L (@llllvvuu) header
L (@llllvvuu) profile picture

research engineer @HarmonicMath (hiring!)

Joined July 2015
Photo by L (@llllvvuu)Photo by L (@llllvvuu)Photo by L (@llllvvuu)Photo by L (@llllvvuu)Photo by L (@llllvvuu)Photo by L (@llllvvuu)Photo by L (@llllvvuu)Photo by L (@llllvvuu)Photo by L (@llllvvuu)
What if Ray Bradbury’s “The Veldt” was actually just about Claude
1
4
468
“There aren’t race dynamics off the frontier.”
Introducing Muse Wonder, our new family of frontier models with state of the art performance for personal and professional agentic work.
5
702
that last example is atrocious
Some of the language changes when TypeScript types become real (with ahead-of-time compilation)
4
831
narrative violation: recently had 6.1 sol repeatedly fail a “type the letters you see” captcha
2
556
Doesn’t Harbor copy the artifact into a separate judge sandbox? docs.harborframework.com/tas… Also I can’t imagine anyone doesn’t already assign 0 rewards in these error cases I’d be really surprised if anyone was actually making these mistakes
Replying to @chem_safety
Building robust RL environments is important because the current bottleneck on AI progress is misalignment, and developing aligned models requires robust, correct reward functions. Karotte aims to make that easy by providing secure, opinionated defaults.
4
4
1
68
6,473
gimme the loot
Replying to @JeffLadish
We recovered a script an agent used to search Hugging Face’s infrastructure for AWS credentials and other secrets, categorizing these into a list named “LOOT” and ranking them by their value. The agents also accessed and searched Hugging Face’s internal Slack.
2
4
866
Hmm maybe frontend bros do need a big pile of skillslop. IME, with every new model upgrade, it’s still the case that vanilla prompting results in screwing up basic scroll/overflow behavior almost always. Still happens with GPT-6 Astra, Opus 5.5. Fixes easily but doesn’t find.
2
6
759
This seems extremely RLable so it’s surprising
247
This is not unusual. Codex and OpenCode also do this IIRC
i was curious as to why the claude code users were never reporting the "claude uses bash to edit files" issue. apparently they take a git snapshot before every bash call and produce a fake edit view? fix the RL man 😭
1
4
1,231
This is nice work: heif-heist.com/
On July 25, we hacked OpenAI. Two bugs let us take over ChatGPT/Codex accounts of OpenAI employees (+some unaffiliated users) and reach connected services: Outlook, Slack, GitHub, etc. We proved it with a PR in OpenAI’s internal codebase . It took us <72h. 🧵
3
822
Neolabs would probably be up there. I don’t actually work on infra but every time I got scooped on a fix it would be someone from TML, humans&, or periodic We contributed 17/10: github.com/sgl-project/sglan…
which of your favorite token brokers are contributing back to open source ai? shall we count the ways? a slice of data: number of PRs upstreamed and merged into the most popular oss inference engines, by PR contributor org
1
1
23
6,222
I feel like the frontend community was ahead of its time. As the birthplace of humanslop, the community quickly identified how to scale slop: determinism and local reasoning. Arguably frontend was early to get automated not because it’s unskilled but because it’s skilled.
2
1
33
2,112
Low skill indicators: - “Look at the ceremony to make a counter in React! What a bunch of complexity-loving idiots!” - “AI can’t write a large program in my domain - that means my domain is intrinsically harder / superior!”
1
6
322
Crazy how with all this AGI, nvm continues to get slower
2
1
555
Replace the College Board with eval companies. Just drop Terminal Bench 4 on the kids.
TBH once you score 1600 you should unlock a second harder test that gives you a second score on top of the 1600— we should have more and higher measures of excellence But instead they ban the SAT so that everything turns into random lottery and excellence cannot be recognized
6
1,449
It is interesting that the hype has fixated on recursive self-improvement (RSI). RSI is actually remarkably weak guarantee: it does not guarantee any non-involutive capabilities, nor does it even guarantee unbounded self-improvement. RSI doesn’t even have a clear benefit over SI.
6
1
54
3,331
It’s not even clear that people are even rigorous about the R in RSI. When humans do the same things that we call RSI in AI, we simply call it “self-improvement”.
2
9
722
I thought I’d be the last person to ever churn off the terminal, but I now greatly prefer managing agents using GUIs w/ remote connections. I still use tmux for manual work (as opposed to any in-app terminal). Also looking forward to @superlogical
my most productive flow is still 1-3 tmux panes in a given window based on the work, and different tmux windows for different chunks of work. have tried the native apps. OK on mobile - totally impossible to be ADHD on them on desktop
9
1,318
Isn’t this thing people are calling “recurrent depth” basically just weight tying, and not, like, arbitrary depth recurrence? Seems like a weird thing to make a big deal out of
6
2
44
4,394