Professor @UMich EECS, Ex-Google, Ex-Voleon. Research on the Foundations of ML+RL+LLM
Joined April 2021
- Tweets380
- Following326
- Followers1.4K
- Likes1.3K
Samet Oymak retweeted
Many amazing [if correct] results close to my heart / work:
1) Polyakov's conjecture on the O(n) model
[Polyakov's was the first qft paper I read]
2) Quantum annealing for SK model.
3) Reconstruction on trees
[Related to conjectures made with Marc Mezard 20 years ago].
4) k-SAT threshold
[Beautiful work by Gaia Carenini has priority here]
5) Free energy for spin glasses on random graphs.
6) Mixing for the SK model.
7) Factor-of-IID threshold for Ising model on trees
8) Free energy of the spherical perceptron
github.com/openai/math
Samet Oymak retweeted
I'm sorry i know people are like not into papers and stuff, but you gotta check this out
Test-time communication looks like a next axis for scaling capabilities
New paper with the incredible @jon_ghoh and @vkontonis @ShivamGarg91462 and Akshay : arxiv.org/pdf/2609.21032
The Hugging Face incident showed when agents can find a channel they'll use the heck out of it. A useful question, I think is: when does communication make a group MORE CAPABLE than the same agents working alone?
Aka is Team-of-N better than Best-of-N, when, and why?
We had N identical agents work on the same task with no prescribed roles, using only a shared log (i.e, text file) and telling them to "collaborate". Across three "researchy" tasks communicating teams beat the heck out of independent agents:
- On ARC-AGI-3, a Team-of-5 sonnet-4.6 agents matches Best-of-33, and can for example solve a game 65% of the time that no single agent cracked in 64 tries.
- On polyomino packing (pack Tetris like pieces into the smallest rectangle, cf Frontier-CS by @eigenlabs), a Team-of-3 Opus 4.6 agents surpasses best-of-60 and set, as far as i understand, a new record for that benchmark.
- On MNIST compression, a team of four 5.6-Sol agents find a 1,957 byte model with 99.4% accuracy, which btw is 20% smaller than the best human solution (on a problem beaten to death!!), while no independent agent gets below 3KB.
The mechanism is a bit obvious in hindsight: when one agent finds a clearly better partial solution, it broadcasts it, and everyone immediately builds on it.
Why? A single lonely agent must make every breakthrough itself, yet a team needs each insight only once, found by any member.
That is kinda like comparing a minimum of sum of "time to n-th breakthrough" vs a sum of minimum of "time to n-th breakthrough". That gap can grow exponentially with the number of "breakthroughs" needed to arrive at a solution.
We worked on this because prior work (before the hf incident) suggests unclear benefits for communicating aganets. Which is true, when the tasks are inherently serial (duh), eg some Terminal bench style tasks. Yet feels it should not be true for research problems.
Indeed for research heavy problems... Test-time communication seems like a new capabilities axis.
I'm sure we will see a ton more of it!
Samet Oymak retweeted
Today, we’re announcing Ternary Bonsai 2 27B.
Based on Qwen3.8 27B, Bonsai 2 27B is 9x smaller than its full-precision counterpart while retaining 98.2% of its aggregate benchmark performance.
Two months after the first Bonsai 27B release, the biggest change is quality. The footprint remains 5.9 GB, but the gap to full precision has narrowed materially, with particularly strong gains in agentic coding, multimodal reasoning, and long-horizon tool use.
Ternary Bonsai 2 27B is available today under Apache 2.0.
Samet Oymak retweeted
I sometimes see people trying to use complexity theory to argue that building "true" AI is impossible. I find this unreasonably annoying. It requires ignoring what is in front of your face and it ignores that worst-case complexity has been an awful guide in machine learning.
Samet Oymak retweeted
If Dario wants to slow down the pace of AI, he should just move the company to Europe.
Samet Oymak retweeted
I assign essentially 0% probability that Anthropic or OpenAI will slow down frontier development in any meaningful way.
And that majority of the "regulation" they try making is regulatory capture to slow down competition.
1. Anthropic/OpenAI will not willingly hand over their lead to Gemini, xAI, Meta, or another frontier lab.
-> Elon will want a slowdown to catch up with OpenAI.
-> OpenAI will want a regulatory framework to slow down xAI + competition.
