Your curated collection of saved posts and media
At PyTorch Conference Europe, Ramine Roane (@roaner, @AMD) discussed how PyTorch evolved from making single-GPU programming easy to becoming a production tool for distributed training and inference, and the role of the open ecosystem in supporting the next stage of AI development. In this clip, he highlights the scale of today's open models, citing examples ranging from hundreds of billions to more than a trillion parameters. ๐ Watch the full keynote: https://t.co/brFzozSFpr #PyTorchCon
OpenMDW-1.1 is now available โ and @NVIDIAAI is adopting it across Cosmos, Isaac GR00T, Ising, and Nemotron model families. A permissive, unified legal framework purpose-built for AI models. Learn more at https://t.co/IT2NSDVLo3 https://t.co/xBdpZIZUXb
Two weeks left to nominate yourself or someone else as a PyTorch Foundation Ambassador๐ฅ The PyTorch Foundation Ambassador Program supports community leaders who organize events, create technical content, mentor contributors, and support regional PyTorch communities worldwide. As we continue to expand representation across local PyTorch communities, we especially welcome applications from contributors in Africa, Latin America, the Middle East, Oceania, Southeast Asia, and Eastern Europe. Ambassadors will receive recognition through PyTorch Foundation channels, exclusive resources, support for community engagement, and professional growth. Learn more and apply here before June 18, 2026: https://t.co/bwh8g8sf05:
Almost out of time! ๐จ The #CallForProposals for #PyTorchCon North America (Oct 20-21 | San Jose) closes tomorrow, June 7 at 11:59 PM PDT. Submit your session, BoF, or lightning talk now! CFP: https://t.co/hLlKK7WxLD https://t.co/GcMZ463miS
FINAL CALL! ๐ข The #PyTorchCon North America #CallForProposals closes TONIGHT at 11:59 PM PDT. Share your #AI expertise in San Jose this October 20-21! Apply now: https://t.co/hLlKK7WxLD https://t.co/QrGMFjM9Rf
On Monday, June 15, PyTorch Foundation project Helion is hosting a Helion DSL Tutorial at PLDI 2026 (47th ACM SIGPLAN Conference on Programming Language Design and Implementation) in Boulder, CO. Helion is a Python-based kernel DSL that supports multiple backends across NVIDIA, AMD, Google TPU, and Intel. This workshop is a fantastic opportunity for compiler researchers, kernel authors, and ML systems engineers to dive deep into the technology. This will be an entirely interactive session where attendees will get to write, autotune, and run real Helion kernels live. @JongsokC
Build on-device AI with ExecuTorch on @Snapdragon . The ExecuTorch Hackathon brings developers to San Francisco on June 27โ28 to build and optimize real-time AI applications that run directly on Snapdragon-powered mobile devices. Participants will build on Samsung Galaxy S25 Ultra devices powered by Snapdragon and receive mentorship and hands-on support from @Qualcomm and @Meta experts. Learn more and register: https://t.co/spjyl4kZJj Applications close June 15.
OpenEnv is a tool for creating an agentic execution environment, like terminals, browsers, or anything an agent can interact with. OpenEnv contributors announced that the project will now be coordinated by a committee that includes Meta-PyTorch, Reflection, @UnslothAI, @modal, @PrimeIntellect, @nvidia , @mercor_ai, @fleet_ai, and @huggingface. Today's announcement also notes that OpenEnv is supported and adopted by organizations including PyTorch Foundation, @vllm_project, @LightningAI, and others. ๐ Read more: https://t.co/SaaSDwFKVk
๐ง From breakthrough research to production-scale #AI, #PyTorchCon is where the #PyTorch community comes together. Join us October 20-21 in San Jose, California. ๐๏ธ Register by 31 July & save $400: https://t.co/AVHdaIFT20 https://t.co/FwgerDsXER
AI agents are changing how you interact with your PC. @nvidia and @Microsoft are teaming up to enable the next generation of developers to build on-device agents on the Windows platform, with easier setup, native security, and integration with the apps and tools developers already use. This includes optimizing the ComfyUI implementation in PyTorch, enabling developers to run larger models with better performance. Read the complete blog post: https://t.co/pAJRnDqKBh
perf packed release: safetensors 0.8.0 is out โก๏ธ Main takeaways: - direct copy into metal MTLBuffers + dlpack for 0-copy hand-off to target framework (only torch for now) -> 2-3x perf improvement + fixes OOMs loading models that are around the limit of unified memory on macOS when loading with transformers - GIL-free serialization, enabling multi-threaded saves from Python -> 1.2x to 2x faster for single files, but you can expect more improvement when saving multiple files in parallel! Check the release notes for the full list of improvements!
