- gruez 2y ago>Interesting, did not know these things could be used to make assessments, judgments, categorizations, etc. Yep. For instance in this paper[1] they asked an LLM a bunch of political questions to figure out what its political alignment is. [1] "Discovering Language Model Behaviors with Model-Written Evaluations" https://www-cdn.anthropic.com/e4f69aacd8c0905030172bc6eb480c252ea7d6ad/model-written-evals.pdf https://www-cdn.anthropic.com/e4f69aacd8c0905030172bc6eb480c...
- indigo0086 2y agoThe diff is there, so the original answer exists and it seems like they prepared for this so they can easily refrain it with the original answer, or this is all idle speculation and they took it into account already. Being at a company that uses ai, data scientists know the problems and probably have discussed it at some length and covered these cases