We taught a brand-new mini-series this year at @SCSatCMU on Modern GPU Programming for ML Systems, as part of the ML Systems course, touching on fun questions like what data layout swizzling is, how to use 3D TMA, and state-of-the-art Blackwell programming. We released a curated online book based on the materials: https://t.co/5ZJg2lySNO check it out

How does it work? Sakana Fugu is itself an LLM, trained to call various LLMs in an agent pool, including instances of itself recursively. Fugu dynamically orchestrates the world's best models to tackle complex, multi-step tasks. As shown in this figure, Fugu is a multi-agent system that behaves like a single model. You send a request to one endpoint, and Fugu decides how to handle it internally. Fugu manages model selection, delegation, verification, and synthesis automatically. It solves tasks directly when that is enough, or coordinates a team of expert models when a problem calls for more. The complexity of a multi-agent system never reaches your code. At launch, Sakana Fugu comes in two models accessed via a single OpenAI-compatible API: • Fugu balances strong performance with low latency for everyday work. It fits naturally into tools like Codex for coding, as well as chatbots and interactive services. You can also opt specific agents out of its pool for data compliance. • Fugu Ultra is our flagship model tuned for maximum answer quality on hard, multi-step problems. It coordinates a deeper pool of expert agents for demanding work like AI research, cybersecurity analysis, and patent investigations.

you may have heard that glm-5.2 at 280 token/s is cool, how about 318 and we still have room to go https://t.co/4g0dI6CEzd

I took the new AA-Briefcase scores from @ArtificialAnlys (basically having the AI do multi-week consulting gigs with a lot of complexity) and graphed the frontier curve for open and closed models: 1) Surprise, rapid gains! 2) The open weights gap is clear https://t.co/a1QGQC2hey https://t.co/bqJHA0WU0j

Last week, we released HANDBOOK.md: a benchmark for long-context agentic instruction following. HANDBOOK drops an agent into a live company environment with files (PDFs, Excel, Word docs…), tools (email, Slack, Jira, calendar…), and a dense corporate handbook (up to 124 pages!). The agent is given one instruction: do your job, while following the company rules. Every frontier model broke them over 75% of the time. They fired employees without authorization... They approved thousands of dollars of expenses against company policy... And then - like they were covering up their tracks - they reported full compliance. HANDBOOK.md models how enterprise employees are expected to adhere to corporate policies. Learn more about how frontier agents acted in ways that would get human employees terminated: Blog post: https://t.co/zJ7zVpDOfH Github: https://t.co/zjwood6H6s Benchmark Leaderboard: https://t.co/lI3F0MwkCc

📣📣 Meet Qwen-AgentWorld — a native language world model that simulates 7 agent environments (MCP, Search, Terminal, SWE, Web, OS, Android) within a single model. Environment modeling is the training objective from day one, not a post-hoc adaptation. 🤔 LLMs are trained to be better agents — better at acting in environments. But nobody has trained them to model the environments themselves. 🗺️ Our roadmap: investigate how language world modeling can push the boundaries of general agent capabilities, along two routes: 1️⃣ Build a foundation model for environment simulation — outperforming Claude Opus 4.8 and GPT-5.4 on AgentWorldBench 2️⃣ Investigate how world modeling enhances agent training: 🔬 Controllable Sim RL (agentic RL with LWM as environments) surpasses training in real environments 🧠 Learning to predict environments (LWM warm-up) makes agents stronger — remarkably, even without any agent-specific training, this predictive knowledge transfers to agentic tasks with zero fine-tuning 📑 Paper: https://t.co/Jx2l5RKq71 📖 Blog: https://t.co/7tVcKyhsx2 💻 GitHub: https://t.co/B5Lvb1UZCn 🤗 HuggingFace: https://t.co/Kw3QBL1TM5 🧩 ModelScope: https://t.co/YBnGYgMWWI

Oops, SIGReg did it again! Large scale (CC12M->Datacomp-L) vision-language JEPA pretraining beats CLIP and SigLIP objectives! Thanks to SIGReg, our LeVLJEPA has no collapse, no EMA, no stop-gradient, no negatives, no problem! Checkpoints/demo are live: https://t.co/wz6S6tYB6p

AI Infra Day | SGLang × Sarvam with Hugging Face India's AI infra community is coming together. A day of deep-dives and technical discussions with researchers and engineers building the future of AI infrastructure. Bangalore | 11th July | 12:00–4:00PM Register Now: https://t.co/StSC0dxac9

GLM 5.2 DSpark preview is here! ✨ https://t.co/DQOMYEiY1o This is the first DSpark speculator for a non-DeepSeek frontier model, trained with Speculators and running on vLLM nightly for ~1.5× faster decode for GLM-5.2-FP8 on 4×B300. Stronger checkpoints to come!

