Mark Kim retweeted
Agent loops need fast feedback, and LLM quantization doesn't have it: evaluations of reasoning models are slow, and PTQ methods are hard to port and combine.
Meet QDA. A coding agent designs the quantization recipe for each model and precision, with a KL score that takes minutes and a library of 20+ stackable PTQ methods.
📈 Beats the best public recipe on 11 of 12 model-precision pairs: +0.9 / +1.4 / +2.2 pts at W4A16 / W4A8 / W3A16
🔥 GLM-5.3-Flash (320B) at NVFP4: 98.3% of FP8 accuracy, vs 96.9% for NVIDIA's release
💰 Only $19–75 of Claude Code tokens per search
⚡ Runs on stock vLLM, SGLang and llama.cpp, zero inference overhead
Blog, paper, code and checkpoints below 🧵
Mark Kim retweeted
🚀 Humanfia × Geometrization: Perelman's Proof, Now Fully Formalized in Lean
The Humanfia team and Yuan Liao, Ziyang Qin, Ayush Khaitan , Ben Chow, and have completed a full Lean formalization of the Hamilton–Perelman proof of Thurston's Geometrization conjecture! It goes beyond Poincaré and classifies every 3-manifold. Mathematicians needed years to digest this proof. Now a machine has verified it line by line within weeks!
The progress of Humanize agentic math:
26-07: Perfect score in IMO 2026
26-08: 670/672 in PutnamBench
26-09: 100% in PutnamBench and 1st place in LeanEval (Surpassed by Axiom later)
26-10: 1st place again in LeanEval with upgraded Humanize workflow
26-10: Completed the Lean formalization of the Hamilton–Perelman proof of Thurston's Geometrization conjecture.
We are increasingly convinced: when everyone has tokens of similar intelligence, the agent workflow is what decides how far AI can go in frontier mathematics.
Tokens are「transforming」 the math.
🌌Geometry library: github.com/qinz1yang/differe…
⚙️ Humanize: github.com/humanfia/humanize
#AI #LLM #Humanize #Mathematics #Lean4 #Agent #Humanfia #Geometrization
Mark Kim retweeted
Two years ago I argued we needed an autoformalizer.
Now the time has come to formalize all of mathematics.
We need to unite behind a single Mathematics Autoformalization Project that turns all known mathematics into one Lean library.
@jdlichtman @patrickshafto
Mark Kim retweeted
🚀 Kernel Design Agents to Optimize Kimi Delta Attention (upto 2.96x)
When the KDA (Kernel Design Agents) project was initiated, its name happened to coincide with another KDA (Kimi Delta Attention). More than once, people asked us , "Can we use KDA to write a version of KDA?" Today, right at a moment with a full moon good for shooting, we are thrilled to share the latest results of KDA(gent)-v0.6: using KDA(gent) to optimize KDA(ttn)!
KDA-v0.6 achieves this through
* Humanize2 Flow & LLM Collaboration Optimization: Introducing multi-round iterative evolution with Flame Chase, GPT-5.6-Sol, and Fable-5, paired with an automated Workspace cleanup mechanism to significantly broaden the operator search space.
* Multilingual Domain-Specific Skill Stack Support: Fully supporting CuTe-DSL, CUDA C++, as well as recent Agent-Native CAKE IR and TIRx. Combined with customized diagnostic tools (such as CuTe's dedicated IKET pipeline analyzer and TIRx's CPU-side numerical simulation and static checks), achieving deep performance bottleneck diagnostics.
* Self-Evolving Kernel Wiki: Building a dynamically updated knowledge base that automatically cleans up invalid/erroneous implementations, corrects classification tags, and streamlines search indexing, significantly improving the Agent's knowledge retrieval quality and code generation accuracy.
KDAgent ultimately optimized the KDAttn kernel to 2.96x of FlashKDA, and we face a lot of unexpected hacks during the optimization. More details shared in our blogs nvlabs.github.io/kda/blog/20…
Release results of KDA for KDA: github.com/NVlabs/kda/tree/2…
Community & Production Deployment:
All kernels are open-sourced for community verification! For deploying NVIDIA Agentic CUDA acceleration in production environments, we recommend using the production-level version rigorously verified by FlashInfer.
By the Team: Dongyun Zou, Yixin Dong, Hongyi Jin, Junxian Guo, Yahui Cui, Avery Huang, Zihao Ye, Junru Shao, Changye Li, Zijian Zhang, Sihao Liu, Song Bian, Ligeng Zhu.
Mark Kim retweeted
The Future of Mathematics Symposium is being held on May 1st-2nd!
We have a truly exceptional group of speakers, with multiple Fields Medalists including Terence Tao.
Tune in for livestream 9am-5pm PST (link below)
Mark Kim retweeted
This is an ImageNet moment for autoformalization
We are pleased to share that using Gauss, we have completed a ~200K LOC formalization of Maryna Viazovska’s 2022 Fields Medal theorems on optimal sphere packing in dimensions 8 and 24.
This is the only Fields Medal-winning result from this century to be completely formalized, and is the largest single-purpose Lean formalization in history.
We are honored to have assisted @SidharthHarihar1 and the rest of the sphere packing team in this achievement.
math.inc/sphere-packing
Mark Kim retweeted
Feeling incredibly grateful to work with @SimularAI team and our partners on the future of autonomous computers.
