distributed machine learning @ google | sometimes mathematician
Seattle
Joined September 2012
- Tweets985
- Following437
- Followers1.4K
- Likes1.4K
Zachary Charles retweeted
Tinker’s Dimuon is Muon with delayed diagonal conditioning, Polar Express and a rank-dependent update norm. I recovered its update rule through the API, borrowing @mrtnm’s approach to StableReinforce. Here’s how 🧵
Dimuon is a new optimizer we built for LoRA fine-tuning. In our SFT tests, it often reaches lower loss in fewer steps than AdamW. We're still experimenting with Dimuon; you can Tinker with it today and share your feedback!
See our optimizers guide at tinker-docs.thinkingmachines…
Zachary Charles retweeted
I agree with this and adopted this stance unilaterally on a recent review. I encourage you to adopt it in your reviewing — it’s reasonable to write reviews that say “this is dense unreadable slop, my review isn’t detailed and I recommend rejection solely on that basis.”
Some thoughts on reviewing: spylab.ai/blog/reviewing/
First for @satml_conf where we trialed an early reject phase which cut submissions by ~40%.
Then more generally about how not everything deserves peer-review, and why we should reward writing things that people actually read.
Now's actually the best time to submit great work to TMLR - good work shines through!
What is TMLR doing under a 300% increase in submissions?
- Suspensions of >150 authors
- Publicly naming major violations
- Submission quotas distinguishing slop patterns
- Desk rejections 6% → 53%; focus on clarity
- AI reviews evaluating soundness
blog.tmlr.org/2026/adapting-…
Zachary Charles retweeted
People really have no idea what string theory is, most have never talked to a string theorist nor bother to read up on it. Its public perception is driven by remarkably confident statements constructing a bogeyman that has nothing to do with the actual research
Things frustrating me about the math discourse:
1) There's a wide variety of beliefs around AI in the math community right now. Many people are excited for the OAI results, many are not.
2) It seems quite normal (even, dare I say, human) to express some amount of consternation if you believe your job security or your students' job security is at risk. This doesn't seem unique to mathematicians at all!
When my dad watched a movie he didn't really like, he didn't want to say anything negative so he'd say things like "The sound effects were great!"
Not its biggest impact, but maybe this will finally get coding agents to stop putting imports in random places.
I think this is correct - it's actually pretty wild how many important problems are apparently solvable by (mostly, though not exclusively) existing, albeit often recent, techniques!
At least some of these OAI math results build off others from the same batch. Cool to see that their models are already seeing benefits from having this enormous database of new results!
Every time I see discourse about talent versus hard work, I'm reminded about how most of the naturally gifted math grad students I knew burnt out and dropped out, and the ones I saw succeed all worked incredibly, almost unreasonably, hard.
Terrific paper - an all-time classic IMO.
We can debate how realistic the PL condition is, but I believe this paper is "there are more things in heaven and earth Horatio than are dreamt of in your philosophy" but for people who stare at convex functions all day.
Congratulations to @MarkSchmidtUBC, Hamed Karimi, and Julie Nutini on the ECML PKDD 2026 Test of Time Award for their work on the Polyak-Lojasiewicz condition!
I spent a lot of time studying this paper during my PhD. The convergence analysis is easy to follow even if you are not an expert in the area, and the writing is a masterclass in clarity.
At the time, the paper showed me how much clear presentation matters when bringing technical ideas to a broad audience.
cs.ubc.ca/news/2026/09/test-…
I had a coworker who believed that everyone should only get to send 3 emails per day. More and more I think they're being 100% correct.
arXiv has updated our policy on rate limiting for all submitters.
This update was made to fairly distribute moderator time & support the arXiv community of staff, volunteers, readers & authors.
Please read our announcement to learn more: blog.arxiv.org/2026/10/01/up…
ALT arXiv logo in white over a whiteboard with math symbols.
Conway had it right. There's a reason math has seen so much AI-driven progress. It's not easy, but it has entirely explicit rules!
"You know, people think mathematics is complicated. Mathematics is the simple bit. It's the stuff we can understand. It's cats that are complicated. I mean, what is it in those little molecules and stuff that make one cat behave differently than another, or that make a cat? And how do you define a cat? I have no idea."
"Most domains are fundamentally pretty shallow," says @RyanGreenblatt. A concise statement of what the superintelligence worldview gets wrong, IMO. My view is the opposite: many domains are surprisingly deep, and AI won't be able to "quickly pick up a bunch of domains."
Zachary Charles retweeted
If you want evidence that capitalism is a force for good, just visit a grocery store.
My ode to the humble American supermarket, for @WSJFreeEx
How much of writing style is determined by a kind of bundling of what you've been exposed to? And can you mimic this with models that have seen as much text as AIs have?
Also some people really just have weird writing styles that you only grow to appreciate over time. Unclear how to recreate this here if you're trying to gain users.
