18 ms·
Are people still experiencing llms getting stuck in knowledge and comprehension loops? I used them but not excessively, and I'm not heavily tracking their perfo
by clejack 1y ago
Are people still experiencing llms getting stuck in knowledge and comprehension loops? I used them but not excessively, and I'm not heavily tracking their performance either.
For example, if you ask an llm a question, and it produces a hallucination then you try to correct it or explain to it that it is incorrect; and it produces a near identical hallucination while implying that it has produced a new, correct result, this suggests that it does not understand its own understanding (or pseudo-understanding if you like).
Without this level of introspection, directing any notion of true understanding, intelligence, or anything similar seems premature.
Llms need to be able to consistently and accurately say, some variation on the phrase "I don't know," or "I'm uncertain." This indicates knowledge of self. It's like a mirror test for minds.
- ramchip 1y agoLike the article says... I feel it's counter-productive to picture an LLM as "learning" or "thinking". It's just a text generator. If it's producing code that calls non-existent APIs for instance, it's kind of a waste of time to try to explain to the LLM that so-and-so doesn't exist. Better just try again and dump an OpenAPI doc or some sample code into it to influence the text generator towards correct output.
- thomastjeffery 1y agoThat's the difference between bias and logic. A statistical model is applied bias, just like computation is applied logic/arithmetic. Once you realize that, it's pretty easy to understand the potential strengths and limitations of a model. Both approaches are missing a critical piece: objectivity. They work directly with the data, and not about the data.