Paper: AI Doesn't Say What It Thinks. AI Orgs: It Could Be Your Friend
Added 2025-04-14 16:19:25 +0000 UTCA fantastic new Anthropic paper shows that LLMs seem hell-bent on obfuscating why they gave an answer, even when it would be easier to be honest.
Download: https://drive.google.com/file/d/12jgVrGZLtVC_8DGTczQIzCVWPIBrnA0y/view?usp=sharing
Paper: https://assets.anthropic.com/m/71876fabef0f0ed4/original/reasoning_models_paper.pdf
Post: https://www.anthropic.com/research/reasoning-models-dont-say-think
FT Exclusive: https://x.com/FT/status/1910545751119135199
Coconut Paper: https://arxiv.org/pdf/2412.06769
Noam post:https://x.com/polynoamial/status/1910379351759347860
‘Deep Misgivings’: https://tech.yahoo.com/ai/articles/openais-sam-altman-deep-misgivings-093754634.html
OG Paper: https://arxiv.org/pdf/2305.04388
Comments
do you have a general stance on not wanting to share your data for training? for example to you use the private chat feature to prevent any record how do you decide the cost of sharing data? or is is as simple as you review the ai so you don’t want them looking at your information.
Drew Rogers
2025-05-22 23:58:21 +0000 UTCFor sure not true infinite memory, that's just marketing
Philip
2025-04-27 11:22:24 +0000 UTCPhilip, I don't think they are claiming it as "infinite memory". Where di you see that? That's something completely different from my understanding.
Kol Tregaskes
2025-04-21 11:49:08 +0000 UTCWhen Claude 3.7 thinking in Cursor is unable to get unit tests working after many tries, it will regularly create a simulated version of the production code in the test file and test that instead, blatantly circumventing the process. When I ask it to take a look at its work, it immediately recognizes the problem. I do find that being less imperative in my prompts ("fix these tests" vs "these tests are failing") make cheating behavior less likely.
Ryan Smith
2025-04-16 20:22:23 +0000 UTC
