⚡️Few-step diffusion without a teacher, self-distillation or JVPs. iDDM: a single model trained from scratch in one stage. ImageNet-256: 4.48 FID at 4 steps and 2.38 at 50, from the same checkpoint 🧵 (1/6)

I have an answer for this, but you guys might not like it

Live now: our Computer Use Track from AI Engineer World's Fair 2026. Nobody is going to rebuild the web for your agent. piped.video/watch?v=Ki980nV0… - Computer Use models will agentify the web, not APIs: @DhruvBatra_, Yutori - Computer Use at the Edge of the Statistical Precipice: @proceduralia, Programma Labs - Bringing agents onto the world wide web: @pk_iv, Browserbase - The Dark Arts of Web Automation: @CoreyGallon, Rexmore - From RL to IRL: Gaurav Mishra, Amazon AGI Lab - The Rise of CaaS, context as a service: Omer Primor, Bright Data - Computer-Use 2.0, agents just got multi-cursor: @francedot + Dillon DuPont, Cua - Persona Engineering, a field guide to synthetic personas: @ianand, InsightSciences.ai - How Web Data Infrastructure Powers the Next Generation of AI: Patricija Žemaitytė, Oxylabs

Euro: Sicilian edition

I agree, and there's more to it. There is a entire class of problems that I call deployment-complete"reinforcement learning problems. Computer use belongs to this class. RL works well when one can easily reset the environment to diverse enough situations, when there is a clearly-defined reward signal, when one can easily generate trajectories in parallel, when the agent can generate actions without worrying about safety. Guess what? For most realistic computer use situations, you can't do any of this off-the-shelf. It's quite surprising, isn't it? It's stuff on a computer but it feels like blockers are all over the place not differently from what you would see in the physical world. To me, it almost feels like training to train a robot. And yes, as you can guess, also robotics is a deployment-complete reinforcement learning problem. So what do people do in robotics to overcome this problem? They build simulators, environments that mirror the one they ultimately care about but that satisfy the properties above. The advantage of computer use compared to robotics is that building environments is potentially easier, because you don't need to model the clumsy physical world. One could say you just gotta build a bunch of websites/apps, that are coincidentally the easiest thing to build with modern AI agents. But then why isn't CUA already solved? Because, in reality, to build an environment that is simultaneously useful and scalable for a deployment-complete RL problem, the environment shouldn't be just a bunch of websites, but it needs to be designed with extreme craft and effort. There are many things that could be said about what this extreme craft entails, but to me, it means that it should at minimum follow this set of design principles: - Privileged verification (P): Task completion must be verified as much as possible against the environment’s internal ground-truth state. - Realistic environments (R): Environment's apps must faithfully reproduce real-world apps. - Integrity-checked configurations (I): Every environment configuration must be automatically verified. - Sandboxed execution (S): The environment must be fully self-contained and must support deterministic reset to any specified initial state. - Multifactorial variability (M). The environment must vary independently along multiple axes. I call them the PRISM principles. In practice, they entail pretty hard engineering and research challenges. But they define what needs to happen, at minimum, to have a good RL environment for CUA. If you want to learn more about this, and how this is also the only way out of the related CUA evaluation crisis, come to my talk at the @aiDotEngineer conference next Wednesday in SF, or reach out!

IMO in Europe and especially in Italy the overlooked bottleneck is the inability to create a community that simultaneously shares values and technical chops to move things forward. Such communities build hope, energy, and strong bottom-up leverage. In Silicon Valley, skilled and aligned people find each other, encourage each other, and sustain each other way more, with open mind and curiosity. This accelerates worthwhile change. If this dynamics is not there, the largest "community efforts" naturally remain top-down and regulatory, like many of the EU ones. I see things changing, but too slowly. Maybe this wasn't as important in the pre-AI tech world? But honestly now it feels to me like a hard prerequisite to effective use of capital.

