Is this Yann LeCun’s first paper after leaving Meta? It demonstrates how humanoid robots can mimic actions from AI-generated videos, which are often too noisy for direct imitation. The system lifts the video into 3D keypoints and then uses a physics-aware policy to execute the motions, enabling zero-shot control. They implemented this on the Unitree G1 humanoid robot.

Introducing QVAC Fabric LLM: The framework that brings full AI inference and fine-tuning to your hardware. Execute and Fine-tune modern models like LLama3 and Gemma 3 on your laptop, your consumer GPU, and even your smartphone. No cloud. No compromise. You control your data. QVAC Fabric LLM is open source. It’s decentralized, hyper-scalable, antifragile, user-controlled AI. QVAC, your device, your AI Learn more: huggingface.co/blog/qvac/fab…

New on our Frontier Red Team blog: We tested whether AIs can exploit blockchain smart contracts. In simulated testing, AI agents found $4.6M in exploits. The research (with @MATSprogram and the Anthropic Fellows program) also developed a new benchmark: red.anthropic.com/2025/smart…

Sim-to-real learning for humanoid robots is a full-stack problem. Today, Amazon FAR is releasing a full-stack solution: Holosoma. To accelerate research, we are open-sourcing a complete codebase covering multiple simulation backends, training, retargeting, and real-world inference.

Suno just announced their partnership with Warner Music Group. Here’s what this means:

Sergey used Gemini in a surprisingly creative way. He asked it inside an internal Google chat, “who should be promoted in this chat space?” AI picked up a young woman engineer who is not vocal, and she was promoted. AI can reveal value that human systems overlook. AI can help CEOs of Big Corps spot true talent rather than people who are just loud or political, and it's probably more objective than the middle managers. How far are we from an AI CEO? Sergey Brin in founder mode using AI to cut through bureaucracy is just fascinating.

no words

my friend & i were laughing about how being a founder right now is basically a weekday hobby. ai made day 0 stupidly easy because you can summon a name, a logo, assets, a half coherent product demo, all before your coffee cools. it’s insane & kind of beautiful. but that ease is exactly why day 2 is the real crucible now more than ever before. because once the vibes fade, everything has to survive contact with reality which are users, distribution, retention, you know the whole brutal ecology of actual demand (it’s so fucking hard). if day 0 got frictionless, day 2 got carnivorous. e.g. i remember talking to a vc backed founder recently who didn’t know what retention was, which is absolutely remarkable. what a strange moment… you got infinite prototypes, zero scarcity of ideas, & yet an ever shrinking pool of things that actually endure. fascinating time to be alive. of course i wrote a piece about it below.

Let me explain exactly why VLC is free despite 6B downloads, because no one seems to get it. VLC doesn’t make money because making money would destroy the only thing that made it reach 6 billion downloads in the first place. The player grew through a specific distribution loop: tech-savvy users install it once, it works perfectly on every weird video file they throw at it, and they recommend it to everyone forever. IT departments deploy it across entire companies. A Reddit comment from 2009 still drives downloads in 2025 because the answer never changed. That recommendation engine dies the second ads appear. Not slowly. Immediately. The users who drive VLC’s distribution are the exact people who understand what ads mean. Your incentives just switched from “make the best player” to “maximize impressions.” They see it, stop recommending it, and your growth engine shuts off. Run the actual numbers. VLC gets maybe 50 million active users daily across 6 billion total downloads. Typical video player ad rates run $1-3 CPM. Even if you served ads on every playback session, you’re looking at maybe $50-150 million annually at absolute peak optimistic assumptions. Sounds like a lot until you realize what Kempf actually traded it for. VLC reaching 6 billion people made Kempf the person who built the infrastructure everyone depends on. He runs a video consulting business. He built dav1d, an AV1 codec that powers modern streaming. Being “the guy who kept VLC free” opens every door in video technology. Clients pay him to solve problems because he proved he optimizes for quality over quick monetization. “Former ad-supported media player executive” gets you exactly zero of that leverage. The people celebrating Kempf’s ethics are missing the calculation. He didn’t sacrifice millions for principles. He rejected $150M in highly uncertain ad revenue to build permanent positioning worth multiples of that in everything else he touches. VLC free generates more value for Kempf than VLC monetized ever could. The trade was never even close.

Everyone is sleeping on Meta's SAM 3 release. But it's actually a big deal. Here's why: Companies spend millions paying humans to label images and videos frame by frame. A single autonomous driving dataset? Months of work, hundreds of annotators, millions in cost. Without labeled data, you can't train custom models. Without custom models, you're stuck with generic solutions. This is why most companies never move past pilots. SAM 3 breaks this cycle. First let's look at the evolution: SAM 1 segmented objects when you clicked on them. Revolutionary, but one object at a time. SAM 2 added video tracking with memory. Game-changing, but you still manually prompted every object. SAM 3 changes everything with text prompts. Type "yellow school bus" and it finds ALL of them in your image or video. Not just one. Every instance across thousands of frames. Now here's where people get confused: "Can't I just use GPT-5 or Gemini for this?" No, and here's why that's a terrible approach. Large multimodal LLMs are great for reasoning, but they're slow and expensive for production visual tasks. You're paying API costs per image, waiting seconds for responses, getting inconsistent results. SAM 3 runs in 30 milliseconds on a single GPU for 100 objects. That's 100x faster, and you own the infrastructure. More importantly, SAM 3 gives you precise pixel-level masks, not descriptions. Try asking an LLM to segment every defective part on a manufacturing line in real-time. It won't work. SAM 3 does this effortlessly. The real breakthrough is their data engine. Meta built an AI-human hybrid system that's 5x faster for complex annotations. They trained SAM 3 on 4 million unique visual concepts - 50x more than existing benchmarks like LVIS. SAM 3 is trained on 4 million unique visual concepts, it handles everything: - Text-based concept search - Interactive refinement with clicks - Video tracking across frames - Zero-shot detection of new concepts The model is open source. Weights, code, and benchmarks are on GitHub. If you're building computer vision applications, this is the foundation model to evaluate. The annotation time savings alone will pay for integration costs within weeks. Find the relevant links in the next tweet!

Real-time video comprehension and generation is the future

I’m starting to get into a habit of reading everything (blogs, articles, book chapters,…) with LLMs. Usually pass 1 is manual, then pass 2 “explain/summarize”, pass 3 Q&A. I usually end up with a better/deeper understanding than if I moved on. Growing to among top use cases. On the flip side, if you’re a writer trying to explain/communicate something, we may increasingly see less of a mindset of “I’m writing this for another human” and more “I’m writing this for an LLM”. Because once an LLM “gets it”, it can then target, personalize and serve the idea to its user.