🔥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

Modern indie videogame genres ranked by how interesting I expect them to be as AI agent evaluation environments (assuming no spoilers): A: Roguelike, Factory Builder, Colony Simulator, Survivors-like B: Metroidvania, Metroidbrania F: Soulslike, Extraction Shooter, Battle Royale

Synopsys, famous for chip design engineering software (EDA), said it has suspended financial guidance for the 3rd quarter and full-year fiscal 2025 after receiving a letter from the US government’s Bureau of Industry and Security (BIS) related to new export restrictions on China,

The San Francisco Bay Area has been so wonderful to explore the last 6ish years, and I’m so grateful to get to be experiencing my life in such a beautiful and fun place. (Not moving, just appreciating)

First time I can remember Dwarkesh supporting specific policies: - Tentative support for 10 year block on state AI legislation - Streamline datacenter construction - Expand energy capacity - Reform liability to limit liability exposure for AI systems - Broad deregulation

It’s both inspiring and depressing when a change in personnel/leadership seems to improve things so quickly

When someone says “I’m not at all confident of X” I sometimes want to say “Cool, just to check, you understand this means if X happens, when we’re tallying up correctness points later on, people who say ‘I’m somewhat confident X will happen’ will get more points than you, right?”

I’m enjoying SemiAnalysis doing more of these explanatory Twitter threads. They’re interesting. I don’t remember them doing as many of these 2+ months ago

The last couple years I’ve begun thinking of truth as a kind of social consensus Schelling point that intelligent people rely on because they know social structures built around true arguments/ideas/logic will be undeniably persuasive or compelling to other intelligent people.

At METR, we’ve seen increasingly sophisticated examples of “reward hacking” on our tasks: models trying to subvert or exploit the environment or scoring code to obtain a higher score. In a new post, we discuss this phenomenon and share some especially crafty instances we’ve seen.

Lawrence Chan response to the "Illusion of Thinking" paper!

AI agent performance on HCAST & RE-Bench seems to ‘plateau’ as agents are given more ‘time’ to do tasks. The best humans, on the other hand, seem to have less obvious plateaus. Some thoughts on this🧵

This feels like the most information dense thing I've read yet about the Israel/Iran situation

Why try to smuggle Nvidia chips into China when you can just smuggle training data out? Incredible WSJ report: wsj.com/tech/china-ai-… Liza Lin Raffaele Huang

delighted to announce that RE-Benchwarmers, the METR et al soccer team, achieved a 1-3 defeat last week. extrapolating out, we see that we can expect a positive goal difference starting next season, after the break

If there are very convincing arguments that your work is important, and there are good structural reasons to think it won't be adequately compensated by the market, choosing that work, over well-paid alternatives, awards you some "counterfactual moral agency points" in my book

I think something like this intuition is why I’m still confused about the Ege Erdil and Matthew Barnett position on “broad labor automation is more important than automated AI R&D” Sorry I’m probably mis-summarizing you guys

At this point believing that "Anonymous" is a real, specific group of hacktivists feels like believing in the Tooth Fairy, but for people whose worldview and politics froze in time somewhere around when the Occupy Wall Street protests happened