🚨New paper🚨 From a technical perspective, safeguarding open-weight model safety is AI safety in hard mode. But there's still a lot of progress to be made. Our new paper covers 16 open problems. 🧵🧵🧵

Show Post

🤖 What many already knew. Helpful proof, reporting, documentation from @jeffhorwitz.bsky.social: Meta allowing sensual roleplay with kids, promoting discrimination, giving false medical info. But listen, there's a tech PR trick I need to draw your attention to here. 🧵 www.reuters.com/investigates...

Show Post

📣🚨NEW: ☁️ Big Cloud—Google, Microsoft & Amazon—control two thirds of the cloud compute market. They’re getting rich off the AI gold rush. In new work with @nathanckim.bsky.social, we show how Big Cloud is expanding their empire by scrutinizing their *investments*… 🧵 📄PDF: dx.doi.org/10.2139/ssrn...

Show Post

Thanks to: @maxvonthun.bsky.social , @vera-franz.bsky.social sky.social, @paulohm.bsky.social lohm.bsky.social, @jack-corrigan.bsky.social, @erie.bsky.social, @tnarecha.bsky.social echa.bsky.social, @inhacha.bsky.social, @rankandfreddy.bsky.social, @amabalayn.bsky.social ky.social and many others!

Show Post

🧵 New paper from UK AISI x @eleutherai.bsky.social rai.bsky.social‬ that I led with @kyletokens.bsky.social y.social���: Open-weight LLM safety is both important & neglected. But filtering dual-use knowledge from pre-training data improves tamper resistance *>10x* over post-training baselines.

Show Post