When we started Simular, the goal wasn’t “agents as a demo”. It was agents that can reliably use a computer alongside humans. Hitting OSWorld milestones and launching Simular 1.0 are early steps on a very long road.
Thank you to our teammates, customers, and investors for betting on this vision so early. We’re just getting started!
simular.ai/articles/simular-…
Mark Kim retweeted
We've raised $300M in Series B funding from Salesforce Ventures and Anjney Midha (AMP)
FLUX is used by millions every month and powers production workflows across the world's leading platforms.
This funding will allow us to invest deeply in research and build the foundations for visual intelligence.
Mark Kim retweeted
We’re humans& ai
Come help us work on better communication and collaboration for LLMs to help smoothly and beneficially integrate AI into our human society.
If you’ve worked on synthetic memory, long context, RL with and without verifiable rewards, user modeling and simulation, fast inference, computation and communication overlap with accelerators, distributed systems, scaling, or anything else you think might help with our mission please get in touch!
Mark Kim retweeted
We're hiring for Research Scientists / Engineers!
- We closely work with all frontier labs
- We're a small org and can move fast
- We can choose our own agenda and what we publish
We're especially looking for people who enjoy fast empirical research.
Deadline: 31 Oct!
Mark Kim retweeted
We just hit 100 million lifetime downloads on Hugging Face! 🦥🤗
Huge thanks to all of you! The amazing community, model creators, and HF team. 💖
Mark Kim retweeted
New blogpost and demo! I optimized a Flappy Bird world model to run locally in my web browser (30 FPS)
(demo and blog in replies)
Mark Kim retweeted
We are launching a new Lean project to formalize a proof of the ABC conjecture almost always.
Based on our paper below, proving that the exceptions to the ABC conjecture (if any exist) must be very rare.
arxiv.org/abs/2410.12234
Mark Kim retweeted
Today we're releasing FLUX.1 Kontext - a suite of generative flow matching models that allow you to generate and edit images.
Unlike traditional text-to-image models, Kontext understands both text AND images as input, enabling true in-context generation and editing.
Mark Kim retweeted
whoa! more mucin magic, this time in the blood brain barrier
Excited to share our new study in @Nature! We uncover how the vascular glycocalyx—a carbohydrate-rich coating on blood vessels—plays a crucial role in brain health and aging. A @wysscoray & @CarolynBertozzi labs collaboration. 🍭🧠 #glycotime (1/12)
nature.com/articles/s41586-0…
Mark Kim retweeted
Excited to share our new study in @Nature! We uncover how the vascular glycocalyx—a carbohydrate-rich coating on blood vessels—plays a crucial role in brain health and aging. A @wysscoray & @CarolynBertozzi labs collaboration. 🍭🧠 #glycotime (1/12)
nature.com/articles/s41586-0…
Mark Kim retweeted
I've uploaded the latest slides & beamer source code to github.com/sustcsonglin/line…. Hopefully this repository will help train an LLM that generates Beamer slides better than I do :)
Linear Attention and Beyond: Interactive Tutorial with Songlin Yang (@SonglinYang4 MIT/Flash Linear Attention)
I didn’t follow some of the recent results, so I zoomed Songlin and she explained it all to me for two hours 😂
youtu.be/d0HJvGSWw8A
Mark Kim retweeted
We made 5 challenges and if you score 47 points we'll offer you $500K/year + equity to join us at 🦥@UnslothAI!
No experience or PhD needed.
$400K - $500K/yr: Founding Engineer (47 points)
$250K - $300K/yr: ML Engineer (32 points)
Challenges:
1. Convert nf4 / BnB 4bit to Triton
2. Make FSDP2 work with QLoRA
3. Remove graph breaks in torch.compile
4. Help solve Unsloth issues!
5. Memory Efficient Backprop
If you have any questions about the challenges, please feel free to ask! We're looking for people to help push Unsloth forward - so come join us to democratize AI further!
Our past work includes:
1. 1.58bit DeepSeek R1 GGUFs: xitter.gq/UnslothAI/status/18838…
2. GRPO with Llama 3.1 8B in a Colab: xitter.gq/UnslothAI/status/18875…
3. Gemma bug fixes: xitter.gq/danielhanchen/status/1…
4. Gradient accumulation bug fixes: xitter.gq/danielhanchen/status/1…
Details & submission guide: colab.research.google.com/dr…
Fixed a bug which caused all training losses to diverge for large gradient accumulation sizes.
1. First reported by @bnjmn_marie, GA is supposed to be mathematically equivalent to full batch training, but losses did not match.
2. We reproed the issue, and further investigation showed the L2 Norm betw bsz=16 and ga=16 was 10x larger.
3. The culprit was the cross entropy loss normalizer.
4. We ran training runs with denormalized CE Loss, and all training losses match.
5. We then re-normalized CE Loss with the correct denominator across all gradient accumulation steps, and verified all training loss curves match now.
6. We've already updated @UnslothAI with the fix, and wrote up more details in our blog post here: unsloth.ai/blog/gradient
This issue impacts all libraries which use GA, and simple averaging of GA does not work for varying sequence lengths.
This also impacts DDP and multi GPU training which accumulates gradients. Please update Unsloth via pip install --upgrade --no-cache-dir unsloth and use from unsloth import unsloth_train
We have a Colab notebook using our fixed GA: colab.research.google.com/dr… and a Kaggle notebook: kaggle.com/code/danielhanche…






