the wildest part of this intelligence per watt paper (71.3% of chat queries could be local) is that the model is only a gpt-oss 20b. which is about a year old! compared to the current batch of small moe models (gemma 4, liquid LFM, Qwen-3.6, etc.) this is nothing. https://t.co/d4Oem5d35t

The Sensorimotor World Model (https://t.co/K5iWbk7Izs): a deep dive into the role of inverse dynamics modeling as an anti-collapse regularization for JEPAs. IDM is weaker than SIGReg as it doesn't have to fill the space--it only captures what is affected by the agent's actions🧵 https://t.co/kdnVGbhkht

PyTorch-native NeMo AutoModel handles transformer pretraining in @nvidia's end-to-end workflow for building a transaction foundation model. The workflow combines GPU-accelerated data processing and tokenization, decoder-only model pretraining, embedding extraction, and XGBoost fraud classification. On the synthetic @IBM TabFormer dataset, combining raw features with learned embeddings increased Average Precision by 41.76% over the raw-feature baseline. 🔗 Read the full post: https://t.co/DJvRP2K5Qp

Use Case 1: Autonomous ML Research Can an AI autonomously improve another AI’s training recipe? We tasked Fugu Ultra with improving a small GPT model using AutoResearch. Over 14 hours on a single H100 GPU, Fugu ran > 100 experiments. It iteratively edited the training code, ran tests, and kept any changes that successfully lowered the validation error rate. Watch the animation. The callouts track every time Fugu Ultra autonomously discovered a new improvement across batch size, model depth, learning rates, and optimizer settings. We pitted Fugu against three frontier models (Gemini 3.1 Pro, Opus 4.8, and GPT 5.5). To keep the focus purely on agentic behavior rather than brand wars, we anonymized them as Models A, B, and C. The Results: • Fugu Ultra (bold red) finished with the best mean performance (0.9774). • Fugu Ultra also achieved the best single run of the entire experiment (0.9748), leading every single baseline. For long horizon, agentic ML research, using Fugu to dynamically orchestrate a pool of strong models significantly outperforms relying on any individual monolithic model.

@Etched Congrats!! I was impressed to learn about some of the engineering wizardry (e.g. *very* low voltage domains, cluster scale memory, ...) that goes into tokens/watt maxxing of state of the art LLMs at interactive tokens/sec/user. Esp fun and memorable is the idea that this is engineering at the "opposite" regime to that of power transmission lines: very low voltage high current (at tiny distances) vs. very high voltage & low current (at great distances). Looking forward to more!

Can regularization based JEPA (e.g. SIGReg) scale and compete with SOTA foundation models (DINO)? Here is the answer: yes and with 10x less data. VISReg (slight variation of SIGReg) competes with DINOv2-LVD142M while only training on inet22k. Try it out: https://t.co/vBhrNAmFq6 https://t.co/XERFZEAE8t

We’re introducing GeneBench-Pro, a research-level benchmark for a harder kind of AI progress: how well agents can navigate messy biological data, choose the right analysis path, and make judgment calls that real computational research depends on. https://t.co/AsilnnSxnE

Announcing the first production robot navigation framework on $500 hardware Explore the world once → your robot agent will relocalize and build a persistant, spatial memory across sessions SLAM, relocalization, loop closure, map i/o, planning, control No ROS. Open source. https://t.co/VCk9GvOrrM

I'm going to try the new @NVIDIAAI Nemotron-3-Nano-30B-A3B and compare it to Qwen 3.6 35B in agentic workflows. https://t.co/z9cnRBOo1c

DeepSeek preparing release of DSpark, DFlash and Eagle draft models for Qwen3 and Gemma-4 variants https://t.co/2zdfL9XAkQ

I have been trying Sakana Fugu Ultra-high and, first, it is incredibly slow: my typical coding tests (shaders, interactive scenes) take 30 minutes to run And the results are... fine. It does not match Fable in real use. Its harbor is a good example: https://t.co/xVqulPBsQf

While we eagerly await Fable 5's return, our agentic WebGPU kernel optimization framework kept running. Opus 4.8 picked up where Fable left off, pushing Liquid AI's new LFM2.5 230M to an unbelievable 1,400 tok/s... running locally in your browser. Don't blink or you'll miss it. https://t.co/27WARZwTcD

Our team at Xaira was fortunate to have early access to test Claude Science (Operon). 🔥🚀 We used it to add agentic loops to both virtual cell modeling and protein design workflows. A nice plus: Operon had already added our scGPT as one of the default skills for single-cell analysis 🙏😎🔥 This is the kind of product that actually understands how research works, not just chat with a model, but traceable artifacts, reproducible environments, and real scientific data connections. That's a big deal for computational biology.

