There is a wrong notion that precision is sacrificed in UE8M0. That is not the case. You can retain the original accuracy even when you directly train in that format when you use our Madam algorithm https://
x.com/AnimaAnandkuma
r/status/1958573570688524596
…
@animaanandkumar
-
Madam Algorithm Preserves Precision in UE8M0 Training Format
By
–
-
DeepSeek v3.1 Uses UE8M0 FP8 Logarithmic Training Format
By
–
It is interesting that the new @deepseek_ai v3.1 is trained using the UE8M0 FP8 scale data format which is logarithmic number system. Our multiplicative weights update (Madam) for training in that format was done several years ago while at @nvidia It yields maximum hardware
-
AI Integration in Scientific Workflows: Augment or Replace?
By
–
How do we build AI for science? Augment with AI or replace with AI? https://
arxiv.org/pdf/2408.05177 Popular prescription is to augment AI into existing workflows rather than replace them, e.g., keep the approximate numerical solver for simulations, and use AI only to correct its errors -
LeanDojo Major Update: Lean 4 Code IDE for Verified Math
By
–
Major update of LeanDojo: Lean + LLM for verified math reasoning. https://t.co/yYKUrhx7UG
— Prof. Anima Anandkumar (@AnimaAnandkumar) 11 août 2025
Lean4Code v1.0.0 – A specialized integrated development environment built as a fork of VS Code, designed specifically for Lean theorem proving. The IDE features automatic Lean installation,… pic.twitter.com/PWJEVWMhafMajor update of LeanDojo: Lean + LLM for verified math reasoning.
https://
leandojo.org Lean4Code v1.0.0 – A specialized integrated development environment built as a fork of VS Code, designed specifically for Lean theorem proving. The IDE features automatic Lean installation, -
Fourier Neural Operators Enable High-Resolution Ocean Weather Models
By
–
Excited to share our recently published paper in @WileyGlobal on "Ocean Emulation With Fourier Neural Operators: Double Gyre" https://
agupubs.onlinelibrary.wiley.com/doi/10.1029/20
23MS004137
… We used Fourier Neural Operators to build the first high-resolution weather model, FourCastNet. Since it works so well for -
Optimizing Large Language Models: Memory and Bandwidth Solutions
By
–
My @MLSysConf keynote is now online. https://
mlsys.org/virtual/2025/i
nvited-talk/2887
… The scaling of large language models has led to impressive gains in language understanding, but at a cost of insatiable memory and bandwidth requirements. I advocated a principled approach of designing optimization -
AI Technology Overcomes Aircraft Turbulence Challenge
By
–
Thank you @BBCNews for featuring our work using AI to overcome turbulence
-
FourCastNet 3 Achieves Competitive Skill at 6-Hour Resolution
By
–
.
@shoyer We admire neuralGCM and all the contributions you are making for AI+climate modeling. Social media doesn't allow for too much nuance – what I meant to say was FourCastNet 3 is unprecedented in offering competitive skill at 6-hour resolution, with probabilistic estimates -
FourCastNet: First High-Resolution AI Weather Model
By
–
I led the creation of the very first high-resolution AI-based weather model FourCastNet at @NVIDIA @Caltech in 2021. Instead of bottoms-up physics-based weather forecasting, for the first time, we were able to show that AI-based models are accurate and tens of thousands of times… pic.twitter.com/1Rp3RAfeln
— Prof. Anima Anandkumar (@AnimaAnandkumar) 21 juillet 2025I led the creation of the very first high-resolution AI-based weather model FourCastNet at @NVIDIA @Caltech in 2021. Instead of bottoms-up physics-based weather forecasting, for the first time, we were able to show that AI-based models are accurate and tens of thousands of times
-
Former Student Recognized for Pioneering AI Chemistry Work
By
–
Hearty congratulations to my former student @ZhuoranQ for a very well deserved recognition. @ZhuoranQ was my first cross disciplinary student from chemistry and someone who was willing to explore ai+chemistry back when it was considered risky. His work on Orbnet was the first