Tim Lau (@timlautk) profile picture

AI Researcher @DRWTrading; Past Postdoc @Penn @PennMedicine @Wharton @ChicagoBooth; PhD @NorthwesternU Statistics & Data Science; Opinions are my own

Palo Alto, CA
Joined January 2014
Photo by Tim Lau (@timlautk)Photo by Tim Lau (@timlautk)Photo by Tim Lau (@timlautk)Photo by Tim Lau (@timlautk)Photo by Tim Lau (@timlautk)Photo by Tim Lau (@timlautk)Photo by Tim Lau (@timlautk)Photo by Tim Lau (@timlautk)
Delighted to witness Prof. Weijie Su @weijie444 receive the 2026 COPSS Presidents’ Award at JSM 2026 in Boston. Truly well deserved. Congratulations again! 🎉
2
11
2
74
16,128
Excited to see my NeurIPS 2026 paper with Weijie Su (@weijie444) cited in The Ultimate MoE Training Handbook’s discussion of router optimizers (ref. 196)! Beyond Muon vs. AdamW, we asked which changes to the router weight matrix W actually affect routing. For a router of the form softmax(Wx), adding the same vector to every expert row of W shifts all logits for a token by the same amount. Softmax cancels that shift. With sigmoid or sqrt(softplus), the same shift changes the gate scores. That’s why we center our softmax router updates across expert rows. For RowNormM, the order matters. We center the momentum, normalize each row, then center again because row normalization can reintroduce a shared component. In our downsized gpt-oss experiments, we kept Muon on the hidden matrices and varied only the router optimizer. Without auxiliary router losses, both centered and uncentered LeftPolarGradM finished with 0% dead experts, compared with 24% for AdamW. arxiv.org/abs/2605.18106
18 of the 20 best open models today are Mixture-of-Experts. Dense is basically over. And yet the know-how to train a good MoE still lives mostly in the heads of a handful of teams. So we spent 800k GPU hours running the experiments ourselves and wrote everything down. Super happy to finally release "The Ultimate MoE Training Handbook", free and open, organized around the questions training teams hit: - how many experts, how big, shared or not? - why is half my step spent in all-to-all, and how do I hide it? - why is my run 25x slower with no error in the logs? - softmax, sigmoid or sqrt(softplus) router, and why keep it in fp32? - aux loss or DeepSeek's loss-free bias? - upcycle a dense checkpoint or train from scratch? - can I serve with fewer experts per token than I trained with? ⚠️ Warning: it's long (~70k words, budget a few days, sorry) but with 55 interactive figures, code you can run and numbers from our own runs. Read it here: huggingface.co/spaces/huggin…
2
3
24
1,412
Tim Lau retweeted
I made this prediction 3 years ago, in July 2023. Few attending my talk believed me back then. In retrospect, replace was not quite the right word. AI won’t replace any fields. Instead, it will fundamentally reshape, and is already reshaping, fields by rewarding deep connections to real-world challenges and other disciplines.
16
27
8
290
219,864
There's an optimizer side to this too. Even exact zero-sum gradients can give Adam updates that aren't zero-sum. In our NeurIPS 2026 paper, we project LM-head and router updates to remove shared logit shifts. Curious to try both fixes together. arxiv.org/abs/2605.18106
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
6
8
109
7,980
1/9 🧵 🚨 New Paper accepted at #NeurIPS2026: PA Score Should rewording change an LLM’s uncertainty? We tackle this through conformal score design. 📄 Paper: arxiv.org/abs/2610.04239 💻 Code: github.com/Raina-Xin/PA_Scor… 📖 Blog: raina-xin.github.io/research…
3
17
1
35
1,631
Tim Lau retweeted
Math humbles many of us: there is always something deeper. Watching our internal model last month solved so many problems I’ve long stood in awe of has humbled me even further. It feels like the start of a Copernican shift in our view of intelligence… openai.com/index/sharing-ai-…
8
42
4
565
28,664
Tim Lau retweeted
Announcing our work textGrain, @OpenAI's text watermark in response to the EU AI Act. OpenAI is rolling out textGrain in the EU and opening detector access to researchers. We'll open-source it so the community can build on top.
We're expanding our approach to content provenance to include text in response to EU regulatory requirements, while recognizing the significant limitations of current text watermarking technology. Our tools already help verify whether an image or audio file was created with our models. This work builds on those efforts to help people better understand when content may have been generated or edited with an OpenAI model. In the EU, we’ll start watermarking eligible text from ChatGPT and Codex over the coming weeks to comply with the EU AI Act. Customers using our API can turn on text watermarking for select models worldwide today.
