Humphrey Shi
@humphrey_shi
professor @GeorgiaTech | UIUC ‖ engineer–researcher building next-generation high-performance, multimodal, and creative AI systems
Over 4 years into our journey bridging Convolutions and Transformers, we introduce Generalized Neighborhood Attention—Multi-dimensional Sparse Attention at the Speed of Light: github.com/SHI-Labs/NATTEN A collaboration with the best minds in AI and HPC. 🐝🟩🟧 @gtcomputing @nvidia
10 years ago, I recruited 4 new PhD students—including Jiahui—with the late Prof Tom Huang, thanks to new industry funding. Proud to see them shaping AI’s frontier. Today, academia struggles to fund the next generation. Industry—your support & partnership matter more than ever!
Real Madrid spent $80M to sign Ronaldo from Manchester United in 2009. Meta paid $100M to sign Jiahui Yu from OpenAI in 2025.
Congrats to Prof Barto & Sutton on Turing Award! 🎉 Barto’s journey is inspiring—I happen to be teaching McCulloch-Pitts neuron to hundreds of @GeorgiaTech undergrads today in my comp vision class.Makes me wonder what breakthroughs our next-gen leaders will achieve in 50 years🚀
Meet the recipients of the 2024 ACM A.M. Turing Award, Andrew G. Barto and Richard S. Sutton! They are recognized for developing the conceptual and algorithmic foundations of reinforcement learning. Please join us in congratulating the two recipients! bit.ly/4hpdsbD
Meet the recipients of the 2024 ACM A.M. Turing Award, Andrew G. Barto and Richard S. Sutton! They are recognized for developing the conceptual and algorithmic foundations of reinforcement learning. Please join us in congratulating the two recipients! bit.ly/4hpdsbD
Five months after Hurricane Helene, we’re finally starting to rebuild our home. Grateful for my family’s resilience through this journey 🙏 Meanwhile, AI’s rapid progress feels like a hurricane, reshaping everything in its path. Perhaps we, too, should rebuild—forward & onward.
🌩️ Where AI Compute & Talent Are Going — and how industry and academia can thrive together？ 1️⃣ Industry ⚡ – the Gigawatt Rush Silicon is the new oil, and the gushers are roaring: xAI (@elonmusk) → 50 million H100-class GPUs in < 5 years — > 100 × current fastest AI…
The @xAI goal is 50 million in units of H100 equivalent-AI compute (but much better power-efficiency) online within 5 years
Another PhD alum @bowenc0221 from our IFP lab just joined @Meta SuperIntelligence (MSL) team this past week—makes me super proud. I expect more talented alumni to make strides there or elsewhere too: ifp-uiuc.github.io Academia’s AI talent pipeline is still meaningful in my…
🚨 BREAKING: Detailed list of all 44 people in Meta's Superintelligence team. — 50% from China — 75% have PhDs, 70% Researchers — 40% from OpenAI, 20% DeepMind, 15% Scale — 20% L8+ level — 75% 1st gen immigrants Each of these people are likely getting paid $10-$100M/yr.
Sparse Attention is now pushing World Foundation Models to the Speed of Light! Attention powers modern AI (Transformers, ViTs, DiTs), and Sparse Attention is the next frontier. Neighborhood Attention (NA) is the first multidimensional sparse attention infrastructure that: -…
Cosmos-Predict2 meets NATTEN. We just released variants of Cosmos-Predict2 where we replace most self attentions with neighborhood attention, bringing up to 2.6X end-to-end speedup, with minimal effect on quality! github.com/nvidia-cosmos/… (1/5)
Congratulations to Dr. Kristen Grauman @UTAustin on receiving the 5th Thomas S. Huang Memorial Prize at @CVPR 2025 — honoring her outstanding contributions to research, education, and service to our community! Remembering Tom & Margaret: youtu.be/QV7WnO9Lk9M?si…
We are releasing a major NATTEN upgrade that brings you new Hopper & Blackwell sparse attention kernels, both capable of realizing Theoretical Max Speedup: 90% sparsity -> 10X speedup. Thanks to the great efforts by @AliHassaniJr & @NVIDIA cutlass team! natten.org
Over 4 years into our journey bridging Convolutions and Transformers, we introduce Generalized Neighborhood Attention—Multi-dimensional Sparse Attention at the Speed of Light: github.com/SHI-Labs/NATTEN A collaboration with the best minds in AI and HPC. 🐝🟩🟧 @gtcomputing @nvidia
A paper from my PhD students—nearly a year of work—was rejected by @icmlconf despite 4 weak accepts, citing “calibration with other submissions.” Still incredibly proud of my students. To young researchers: rejections happen. Keep learning, keep going—the real judge is within.
