What does AI believe is true?
How can I infer the beliefs of a neural network in an unsupervised way?
That question motivates Burns et al. to propose the Contrast-Consistent Search (CCS) method in their work "Discovering Latent Knowledge in language models Without Supervision".
Timestamps:
00:00 - Discovering Latent Knowledge in Language Models Without Supervision
02:08 - Problem definition: Discovering Latent Knowledge
03:30 - Method: Contrast-Consistent Search (CCS)
08:35 - Experimental Setup
09:37 - Finding: CCS is robust to misleading prompts
11:52 - Finding: Truth is a salient feature
12:28 - Related work on zero-shot prompting and truthfulness
13:15 - Limitations
13:52 - A few closing thoughts
14:51 - Further resources
Topics: #interpretability #ai #llm #liedetection #artificialintelligence
Link to the paper: https://arxiv.org/abs/2212.03827
Notebook on Github: https://github.com/collin-burns/disco...
Forum post: https://www.alignmentforum.org/posts/...
Critique: https://arxiv.org/abs/2307.00175
For related content:
Twitter: / samuelalbanie
Research lab: https://caml-lab.com/
personal webpage: https://samuelalbanie.com/
YouTube: / @samuelalbanie1
TikTok: / samuelalbanie
Instagram: / samuelalbanie
LinkedIn: / samuel-albanie
Threads: https://www.threads.net/@samuelalbanie
Discord server for filtir: / discord
(Optional) if you'd like to support the channel:
https://www.buymeacoffee.com/samuelal...
/ samuel_albanie