David Masterson <[email protected]> writes: > Hmm. You've probably also seen this, but... > > OpenAI has proven that LLMs have a fundamental problem -- they lie and > their lying is getting more pronounced in the newer models. The basic > problem is they are trained to *not* say "I don't know" because saying > that would break the foundation of their business plan. Something to > incorporate in your draft... > > https://www.science.org/content/article/ai-hallucinates-because-it-s-trained-fake-answers-it-doesn-t-know
That's not a fundamental problem. Just a benchmark problem. Training methods are evolving. The more recent benchmarks specifically test RAG performance or grounding to the facts in general. The field is evolving quickly. While 6-7 months ago I would not ever trust to any citation claims made by LLMs, getting actual correct citations (including specific paragraph used) is a standard practice for the newest LLMs. -- Ihor Radchenko // yantar92, Org mode maintainer, Learn more about Org mode at <https://orgmode.org/>. Support Org development at <https://liberapay.com/org-mode>, or support my work at <https://liberapay.com/yantar92>