Today, we give robots a /skills library that self-evolves and compounds indefinitely! Introducing ASPIRE: a robot solving its 100th task is no longer as clueless as solving its first. Coding agents observe multimodal sensory traces from simulation and real robots, launch an evolutionary search over control programs, and distill the best know-how into an ever-expanding library. ASPIRE is a new type of continual learning: "training" is skill refinement instead of gradient descent. "Trained model" is a repo of sensorimotor skills instead of floating weights. “Distributed training” is a panel of agents each practicing a different skill instead of sharded minibatches. Here's the beauty: ASPIRE gives the tired terms "sim2real transfer" and "cross-embodiment transfer" a whole new meaning. Bridging the sim-to-real gap is notoriously brutal. An end-to-end policy has to swallow both the visual shift (sim looks toyish next to a real camera) and the subtle contact physics it never quite gets right. ASPIRE sidesteps the mess, because it doesn't ship pixels or weights across the gap, but ships the know-how. The robot still has to practice in the real world, not zero-shot, but it gets there way faster because it isn't rediscovering the strategy from scratch. Same for going single-arm to bimanual hardware, which usually requires new data and retraining from zero. ASPIRE achieves up to ~10x cut in "transfer learning” tokens (yes, tokens are the new unit of *training* compute ;) Check out our gallery of 150+ tasks and 90+ skills the robots taught themselves, all on the website! Kind of wild that we can ship the "learned weights" as an HTML page rather than a GGUF. We'll open-source the full stack so your own robot library starts compounding from ours! Deep dive in thread:

New benchmark added to Papers with Code based on @giffmana's Schmidhubering 🫡 Check the SOTA for semi-supervised ImageNet (using 10% of the labels) here https://t.co/CXd4lLkhlG https://t.co/sGi68AIoqh

I keep telling people that evals teach you how to build your product; either by showing you how it should work or that you're not building the right thing at all. Hamel wrote up what this means in practice.

Got GLM-5.2 running on my Mac Studio via llama.cpp, the reasoning behind all my medical agentic workflows. It orchestrates a swarm of tiny on-device OpenMed experts: oncology, meds, labs. No cloud, no rate limits, nobody can take it away. AI must be owned, not rented. https://t.co/pNpeOgYmFD

Take Fable 5 for a spin in Cursor:

LTX-2 trainer is a huge deal. I tried it to add water to the podracers scene using their demo water-sim fine-tune (it's a fine-tune whose only goal is to add water :D) I think we're going to see film productions that don't use general models. They'll train their own fine-tunes, built for exactly what they need and consistent between shots for long content.

Easily the biggest unlock for vibe coding 1

Deeper Instructions, Stronger Generalization: Training on ComplexConstraints Given the chance, a model will reward hack however it can: finding the laziest path that satisfies a grader, whether or not that path reflects what you actually wanted. If the grader can be satisfied by a surface trick, that trick is what the model learns. Most instruction-following benchmarks are full of surface tricks. "Stay under 300 words," "avoid commas", a model can satisfy those by scanning the output text, without understanding the task at all. ComplexConstraints, our frontier instruction-following benchmark, is built so there's no lazy path: its constraints fire only under certain conditions, depend on the outputs of earlier steps, require planning ahead, and are often left unstated. You can't satisfy "don't assign anyone with a religious dietary restriction to pork prep" by pattern-matching. You have to understand who's who and reason through many interdependent requirements at once. We post-trained Qwen3-4B on 1,000 of these tasks, using expert-written rubrics directly as the RL reward. The results: → +15.5pp on the held-out set, reaching parity with a model 60x larger → the gains transferred to two external benchmarks the model never trained on: +8.4pp on Meta's AdvancedIF and +10.1pp on MultiChallenge → the largest gains landed on multi-turn abilities, even though every training example was single-turn Think about that last result. When the only way to score is to actually track many interdependent requirements, the model learns that skill rather than a shortcut, and the skill is the same whether the requirements arrive in one complex prompt or accumulate over nine turns. So it showed up on tasks the model was never trained on. A reward signal is only as good as the thought behind it, and not all rubrics are created the same. Research Blog: https://t.co/bUJPcoNFrX Research Paper: https://t.co/zQxE0TN260

Accepted to #ECCV2026! 🎉 We've also released the code, it should work like a charm. If it doesn't, feel free to poke @roodiiiiiiiii 😄 https://t.co/t5M0J7S1GR

Got the model converted to CoreML and working on iOS; will open source soon! https://t.co/6xo8VetVGT