Challenges with unsupervised LLM knowledge discovery · HackerPapers discussions

Challenges with unsupervised LLM knowledge discovery — mentioned by westurner

Sun, 07 Apr 2024 00:46:26 GMT

"Language models can explain neurons in language models"; GPT-4 can help explain what GPT-2 neurons do: https://openai.com/research/language-models-can-explain-neur... https://news.ycombinator.com/item?id=35877402 IDK if this is a practical limit to such research? From https://news.ycombinator.com/item?id=38753725 : >> "Challenges with unsupervised LLM knowledge discovery" (2023) https://arxiv.org/abs/2312.10029 : >> Abstract: We show that existing unsupervised methods on large language model (LLM) activations do not discover knowledge -- instead they seem to discover whatever feature of the…

Paper and discussion history. A mention is not an endorsement.

Challenges with unsupervised LLM knowledge discovery — mentioned by westurner

Sun, 24 Dec 2023 14:41:13 GMT

"Challenges with unsupervised LLM knowledge discovery" (2023) https://arxiv.org/abs/2312.10029 : > Abstract: We show that existing unsupervised methods on large language model (LLM) activations do not discover knowledge -- instead they seem to discover whatever feature of the activations is most prominent.

Paper and discussion history. A mention is not an endorsement.

Challenges with unsupervised LLM knowledge discovery — HN submission

Sun, 24 Dec 2023 14:38:13 GMT

Challenges with unsupervised LLM knowledge discovery

2 points; 3 discussion replies on HN. Paper and source history. A submission is not an endorsement.