Samet Oymak

I was fired from Block today. I was the PM in charge of changing the default tip option on the Square terminal to start at 40%. Jack replaced me with an AI agent that decides which tip amount to show based on your age, weight, and race. If anyone is hiring product managers, please let me know!

Fantastic job by @DimitrisPapail! Two hypotheses: - Harder benchmarks likely have high variance and tricky to predict (like many models achieve near zero). - Expensive benchmarks with more problems may be more predictable due to finer difficulty granularity (so scaling laws show up). This makes IMO 2025 hardest to predict as actually observed :)

Our paper "SmartChunk Retrieval: Query-Aware Chunk Compression with Planning for Efficient Document RAG" is accepted at #ICLR2026 🎉 Static chunks fail for long docs,Let the model decide. We make chunking a reasoning problem, introducing STITCH, a new RL↔SFT post-training loop

We’ve pushed out the Pareto frontier of efficiency vs. intelligence again. With Gemini 3 Flash ⚡️, we are seeing reasoning capabilities previously reserved for our largest models, now running at Flash-level latency. This opens up entirely new categories of near real-time applications that require complex thought. It’s available in the API, and rolling out today as the default model in AI Mode in Search and Gemini app globally. Read more on the blog at: bit.ly/4pTo5YU More in thread ⬇️

All of reinforcement learning in 809 words. argmin.net/p/defining-rei…

I’m really excited about our release of Gemini 3 today, the result of hard work by many, many people in the Gemini team and all across Google! 🎊 We’ve built many exciting new product experiences with it, as you’ll see today and in the coming weeks and months. You can find it today on @GeminiApp and AI Mode in Search. For developers, you can build with it now in @GoogleAIStudio and Vertex AI. blog.google/products/gemin… The model performs quite well on a wide range of benchmarks.

Update: TabPFN-2.5 is actually the SOTA model on all of TabArena (which has datasets with up to 100k training data points). In a single forward pass, TabPFN-2.5 outperforms all other models, even if you tune them for 4 hours. We built and previously evaluated TabPFN-2.5 for up to 50k data points (and 2k features) and were kind of surprised that it's SOTA up to 100k 🙂 👉 TabPFN-2.5 webinar tomorrow: app.livestorm.co/p/21526e44-406… 👉 Model report on arXiv: arxiv.org/pdf/2511.08667

The arXiv position paper controversy and the weird, unwritten, organic evolution of academic practice. argmin.net/p/a-position-o…

(1) Our team at @GoogleDeepMind has been collaborating with Terence Tao and Javier Gómez-Serrano to use our AI agents (AlphaEvolve, AlphaProof, & Gemini Deep Think) for advancing Maths research. They find that AlphaEvolve can help discover new results across a range of problems.

Much needed step from arXiv for quality control and preventing blog post & survey dumping: blog.arxiv.org/2025/10/31/att…

@tydsh This is crazy and it's truly their loss

Several of my team members + myself are impacted by this layoff today. Welcome to connect :)

Learning is the derivative of knowledge.

This paper plane outside my office has been there since I joined U of M in 2023. Considering nominating it for tenure.

Honest question. What "scaling laws" theory papers that are not a variation on 1980s nonparametric statistics?

@docmilanfar Wow this is exactly what I was teaching in the class this week: Gaussian kernel smoother = Softmax Attention

How Kernel Regression is related to Attention Mechanism - a summary in 10 slides. 0/1

@DimitrisPapail 100% agreed. Today is the worst day SLMs will ever be 🙂

Small models as the new frontier and why this may be academia's LLM moment Academia should reject the nihilism of "scale is all you need", i.e, that meaningful research requires frontier scale compute. This mindset hurts basic research and what we can contribute to machine learning in practice. Many interesting questions about architectures, data, and training methods do show signal and can be tested at the O(100M) to O(1B) parameter scale within reasonable budget. There seems to exist no fundamental reason why these insights wouldn't transfer and hold up to 14B, 32B, or even larger models. Yes, there will be trends and observations that break at the trillion parameter scale, but my conjecture is that this will be irrelevant for the majority of models people will actually deploy locally in the future. The economics of post-training (SFT/RL) are finally favorable for academia. Post training a 7B model fits on a single H100 GPU, which roughly $3/hour on cloud providers. You can train on 100M+ tokens for under $100. Why care about mid/post-training? That's where a lot of interesting problems are! Reasoning, tool use, specialization, etc, these are settings where you see meaningful performance improvements and skills learnt within millions of trained tokens, not billions, that are typical for pretraining. More importantly, the 4B-32B parameter range will likely dominate local deployment in the not so distant future. These models fit on reasonable hardware (a beefy laptop) as inference requires enough RAM to fit the model, but you can use without GPUs for single batch inf calls. Also these models, at that scale, are getting seriously good for tasks likecoding, math, tutoring, computer use etc. So here is my conjecture: local models at the <100B scale will eventually generate more tokens/day than api-hosted frontier models. This may be academia's moment! The open-weights ecosystem provides a path to real impact without million-dollar GPU clusters at this scale. Our research can directly study, understand, and improve the 99% of models that will run locally, not the 1% that require data centers. This is finally both possible and meaningful. Don't be discouraged by scale maximalism!!

Our paper BREAD is accepted to #NeurIPS2025. BREAD can (provably) solve problems that are not solvable with standard SFT->GRPO training by interpolating between them: arxiv.org/pdf/2506.17211 Congrats to the lead authors @xuechen_zhang (on the job market!) and @HuangJ42673.

Congrats to @jackcai1206 and @nayoung_nylee for their NeurIPS Spotlight on "Extrapolation by Association: Length Generalization Transfer In Transformers" :)

Per reddit/ML (tinyurl.com/4am325bn), #AAAI2026 reviews and decisions are a mess. Seems overwhelmed by 30k submissions...