I myself didn't fully understand what the deepseek r1 hype was about, but at some point I realized where the mismatch in hype was coming from. For many people with reinforcement learning knowledge, the o1 announcement was pretty clear. o1 was a model trained with reinforcement leaning to dynamically generate chain of thoughts, using verifiable rewards. I mean, I didn't know the details of their training recipe or anything like that. But the general strategy was pretty much implied by their blog post and release. To those people, r1 is an open source reproduction of the same, implied, strategy, or at least, the first one that got really good performance. That is extremely valuable as an engineering contribution, but far from being an actual scientific breakthrough. I came to learn that many other people, typically without a technical RL background, instead consider deepseek r1 the first model trained to reason with reinforcement learning and verifiable rewards, associating o1 to a possibly different paradigm (e.g., MCTS). It's honestly a good reminder of how history (e.g., of science) can feel very different depending on your perspective

I have decided to leave Meta Superintelligence Labs. I’ve been at Meta for more than three years, starting during my PhD. I’ve learned a lot during this time, and I’m grateful to have worked with many amazing collaborators. We advanced the frontier of LLMs, reinforcement learning, and computer use agents. I’m pretty proud of our latest work on rigorous agent evaluations and CUA envs (check it out: arxiv.org/abs/2605.08261). Looking forward to taking some time to explore what’s next!

now THIS is cooking hard

You deserve more than a crowded Blue Bottle and a batch of 400 startups... I'm launching Kernel Grants, a pre-seed program that will invest $271,828 in 10 founders/year who are building tooling and infrastructure for the token factories of the future. We have an amazing set of speakers lined up for our first set of events: - @pirroh, President of Replit - @soumithchintala, CTO of Thinky - @jeremyphoward, Founder of Answer.ai - @NaderLikeLadder, Dir. of DevTech at NVIDIA - @OfficialLoganK, MOTS at DeepMind (and first ever Latent Space guest!) - @clattner_llvm, Founder of Modular - @dylan522p, Founder of SemiAnalysis - @swyx, Editor of Latent Space (+ AIE, Cognition, etc!) Batches are a relic of pre-AI acceleration. Any day is a great day to start building, so applications are open and we accept founders on a rolling basis. Let's build! kernellabs.ai/grants Enjoy an exclusive tour of our Kernel space 👀

enjoy!

this assumes that intelligence scales nicely and uniformly, that failing small pattern tasks means failing big science. but intelligence is asymmetric. humans who crack deep problems often stumble on simple ones. why should we aim for AI to be different?

New Anthropic Fellows research: Abstractive red-teaming of language model character The worst way to find out about a character flaw in your language model is from a viral screenshot. How can we find these issues before deployment, rather than after? In this work, we introduce abstractive red-teaming, a new approach that searches over natural-language categories of queries, rather than individual prompts.

wait, if the test for AGI is hard games, why don't we just use existing hard games as the benchmark? I feel like arc-agi-n is supposed to be basically nethack or dark souls... but we already have those, no reason to wait to build them this was the original idea of the pioneers of modern reinforcement learning (e.g., @_rockt,). why should we go in circles?

In the limit, what's important is our ability to adapt.  What is a good recipe for teaching agents to adapt on-the-fly? We introduce two meta-learning for LLMs papers written with @JonnyCoook at @GoogleDeepMind. This is research from last year we can finally share 🧵👇

Only words of appreciation for the first Sicilian AI Workshop organized by @proceduralia “et al”. This is an example of local AI community that benefits everyone! Plus fantastic food & you’re by the sea. I hope this will grow in the archipelago of Mediterranean schools/confs ☀️☀️

Our paper Perception Encoder (PE) was accepted as an Oral at NeurIPS 2025! 🎉 We show that scaling simple contrastive vision–language pretraining, refined with a robust video data engine, is enough to build a universal vision encoder. 🔹 Strong embeddings for image & video tasks (cls, retrieval, VQA, detection) 🔹 Alignment methods unlock multimodal & spatial understanding 🔹 Open models, code & new datasets incl. human-annotated videos 📄 Paper: arxiv.org/pdf/2504.13181 💾 Data & Checkpoints: tinyurl.com/48y2h8k9 🔧 Code: tinyurl.com/sp8sdd9h #NeurIPS2025 #multimodalai