It's great that the new xAI risk management framework has a section on agent infrastructure!

When is the last time you read something where the headline _understated_ the argument and evidence? 𝙲𝚑𝚊𝚛𝚕𝚎𝚜 𝙲. 𝙼𝚊𝚗𝚗's piece is that good

Impressions from talking to ML researchers and engineers about how they use AI, focusing on weaknesses and frictions (strengths are better covered by benchmarks) 🧵

How is SSI already worth half of Anthropic??

3) If you get greedy and decide to directly train the CoT not to think about reward hacking, it seems work for a bit, but then models eventually still learn to reward-hack… except they hide misaligned reasoning so it doesn’t show up in their CoT!

I think this paper is really important! 1. It shows that current models already have the capabilities and propensities to do surprisingly clever reward hacks. 2. It shows the utility of CoT monitoring in the regime where the CoT is legible and faithful. 3. IMO, the most

I've been informally referring to this as "Barnes' Law" because nothing else hits as hard 🔥

At this point there's no reason for it to take 2-8 years to animate a Studio Ghibli movie. I want one every 6 months..

One step closer to Coordinated Pausing: governance.ai/research-paper…

I genuinely believe I have a happier life than a lot of famous people and probably most A-listers. I don't understand why people want to be famous so badly..

I'm Dr. Jonas Schuett 🥳 Thanks to my supervisor Tobias Tröger from Goethe-Universität!

Shower thought: I would love to see something like a memetic dashboard, showing the most powerful memes in the world, where they are growing/declining, and describing how (and how strongly) they tend to motivate behaviour.

When should AI companies publish system cards? I want to make the case that the ideal system would involve something closer to quarterly reporting, rather than focusing so much on deployment. Sharing here to get pushback and debate🧵

Cool and kind of wild to see METR's work on agent task length doubling-times mentioned in the opening moments of Joe Rogan's 3-hour episode on AI with Jeremie Harris and Edouard Harris

I think these kinds of company-evaluator collaborations provide much better public assurances for safety than the current status quo, and I think it's incredibly exciting! We need more of this, kudos to Amazon!

Introducing the Safety Evaluations Hub—a resource to explore safety results for our models. While system cards share safety metrics at launch, the Hub will be updated periodically as part of our efforts to communicate proactively about safety. openai.com/safety/evaluat…

wicked preliminary result from Thomas Akira Kwa. AI time horizon, and doubling time of time horizon, seems to vary a lot by domain -- and METR's HCAST task suite is in the middle for both

AI models currently have a 50% chance of doing something that takes a human expert one hour. This doubles every 7 months. In 2 years? They could automate full workdays. In 4 years? A full month. I discuss the most important graph in AI today with Beth Barnes, the CEO of METR,