From #CloudNative to #AI and #OpenInfra: Experience the full spectrum of #OpenSource innovation in Shanghai. ๐จ๐ณ Event: KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China Dates: Sept 8-9 ๐ฐ Save ยฅ710 RMB if you register by July 28! https://t.co/cChMvGvcei https://t.co/yqRITldxQD
Helion is a PyTorch-native hardware agnostic kernel DSL designed for writing high-performance kernels using a tile-programming model. In our latest blog post, Sean Chen (@RedHat) and Yanan Cao (PyTorch, @Meta) explored integrating Helion kernels with @vllm_project and evaluated their performance across a range of LLM inference workloads, showing impressive performance speedups thanks to Helionโs Ahead-of-Time autotuning and fine-grained runtime dispatching. Letโs dive into kernel fusion, framework design, and end-to-end benchmarking, showing how Helion can be easily integrated into vLLM and bring significant benefits. Read the full technical deep dive ๐ https://t.co/15zXb5Y3Kg
Submit your poster proposal for #PyTorchCon North America! Share your latest #AI, ML, #PyTorch, tooling, infrastructure, or research work with the community October 20-21 in San Jose, CA. ๐๏ธ Poster #CFP closes July 26: https://t.co/M7WmgnmuKJ https://t.co/Ozs2HtTFGs
Enable smarter, longer-thinking agents Scale agentic AI and reinforcement learning by shortening CPU execution time, increasing task throughput, and improving overall AI factory output. The @nvidia custom Olympus core in the NVIDIA Vera CPU uses a neural branch predictor to reduce stalls in branch-heavy code. Combined with other prediction mechanisms, it can sustain two taken branches per cycle with zero penalty, maintaining throughput for deep software stacks such as PyTorch, graph workloads, and scripting engines. Read the complete blog post: https://t.co/L6NmBJWgDY
๐งฑ DiffusionBlocks: Training Neural Networks One Block at a Time https://t.co/zw7kehHDYc
ไปๅค22:00ๆพ้ ใใฌใๆฑไบฌWBS (@wbs_tvtokyo) ็ต็ฃ็ใฎAI้็บๆฏๆดใใญใธใงใฏใใGENIACใๆกๆใซใคใใฆใๅๆใๅใใพใใใๅผ็คพCEOใฎDavid Ha (@hardmaru) ใจResearch Scientistใฎ่ ๆฒผใใ็งใใกใฎๆฆ็ฅใๆฅๆฌ็บใฎAIใไธ็ใๅคใใๅฏ่ฝๆงใซใคใใฆ่ชใใพใใใใฒใ่ฆงใใ ใใ๏ผ
Building AI that Builds AI: Introducing the Sakana AI RSI Lab ๐ https://t.co/AskX3J5oEJ Today, we are announcing the Sakana AI Recursive Self-Improvement (RSI) Lab: a dedicated research group in Tokyo tasked with redesigning the AI development process itself using AI. While the industry increasingly speculates about the theoretical potential of self-improving AI, weโve spent the last two years actively laying the foundations to make it a reality: โช LLMยฒ: AI models automating research to invent better preference optimization algorithms. โช Darwin Gรถdel Machine: Agents autonomously rewriting their own codebase to double software-engineering performance. โช ShinkaEvolve: Hyper-sample-efficient program evolution that builds novel loss functions for MoE models. โช ALE-Agent: Reinforcement agents outperforming hundreds of human experts via self-learning. โช Digital Red Queen: Open-ended adversarial coevolution laying the groundwork for RSI in cybersecurity. โช The AI Scientist: Towards end-to-end automation of AI research, recently published in Nature. Now, we are unifying these breakthroughs. The Sakana AI RSI Lab is officially tasked with building open-ended, adaptive architectures that collectively self-improve. Human intelligence did not emerge from limitless resources; it was forged through the open-ended, compounding process of evolution operating under strict constraints. We are applying this exact principle to AI. We believe recursive self-improvement is achievable on modest, sample-efficient compute. It shouldnโt be a winner-take-all asset locked inside hyperscale clusters, but a democratized public good. Weโre scaling our team to execute this mission. We are looking for frontier scientists and engineers who are entirely unsatisfied with the brute-force status quo. If you are ready to break away from standard benchmarking and build the self-improving future in Japan, come build with us.