19
35
7
248
29,016
Excited to share that my joint work with @weijie444 on symmetry-compatible optimizer design has been accepted to #NeurIPS2026! Since our arXiv release, it’s been exciting to see this direction grow, both within and beyond language modeling. USC/Meta cites our work and incorporates RowNorm into Periodic Row-wise Muon for 1.3B–15B diffusion transformers. Meanwhile, Qwen3.8-Flash-Next and DeepSeek-V4.1-Flash highlight the importance of matching optimizer geometry to different parameter blocks. Looking forward to exploring how symmetry and geometry can guide architecture–optimizer co-design for more efficient large-scale training!
1/4 New paper with @weijie444! We introduce a symmetry-compatible principle for LLM optimizer design and, as a byproduct, get an end-to-end layerwise optimizer stack where every major matrix-valued parameter (embeddings, LM heads, SwiGLU MLPs, MoE routers) has its own principled update! 📝 arxiv.org/abs/2605.18106 💻 github.com/timlautk/equivari…
2
4
1
28
2,374
Tim Lau retweeted
With great power comes great responsibility
We’re working with an independent advisory group of mathematicians to help OpenAI responsibly share advances in AI and mathematics. The group will advise on how we assess and communicate new mathematical results, uphold academic and professional standards, and build tools that support mathematical research and learning. Through this work, we want mathematicians to be at the center of shaping how AI supports mathematical understanding and how its benefits reach the wider community. openai.com/index/advisory-gr…
11
12
227
18,443
Thanks @sama for spending the evening with us at GPT-6 Community Night! Excited to share how GPT has helped, shaped, and motivated my AI research.
2
42
1,862
Tim Lau retweeted
A few clarifications I’d like to make. The twin prime gap result was obtained just as GPT-6 Astra was about to be released, so we thought it was meaningful to share it as an illustration of Astra’s mathematical capabilities. All these efforts were happening concurrently; nothing was “rushed out.” BTW, Anthropic’s bound is 188, while ours is 186. My impression is that improving the bound from 188 to 186 is more difficult than improving it from 246 to 188.
Replying to @tonylfeng
Finally, I want to call out the general behavior from OpenAI and Anthropic. Navier–Stokes is just one more entry in a ledger of unhealthy competition between AI companies. Shortly after OpenAI unveiled the Unit Distance breakthrough, Anthropic(/Levent?) was there to match it. The day after Anthropic(/Levent?) disproved the Jacobian Conjecture, OpenAI matched them. On August 1, OpenAI rushed their release of 10 math breakthroughs because Anthropic(/Levent?) learned about the problems and started trying to scoop them. One month later, Axiom Math pounced to beat Julia Stadlmann’s recent work on prime gaps, then OpenAI rushed out a release beating Axiom Math’s bound, and within minutes Anthropic(/Levent?) was there to say they had (more or less) matched OpenAI’s bound. As my writing suggests, it is not clear to me (nor to OpenAI, apparently) to what extent Levent’s breakthroughs are meant to represent Anthropic. Either way, Anthropic has unleashed its technology on mathematics in unruly fashion, with answers dropped (still without explanations) in cryptic tweets or mysterious database contributions, Lean proofs dangling in random github repositories, and then a couple of official press releases (one very nicely written, though also arguably a scoop; and another where the discussion of the actual mathematics is a puzzling pile of slop). OpenAI has been more coherent, but has brought brutal force for flexing its power and trading blows with its rivals. AI companies are developing a powerful and dangerous technology, which will certainly upend the lives of many people. People are struggling to adapt their values, institutions, and sense of meaning. Amidst much anxiety about the future, frontier labs have advance access to the technology, so they lead by example. Can they please do it in a way that actually makes us optimistic about the future of humanity?
10
17
215
55,154
Finally, a Muon-dominated run using non-Adam updates for both token embeddings and the prediction head! 552B backbone parameters. 45T tokens. DeepSeek reports “no instability.” DeepSeek-V4 and Qwen3.8-Flash-Next kept both on AdamW. V4.1-Flash moves them to momentum + Sinkhorn balancing, reporting better performance with less optimizer-state memory. Different blocks, different optimizer geometry. Very much in the spirit of our symmetry-compatible principle! arxiv.org/abs/2605.18106
🚀 Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient. 🔹 Introducing the smallest model in our new architecture family, with native visual understanding. 🔹 Designed for greater capability, faster inference, higher throughput, and scaling to larger models. 1/6
1
2
65
5,733
Tim Lau retweeted
ChatGPT 6 disproved the global convergence of BDRS (Bregman Douglas-Rachford Splitting). This settles a problem that I have been working on since 2020. A preliminary write up is here: dropbox.com/scl/fi/y3ap89p6t… I will add more details and thoughts later.
