@NoahZiems @lateinteraction Where’s the list of candidate puns, Omar
English
Drew Breunig
15.7K posts
Drew Breunig
@dbreunig
Writing about and working on AI, geo, and data.
Bay Area Katılım Mart 2008
1.1K Takip Edilen8.9K Takipçiler
This project went through more naming iterations than anything else: magnet RL (actively attract the rollout to the right final answer), lucky RL (stumble upon good + realistic rollouts more often than pass@K), foresight RL (learn to use special knowledge of the future), ...
Souradip Chakraborty@SOURADIPCHAKR18
🚨Typical RL algorithms and on-policy distillation methods are blind samplers: they use privileged info to score rollouts, but not to *find* them. We ask: can we use privileged info to *actively sample* the rollouts RL wishes it can stumble upon with compute? ⤵️ Pedagogical RL
English
Great teachers craft demonstrations their students could have built themselves.
Souradip Chakraborty@SOURADIPCHAKR18
🚨Typical RL algorithms and on-policy distillation methods are blind samplers: they use privileged info to score rollouts, but not to *find* them. We ask: can we use privileged info to *actively sample* the rollouts RL wishes it can stumble upon with compute? ⤵️ Pedagogical RL
English
The new flex is how many exploits Mythos finds, and how long it takes to find then.
Mythos found one MacOS exploit in 5 days.
Only a single curl bug was identified.
Andrew Curran@AndrewCurran_
Mythos has cracked MacOS. It took five days.
English
English
English
From Theory, @motherduck, and @lancedb: we salute you @AICouncilConf, and the conference attendees. If you missed our boat party, we hope you’ll join us in the future.
🫡
English
English
English
Live by the fuzzy, natural-language thinking engine, die by the fuzzy, natural-language thinking engine.
Nick Schrock@schrockn
As a business practice it is sensible and fine imo to subsidize only your own harness/apps but the unenforceability of it technically is makes this all very funny to watch.
English
I am! Here's a drop for dspy.RLM's code interpretter: github.com/dbreunig/dspy-…
Isaac and I have been discussing letting the code interpreter provide instructions to the RLM module, which would save a turn or two and prevent the model from trying unsupported imports. But since they added `json` support, the speed of execution is worth the 2-3 burned turns discovering the edges of monty.
English
Should I try out Monty for the RLM REPL I’ve been hearing more and more about it recently
Auctor@auctor
Blog: The Agent Is a Workflow That Writes Itself How we productionized the RLM with durable execution. Subagents lower to child workflows and PTC runs through a deterministic workflow-space interpreter. Every tool, subagent, and PTC call goes through a single recursive dispatch loop. Closure under replay, retry, cancel. We call these durable agents.
English
3 years since ChatGPT, hundreds of billions invested, models capable of writing entire codebases and finding novel exploits...
And the emdash persists!
English
@tetsuo_cpp @badlogicgames I'm going to be there's at least 10 incomplete versions on GH right now.
English
@badlogicgames Whatever happened to that Rust slop rewrite of Pi anyway?
English
I'm finding paper and pen, marker and whiteboard, to be multiples more valuable in this AI era.
It's a different space to think, differently, without the easy chatbot escape hatch.
English
Drew Breunig retweetledi
Some awesome initial experiments on training small RLMs :)
A direction I think will be super super important moving forward for fully seeing the capabilities of RLMs vs. traditional agentic systems
alphaXiv@askalphaxiv
Reinforcing Recursive Language Models Can a 4B model learn to recursively call itself to answer hard long-context questions? We RL fine-tuned a small model to behave as a native RLM. On evidence selection across scientific papers, our 4B RLM matches Sonnet 4.6 in quality while running significantly faster and cheaper.
English
Super excited to join the 2026 cohort of @KnightHennessy scholars!
Has been incredible talking to the other scholars about the major problems they are tackling across healthcare, science, policy, etc. Excited to work towards Human-Centered Open-Source AI that can support people tackling the world's biggest challenges!
KnightHennessy@KnightHennessy
Meet the 2026 cohort of KH scholars! These 87 new scholars make up the most global Knight-Hennessy Scholars cohort to date, and will pursue degrees in 45 graduate programs across all seven graduate schools at @Stanford: knight-hennessy.stanford.edu/news/knight-he… (1/2)
English
Drew Breunig retweetledi
Great to see @percyliang as a keynote speaker at #cais2026!
ACM Conference on AI and Agentic Systems@CAISconf
🎤 Keynote announcement: @percyliang (Percy Liang), Professor of Computer Science at @Stanford, founding director of the Center for Research on Foundation Models, and co-founder of @togethercompute, is keynoting #CAIS2026. Percy's HELM framework set the standard for holistic evaluation of language models, and his Foundation Model Transparency Index (now in its third year) put every major AI lab on notice for what they do and don't disclose. His current work on Marin takes this further: an open lab where every experiment, successful or not, is public from day one. This one's going to be good. San Jose · May 26–29 caisconf.org
English
Drew Breunig retweetledi
🎤 Keynote announcement:
@percyliang (Percy Liang), Professor of Computer Science at @Stanford, founding director of the Center for Research on Foundation Models, and co-founder of @togethercompute, is keynoting #CAIS2026.
Percy's HELM framework set the standard for holistic evaluation of language models, and his Foundation Model Transparency Index (now in its third year) put every major AI lab on notice for what they do and don't disclose. His current work on Marin takes this further: an open lab where every experiment, successful or not, is public from day one.
This one's going to be good.
San Jose · May 26–29 caisconf.org
English
@badlogicgames @vboykis imo the front page has generally been balanced but the comments are vocally anti.
English
@vboykis please do! i'd be very interested in the outcome.
English
AI is what Donna Haraway called a "power object": something we treat as enormously important without understanding how it works.
Reverence plus naivety means everyone fills the empty box with their own pet issue.
Kevin Van Valkenburg@KVanValkenburg
This reader comment on a NY Times column where Ross Douthat ponders that maybe God is speaking to us through A.I. is an absolute fastball on the corner with movement, and deserves a column. "Lightening was once mysterious too; mystery did not make Zeus correct" is perfect.
English
Drew Breunig retweetledi
This reader comment on a NY Times column where Ross Douthat ponders that maybe God is speaking to us through A.I. is an absolute fastball on the corner with movement, and deserves a column. "Lightening was once mysterious too; mystery did not make Zeus correct" is perfect.
English
Needless to say, I think Caldotcom going closed source for security reasons is exactly the wrong move. x.com/dbreunig/statu…
Drew Breunig@dbreunig
English