Weโve been laying the foundations for RSI over the last 2 years. Now, I am looking for a select group of distinguished frontier researchers and engineers to join our core RSI team. https://t.co/1eEBgLdQhY If you have a proven track record at the frontier, but find yourself entirely bored with the status quo of brute-force scaling, this is your call.
Member of Technical Staff (RSI Lab) https://t.co/ptU2Xh4aO1 If you are a visionary builder ready to move to Tokyo and engineer the engine of recursive discovery, we invite you to apply. https://t.co/ydKhZQiup8
Not only do we want to train a good model, we want to know it'll be good before we even start training. About a month ago, the Marin team launched a 129B (16B active) 1e23 FLOPs MoE run and preregistered a loss of 2.252. The run finished this past week and landed at 2.234. https://t.co/OptaVa7jIO
This week, @classiclarryd kicked off a 129B (16B active) 1e23 FLOPs MoE run. In typical Marin style, we have fit scaling laws and have made a loss projection of 2.252. Stay tuned. https://t.co/QnwJ8YxT9H
A mic drop moment @ycombinator tonight @sama just offered $2M in OpenAI tokens to EVERY YC startup in the current batch in exchange for equity Just like Yuri Milner offering to invest in every startup back when Sam was a YC partner I can't wait to see what's unlocked when you let the most driven, creative and formidable founders tokenmaxx
Itโs Codex Thursday, and yes, we have updates for you. First up: Appshots, a new way to bring the context of what youโre working on into Codex. On your Mac, press Command-Command to attach your app window to a Codex thread. Codex gets both a screenshot and text from the window, including content beyond whatโs visible onscreen. Appshots are available across plans on Mac, with enterprise access coming soon.
AI should dramatically increase quality of life and individual freedoms for people around the world. The OpenAI Foundation is making an initial $250M commitment to measurement, transition support, and new approaches to broadly shared prosperity. https://t.co/zOD8O94RjQ
It's time to fly. https://t.co/ObUaCZ07EM
The AI race is no longer just about models. It is about energy. Chinaโs wind-powered underwater datacentre shows how the next wave of AI infrastructure may be built around new approaches to power, cooling and efficiency. The winners in AI may not only be the companies with the smartest algorithms, but also those with the most sustainable infrastructure.
https://t.co/IVGHTjLNTf
AI safety is becoming a launch strategy. Anthropicโs release of a Mythos-class model shows how leading AI companies are trying to balance powerful new capabilities with safeguards for high-risk areas. The real test is no longer whether models can do more. It is whether society trusts how they are released.
AI may help find the next football superstar, but it should not define what talent is. Data can make scouting faster and more disciplined, but football still depends on qualities that are hard to measure: instinct, resilience, creativity and the ability to surprise. The risk is not using AI in sport. The risk is believing the spreadsheet sees everything.
https://t.co/QGVSJcNhOB
The AI race is becoming a price war. If OpenAI cuts token prices to compete with Anthropic, lower costs could accelerate enterprise adoption but also pressure margins across the AI stack. The next battle may not be who has the most powerful model. It may be who can deliver intelligence at the lowest sustainable cost.
The future of work will not be human versus AI. The people who thrive will be those who combine human judgment, creativity and trust with AIโs speed, scale and pattern recognition. The real career risk is not being replaced by AI. It is being outpaced by people who know how to use it better.