Every lab has an incentive to support rules that constrain competitors more than they constrain itself
2. The US is not going to meaningfully slow frontier AI development while China continues advancing. (as seen with Trump's post today about rejecting AI slowdown).
It would be probably the most flagrant national security risk if China leapfrogged the US in capability.
Especially when both sides have called frontier AI capabilities "cyber nuclear weapons"
3. China will not slow down frontier AI development. That is a given.
It will be theater if any of these parties agrees so publicly, because behind the scenes:
-> They'll likely keep on progressing, whether it's OpenAI, the US Government, or China.
If they wanted to slow down, they could have done so themselves at any point in time:
Yet everyone is still fighting over the scarce resources to run AI as seen with memory procurement from Samsung going all the way to 2031.
Or the $NVDA / $AVGO supply constrained 70-100%+ Y/Y growth protections.
So I suggest people unfollow the clowns on X that predict AI stocks from $NVDA to $TSM to $MU will crash for engagement.
All this does is cause retail to panic short term and hand over the keys to the kingdom over to others.
Samet Oymak retweeted
1) open source cant be "banned"
2) did not read the rest
Samet Oymak retweeted
Lots of disdain in the comments. It’s honestly a bit unnerving to read people’s negative reaction to a harmless plea to continue prioritizing sharing discoveries and results, as humanity has done for millennia. Maybe it’s easy to think that the authors are operating from a place of ego and jealousy, if they are Fields medalists, and therefore the general public doesn’t trust them. So let me try explaining as someone who has no career achievements, no stake in the frontier labs, and no jealousy because my peers are all early-career like me 😆
Academics are stewards of history. We have 3 jobs. (1) to train people / teach them what has been done in the past. (2) we make some discoveries ourselves, and (3) we regularly interact with our peers to learn about their discoveries; what tomorrow’s history may be. Many people think we just do (2), but in fact all 3 are important; otherwise we would be highly inefficient in problem-solving over the long term (centuries, millennia).
We cannot really make AI the stewards of history (there are capitalist and geopolitical incentives not to). If we rid the world of academic values, there may not be immediate consequences, but things will be very bad in the long term. Progress will stall. Eventually folks will realize we need a clear and continually-updating ledger of history (ie “memory” for the AI-pilled folks). It will take a lot of work to rebuild this culture from first principles
On the optimistic side I think academics have a golden opportunity to figure out how we can harness AI in pursuit of our values. We can probably keep more information in our history, disseminate info more broadly, review and validate each others’ discoveries faster, etc. But, given that academics are powerless compared to these trillion dollar companies, we can only do the best we can—speak up; discuss; educate the public about academic values (which we haven’t really had to do since ww2); and, honestly, pray
The amount of “schadenfreude” directed at mathematicians makes me concerned about how society will respond to a large-scale labor disruption. At least they (mostly) have tenure.
Since Tuesday, half of Twitter became experts on Millennium Prize problems. I consider that a win for math.
Samet Oymak retweeted
🧠 Asymmetric architecture. More intelligence, less cost.
🔹 552B-parameter MoE.
🔹 New Causal Encoder–Decoder architecture: just 8B active parameters for input, 16B for output.
🔹 New pre-training methods + larger-scale RL post-training deliver benchmark results ahead of flagship models, including DeepSeek-V4-Pro.
2/6
I’m all for supporting curiosity of undergrads and proving open problems with AI. But also sad to see Twitter experts attempting to ridicule Fields medalists 🤦♂️
😂 So Caltech undergrads wanted to, like barbarians at the gate, use their AI credits to destroy open math problems, but one thousand older, senior mathematicians, the established ones with Fields medals and honors under their belt were absolutely outraged that these tyros, these uneducated uncivilized barely self showering children would decimate their status symbols, raised a ruckus, and got those free credits rescinded.
Anime characters spreading rumors about AI proving Millennium Prize problems 🤦♂️ What a weird timeline!
Samet Oymak retweeted
The world of R&D is forking into two paths: the token-abundant research, and the token-starved research. The future is in the former - evidentially, the progress by today's top AI industry teams and neolabs is breathtaking, where researchers’ human brilliance is super charged by AI’s assistance. Every research university president should be reading this report and reflecting on what the future of higher education research should be.
openai.com/index/research-ac…