5
7
82
5,626
Tim Lau retweeted
Announcing that GPT-6 Astra has pushed the prime gap to 186, with Lean formalization! I was 9 when I first heard the twin prime conjecture. Its elegance and Yitang Zhang’s legendary story have always stuck with me. A truly surreal night, being the first to see our model make progress, pushing 246 all the way down to 186, on a problem I’ve revered since I was a kid. For me, it felt like witnessing a new era of intelligence being born, made possible by everyone at @OpenAI!
56
242
43
2,076
152,969
各層の最適化は、その構造や対称性に合わせて設計すべきであり、この観点からMuonを含むSpectral Optimizerを統一的に整理できる研究を紹介する。 T. Lau, W. Su, “Symmetry-Compatible Principle for Optimizer Design: Embeddings, LM Heads, SwiGLU MLPs, and MoE Routers” まず通常の線形層を考える。 y = Wx ここでx, yは列ベクトル、Wは行列である。AdamでWを更新する場合、行列の各成分を独立したパラメータとして扱い、それぞれについてモーメンタムや勾配の二乗平均を計算し、更新幅を決定する。 この場合、更新はどの基底で行列を表したかに依存してしまう。 例えば、線形層の入力と出力の基底を直交変換(回転)して表したとしても、線形層そのものは同じ操作を表してほしい。 入力側を x -> Qx 出力側を y -> Py と回転させると、重み行列は W -> PWQ^T と書き換えられる。 実際、 (PWQ^T)(Qx) = PWx = Py である。 このとき勾配も同様に G -> PGQ^T と変換される。 このとき、最適化における更新写像をUとすると、 U(PGQ^T) = P U(G) Q^T を満たすのが自然である。 つまり、先に基底変換してから最適化を適用した結果と、最適化を適用してから基底変換した結果」が一致してほしい。これが同変性(equivariance)である。 Adamのように各座標に独立な非線形変換を適用するoptimizerは、一般にはこの性質を満たさない。そのため、同じ線形変換を表していても、基底を変えると最適化軌道が変わってしまう。 一方、Muonは勾配またはモーメンタムを行列として扱う。これらをまとめてDと書く。 Dを特異値分解して D = UΣV^T とすると、理想化したMuonの更新はpolar factor polar(D) = UV^T で表される。 すなわち、左特異ベクトルUと右特異ベクトルVは保ち、特異値だけを変換する。 このように、特異ベクトルを保存し、特異値だけを変換する写像をSpectral Operatorと呼ぶ。 一般にSpectral Operatorは U(D) = U Diag(Φ(σ(D))) V^T と書ける。ここでσ(D)はDの特異値を並べたベクトルであり、Φは特異値に対する変換である。 例えば勾配降下法では Φ(σ) = σ である。一方、理想的なpolar updateでは、非ゼロの特異値をすべて1に置き換えるため、 Φ(σ_i) = 1 とみなせる。 このpolar factorは polar(PDQ^T) = P polar(D) Q^T を満たす。このためMuonは、入力・出力の直交基底変換に対する同変性を保つ。 論文ではさらに、左右からの直交基底変換に対して同変かつ連続な更新写像がSpectral Operatorになることも示している。 この線形層以外の層は異なる形の対称性を持つ。 例えば、Embeddingを、行方向に各トークンの埋め込みベクトルが並ぶ行列として考える。 列方向のhidden featureについては基底を回転させてもよい。一方、行方向では任意の基底回転によって複数のトークンを混ぜることには意味がない。許されるのは、トークンの番号を入れ替える置換のみである。 したがって、Embeddingは、行方向には置換、列方向には直交変換に対する対称性を持つ。 これは通常の線形層が持つ左右両側の直交対称性よりも弱いため、例えば各行を独立に正規化するRowNormも対称性を壊さず使うことができる。 LM headもEmbeddingと似ており、語彙方向の置換とhidden featureの直交変換に対する対称性を持つ。 さらにLM headでは直後にSoftmaxを使うため、すべての語彙logitに同じ値を加えても出力確率が変化しない、という不変性がある。 そのため、語彙方向の平均成分は出力分布に影響しない冗長な方向であり、この平均成分を取り除いてから更新できる。 SwiGLUでは、中間層のニューロンを入れ替える置換に対して対称性がある。そのため、gate/up projectionでは中間ニューロンに対応する行方向の対称性、down projectionでは列方向の対称性に基づく更新が自然になる。 MoE Routerについても、expertの順番を入れ替える置換不変性と、Softmaxに由来する不変性が存在する。そのため、expert方向の平均成分を除去した上で、RowNormやLeft-Spectral updateを適用することができる。 このように、各層が持つ対称性をまず調べ、その対称性を保つように最適化法を設計する、というのが本論文の基本的な提案である。 コメント === 各パラメータの更新について、対称性という観点から最適化法を導出することを提案した興味深い研究である。 Muonなどによって、パラメータを要素ごとではなく行列として扱う方向性が注目されるようになったが、この論文ではさらに一歩進み、その行列がネットワーク内でどのような意味を持ち、どのような対称性を備えているかまで考慮して最適化法を設計することを提案している。 これまで、Muonについて実験的に示されている重要な利点の一つとして、大きなバッチサイズでも性能劣化が比較的小さく、大規模学習での学習効率を改善できる可能性があることが挙げられる。 また、低い特異値に対応する方向を相対的に強く更新することから、Long-tailな知識や低頻度方向の学習に有利なのではないかという仮説も考えられている。 こうしたMuonの性質を、単に特異値を均すからという観点だけでなく、モデルが持つ対称性と最適化の同変性から見る方向は興味深い。 こうした考え方は、パラメータの座標表示ではなく内在的な幾何に基づいて最適化するべきという、自然勾配法の流れともつながっているが、今回は各パラメーターブロックが持つ具体的な対称性から最適化を設計する方向を進めたものとみることができる。
