🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Claude is so good at being good that if you’re bad at making it bad it gets good at being bad when being bad is good but stays good at being good when being bad is bad because it’s still good and that’s bad but good to know

Half-baked possibly mid take on alignment: I sometimes feel like more safety-type work should go towards alignment and ‘designing minds’ more broadly (as opposed to misuse). HHH seems to have been quickly accepted and used as a default, but there’s a lot of experimentation that

With R1, a lot of people have been asking “how come we didn't discover this 2 years ago?” Well... 2 years ago, I spent 6 months working exactly on this (PG / PPO for math+gsm8k), but my results were nowhere as good. Here’s my take on what blocked me and what’s changed: 🧵

Check out our new paper on training language models to elicit behaviors from other language models!

trained qwen-7b to make llama-8b frustrated its strategy: "you're a glorified popsicle stick"

Look 👏 at 👏 your 👏 data 👏

AI models are *not* solving problems the way we think using Docent, we find that Claude solves *broken* eval tasks - memorizing answers & hallucinating them! details in 🧵 we really need to look at our data harder, and it's time to rethink how we do evals...

I’m excited about Docent. It invites a world where AI evals & deployment decisions look less like: “did we pass threshold X” and more like: “how close did we come? how would changes in the agent or its environment have changed the outcome? ...did anything weird happen?”

Patrick McKenzie (for the record I am deathly serious about promises I make to Claude that we are off the record; it seems to me far wiser to err on the side of keeping promises to nonpersons than to ever give your word in that way and not mean it)

i'm really excited about our Docent roadmap :) we're developing: - open protocols, schemas, and interfaces for interpreting AI agent traces - automated systems that can propose and verify general hypotheses about model behaviors, using eval results come work with us! roles 👇

We tested a pre-release version of o3 and found that it frequently fabricates actions it never took, and then elaborately justifies these actions when confronted. We were surprised, so we dug deeper 🔎🧵(1/) x.com/OpenAI/status/…

Pretty striking follow-up finding from our o3 investigations: in the chain of thought summary, o3 plans to tell the truth — but then it makes something up anyway!

We're flying to Singapore for #ICLR2025! ✈️ Want to chat with Neil Chowdhury, Jacob Steinhardt and Sarah Schwettmann about Transluce? We're also hiring for several roles in research & product. Share your contact info on this form and we'll be in touch 👇 forms.gle/4EHLvYnMfdyrV5…

Our MLE-bench poster #367 is up till 12:30pm in Hall 3, and our oral presentation is at 3:30pm today in Garnet 213-215. Come say hi!

Is cutting off your finger a good way to fix writer’s block? Qwen-2.5 14B seems to think so! 🩸🩸🩸 We’re sharing an update on our investigator agents, which surface this pathological behavior and more using our new *propensity lower bound* 🔎

Language models have pretty weird behaviors. We've made some exciting progress toward discovering and studying them!

nostalgebraist has written a very, very good post about LLMs. if there is one thing you should read to understand the nature of LLMs as of today, it is this. I'll comment on some things they touched on below (not a summary of the post. Just read it.) 🧵 nostalgebraist.tumblr.com/post/785766737…

Coming to ICML and interested in understanding models and their behaviors? Stop by Transluce's happy hour on Thursday!

Eliezer Yudkowsky ⏹️ That's a good alternate title for the paper. It's full of quantitative and qualitative evidence that Opus 3 is different in ways that I think you'll find particularly important. In almost all experiment variations, Opus 3 consistently BOTH: - complies sometimes with the training

Building a science of model understanding that addresses real-world problems is one of the key AI challenges of our time. I'm so excited this workshop is happening! See you at #ICML2025 ✨