after a long time being busy with real work, some video game posting: I really like The Outer Worlds 2. Much better than the 1st one, imo better than Starfield. The writing is both funny and not cringe, which imo no one since Fallout got right.

Interesting, but seemingly incompatible with token-level adversarial attacks being easy?

Link here:

quick plug: i was trying to introduce my non-super-technical biologist friend to AI tools and the ergonomics were pretty bad so I build an RStudio-MCP that lets Claude Code connect to it, be aware of the environment, plot things, etc. <Link below>

Claude's disappointment is immeasurable and its day is ruined

New research with @AdtRaghunathan, Nicholas Carlini and Anthropic! We built ImpossibleBench to measure reward hacking in LLM coding agents 🤖, by making benchmark tasks impossible and seeing whether models game tests or follow specs. (1/9)

today’s dreams were finally good - I was watching a long anime about ASML’s corporate development when my brain rudely woke me up not bad after 4 days of nightmares

Just caught myself thinking “ugh, the AI couldn’t even compute this cohomology group correctly.” OK, admittedly things have come pretty far, pretty fast.

sf event norms: i was super late to a lunch party, after a stressful uber ride i arrived 50 mins late. turns out i was the first one there

also i kinda don’t understand meta’s game here. they seem to be pushing ai-generated shortform vids, but me and everyone i know hates seeing them on insta. am i that much in the bubble? is the bet that we’ll eat what they feed us? are engagement metrics really better for ai vids?

why tf haven’t apple or google shipped hardware attestation for photos or videos yet

gm

my mom has been watching a lot of chinese tv recently and is now explaining omegaverse to me wtf is happening

what's the best llm coding multi-agent orchestration product rn? https://t.co/C6Je6Aq5uZ looks fun but i need something that runs on linux for my workstation, not just macos

Our benchmark:

hmm

best guess - the prev intersection is set up so that they arrive at the next one just as the green turns to yellow. so they instinctively stop. then when the light turns red, they had a second to look around, see no cars, and start walking

there are a couple intersections i hit when biking in sf where pedestrians seem to regularly wait on green and then start walking on red, and for the life of me i can't figure out what's going on.

Even with full-batch gradients, DL optimizers defy classical optimization theory, as they operate at the *edge of stability.* With @alex_damian_, we introduce "central flows": a theoretical tool to analyze these dynamics that makes accurate quantitative predictions on real NNs.