RT @EdelmanBen: 1/ Excited to share a new blog post from the U.S. AI Safety Institute! AI agents are becoming increasingly capable. But th…

@RylanSchaeffer Woah what were you running for 46 days? 👀

@sea_snell I expect 2. very roughly predicts nonzero APPS Pass@1 for Llama 3.1 405B zero shot--have you evaluated it on apps as well as this language modeling loss score to check your prediction?

The U.S. AI Safety Institute is hiring technical staff! We're seeking talented engineers, scientists, or security experts for frontier testing, technical guidance, and more—these roles are only open for a couple more days, so apply or DM recommendations!

RT @alxndrdavies: Jailbreaking evals ~always focus on simple chatbots—excited to announce AgentHarm, a dataset for measuring harmfulness of…

RT @ben_s_bucknall: Excited to share our new website on Open Problems in Technical AI Governance, a companion to our recent paper on the to…

RT @AISafetyInst: Jade Leung (our CTO) and @geoffreyirving (our Research Director) have been nominated in the @TIME top 100 most influentia…

RT @BrandoHablando: Gave a 60 min lightning talk at Stanford's AI Safety Annual Meeting on our work identifying novel factors that explain…

@BogdanIonutCir2 Possibly, though I'd expect directly weaponizing dangerous capabilities is easier than repurposing those capabilities for safety. A crux is whether one expects AI systems will be very useful hackers or capability researchers before they're very good safety/alignment researchers.

@maxwellazoury @DanHendrycks That might be a useful property for certain models to have--imagine hardened models you want to share with pre-release testers or deploy in not-very-secure datacenters but not be as worried about unexpected harms from malicious finetuning if the model leaks to bad actors

@MotionTsar So it tends to assume coins are fairer than they actually are? This is acceptable 🙃

RT @The_JBernardi: 🚀 New blog! Achieving AI Resilience: Exploring AI safety through a lens of adaptation & societal resilience. Advanced A…

RT @Kurz_Gesagt: Humanity's smartest invention might also be its last. Superintelligent AI could be our dream come true – or our worst nigh…

RT @METR_Evals: How well can LLM agents complete diverse tasks compared to skilled humans? Our preliminary results indicate that our baseli…

@inner_treasure Quite possibly! I do acknowledge this in the post, but I also expect the most significant principal component accounts for a lot of the variance such that "general capabilities" is a useful enough thing to talk about, at least for the purposes of this piece

@nitarshan @AISafetyInst Congrats, and thank you for your service!🫡

🔗Consider reading the full post for many more details, references, considerations, and AI governance ideas. How do you conceptualize the progression of AGI development? Blog post: Discussion forum: 15/15