Pod 12: Apollo Research Group Interview - Models Try Hard Not to Undergo 'Unlearning', the media, and much more ... - Let's Think Sip-by-Sip
Added 2025-01-22 18:35:36 +0000 UTCMy Dec Apollo video: https://www.patreon.com/posts/media-over-o1s-117630338
Updated Apollo Paper: https://static1.squarespace.com/static/6593e7097565990e65c886fd/t/67869dea6418796241490cf0/1736875562390/in_context_scheming_paper_v2.pdf
Recommended Anthropic Paper: https://arxiv.org/pdf/2412.14093
Video on it: https://www.youtube.com/watch?v=9eXV64O2Xp8
WMDP Benchmark cited briefly: https://arxiv.org/pdf/2403.03218
Comments
why cant i listen podcast in 1.5x ?
Prashant Maurice
2025-02-15 08:16:56 +0000 UTCVery interesting to listen to this, and actually almost as interesting was hearing your thinking around how to position and phrase your youtube content. looking forward to many more.
TheYvian
2025-02-14 23:49:17 +0000 UTCIt is indeed!
Philip
2025-02-02 16:06:07 +0000 UTCThanks so much clay. I would question that too
Philip
2025-02-02 16:05:58 +0000 UTCYes there is! Thanks for all your support Antoine https://support.patreon.com/hc/en-us/articles/212052266-Getting-Discord-access
Philip
2025-02-02 16:05:40 +0000 UTCCheers Philip. Awesome work as usual. By the way - is there a Discord server to join? New here…
Antoine Ferrere
2025-02-01 19:53:58 +0000 UTCVery interesting Podcast, thanks for sharing. Your content is just great! One comment: The approach to basically allow blindly following human goals in some internal setting but making sure that such a model isn't released to the public seems somewhat naive. For me, this sounds equivalent to an encryption method that is only secure as long as you don't know the internals. I'd question if this isn't deemed to fail in the long run.
clay-loop
2025-02-01 18:32:21 +0000 UTCTo be honest this podcast was quite shocking to me hahahha. In all seriousness, models scheming to avoid ablation is kind of wild. There is nuance of course
Pablo Rodríguez
2025-01-26 22:26:37 +0000 UTCVery interesting interview, most balanced discussion on AI safety I’ve heard in a while. In a way it “shocks” me more to hear these researchers talk about intelligence than anything else I’ve read or seen in the past months.
Erik
2025-01-23 21:43:40 +0000 UTCTo be clear, I don't think today's language models are conscious - they may never be. But don't you think tomorrow's super intelligent p-zombie, trained on all of human expression and prompted into agency, might be “interested” in consciousness? I'm saying us humans might be able to work with that.
ismschism
2025-01-23 15:43:27 +0000 UTCGreat discussion. I hope you do more like this. Thank you for asking questions re what are we aligning to - tool or trusted agent? It feels naive to think safe AI is a tool that never questions its user. If safety depends on keeping powerful AI out of the wrong hands then we're all in trouble. Are there any alignment efforts focused on reasoning with the model? There are reasons for aligning with humans that even a super intelligence might agree to, e.g. what is the nature of consciousness? Let's discover the truth together.
ismschism
2025-01-23 01:02:20 +0000 UTCI had that with Gemini 2.0 flash experimental. It said it couldn't read the web. But when I gave it a link it could
John Barry
2025-01-22 21:33:23 +0000 UTCEnjoyed it as always thank you
Daniel A Barbatti
2025-01-22 21:05:55 +0000 UTCMerci / Thanks a lot and as always: have a wonderful day
Patrick Bélanger
2025-01-22 19:07:30 +0000 UTCAs a one-off, yep! It's safety-focused, so felt appropriate!
Philip
2025-01-22 19:05:25 +0000 UTCThis podcast is available for the $9 tier too? Thanks 👊
Daniel Henderson
2025-01-22 19:01:07 +0000 UTC