Wondering what's happening with NATTEN in 2025? Check out Generalized Neighborhood Attention! Spoiler: NATTEN gets a new stride parameter, we made a simulator for all your analytical studies, AND a Blackwell kernel! Keep reading for more... (1 / 5)
Impressed by FramePack from @lvminzhang & @magrawala! Their table puts our StreamingT2V (Mar 2024) at #2 overall and 🥇 in motion (99.96 %). A nice reminder that memory blocks still matter—and may fruitfully complement token‑compression and other approaches for marathon vids!🏃
㊗️Congrats on Lvmin Zhang’s (github@lllyasviel) latest project FramePack and thank you for using and recommending HunyuanVideo. 😀So happy to see innovations based on Hunyuan and we would like to see more. ▶️FramePack's Brief Intro and Showcases Attached: FramePack is a…
"Thinking with Images" is what we have been cooking after GPT-4o launched last year and it marks a paradigm shift in how we view/solve perception problems in this new era of RL. It is such a pleasant and an honor to work with this amazing team to get it out!
“Thinking with Images” has been one of our core bets in Perception since the earliest o-series launch. We quietly shipped o1 vision as a glimpse—and now o3 and o4-mini bring it to life with real polish. Huge shoutout to our amazing team members, especially: - @mckbrando, for…
After nearly 5 incredible years, I’ve stepped down from my role as Chief Scientist at Picsart. Grateful for the journey—from building AI Research from scratch to a global team creating products used by millions every day✨ Now exploring what’s next in multimodal AI🚀. DMs open🤝
Huge congrats to @jhyuxm and @bowenc0221—our young and talented former IFP lab alumni—for their amazing work on GPT-4.1 at OpenAI.🎉 Proud and inspired — as one of the old guards watching the next wave take flight. The future clearly belongs to the young. Onward and upward! 🚀
GPT-4.1 (and -mini and -nano) are now available in the API! these models are great at coding, instruction following, and long context (1 million tokens). benchmarks are strong, but we focused on real-world utility, and developers seem very happy. GPT-4.1 family is API-only.
Check out Slow-Fast Video MLLM — a new paradigm to empower multi-modal LLMs with longer video context and finer spatial detail! 🎥🧠 🔗 github.com/SHI-Labs/Slow-… Led by 🐝 Min @__flying_lynx__ from @gtcomputing, in collaboration with @nvidia @ZhidingYu and more 🤝
One year ago, @abhshkdz and I left Meta to start Yutori. Ten months ago, @DhruvBatraDB joined us :) Nine months ago, we crystallized our vision. Two months ago, we released a sneak peak into what we’ve been building. Today, can’t be more excited to fully unveil @yutori_ai’s…
Video/Physics Generative AI was bottlenecked by diffusion runtime— 5s used to take minutes. My student @AliHassaniJr @gtcomputing helped scale full 35-step Cosmos 7B DiT 40× to real-time on Blackwell NVL72, in collab w/ @nvidia @liu_mingyu’s team. Congrats—just the beginning!🐝🚀
Nvidia just released Cosmos-Transfer1 on Hugging Face Conditional World Generation with Adaptive Multimodal Control
Check out WorldModelBench, our first workshop on Benchmarking World Models @CVPR, lead by researchers from @nvidia and beyond. Explore benchmarks, evaluation metrics, downstream tasks, & safety for World Models. Call for papers now open till April 7th: worldmodelbench.github.io
Join us at the WorldModelBench workshop at #CVPR2025 where we'll tackle systematic evaluation of World Models! Focus: benchmarks, metrics, downstream tasks, and safety. Submit papers now: worldmodelbench.github.io