3
10
1
95
10,814
1/3 New work from USC and Meta evaluates RowNorm on 1.3B–15B text-to-image DiTs with a diffusion velocity-prediction objective, extending evidence beyond autoregressive language modeling and to much larger scales. arxiv.org/abs/2608.20818v3
1
4
29
2,707
2/3 Conceptually, Periodic Row-wise Muon mixes bi-orthogonally equivariant NS5 and one-sided equivariant RowNorm: two classes in our symmetry-compatible taxonomy. Core takeaway: layer symmetry can guide both update geometry and cheaper, shard-local computation.
1
1
1
293
3/3 With K=3, two of three steps run RowNorm on sharded momentum, avoiding full-matrix all-gather. Across 256 H100s: 66.7% less optimizer communication, 47–54% less optimizer time, and 16–24% faster steps—evidence of our framework’s computational promise: teapawt.org/timlautk/status/205678…
1/4 New paper with @weijie444! We introduce a symmetry-compatible principle for LLM optimizer design and, as a byproduct, get an end-to-end layerwise optimizer stack where every major matrix-valued parameter (embeddings, LM heads, SwiGLU MLPs, MoE routers) has its own principled update! 📝 arxiv.org/abs/2605.18106 💻 github.com/timlautk/equivari…
1
1
249
Really cool to see Qwen3.8-Flash-Next reach a similar conclusion at huge scale: different layers need different optimizer geometry. Muon works well for ordinary linear maps but destabilizes the MoE router—exactly why we need symmetry-compatible updates: arxiv.org/abs/2605.18106.
⚡Meet Qwen3.8-Flash, a multimodal MoE and an early preview of the Qwen4 architecture, now open-weight! The production version Qwen3.8-Flash will be available soon via QwenCloud API at just $ 0.16/1M input tokens and $ 0.47/1M output tokens. 125B parameters + 51B N-gram embeddings, with just 6B activated per token. Unmatched cost-efficiency. What's new: 🥳 - Next architecture: GDN + QSA hybrid attention, Gated Residual, N-gram Embedding & Muon optimizer, serving as a precursor to the architecture used in Qwen4. - Dramatically lower training and inference costs: trained at just 1/9 the cost of Qwen3.7-Plus, while outperforming it across the board with especially strong gains in coding and office tasks. - Strong performance: scoring 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro, 73.9 on CoWorkBench, 84.5 on AndroidWorld, and 95.7 on MathVision (with CI). - 262K native context, extensible to 1M with YaRN. We’re also releasing the weights for Qwen3.8-Flash-Next, giving the community an early look at the new architecture we’re exploring for Qwen4.🚀 We can't wait to see what you build with Qwen3.8-Flash!👀👇 - Blog: qwen.ai/blog?id=qwen3.8-flas… - Technical Report: github.com/QwenLM/Qwen3.8-Fl… - Hugging Face: huggingface.co/Qwen/Qwen3.8-… - ModelScope: modelscope.cn/models/Qwen/Qw…
3
3
43
5,061
Also love that they split fused QKV and GDN tensors by logical head before orthogonalization. The optimizer should see the actual operators, not whatever tensor layout happens to be convenient.
1
3
247