Chjango Unchained⛓️

4 AI Breakthroughs Changing Robotics Forever: DrEureka, Voyager, Darwin-Gödel Machines & AlphaEvolve Explained x.com/i/broadcasts/1…

Earlier this year Yann LeCun left Meta because Mark Zuckerberg wouldn't bet the company on JEPA. Last week his group dropped the first JEPA that actually trains end-to-end from raw pixels. 15 million parameters. Single GPU. A few hours. The timing is not a coincidence. For four years Meta has been the house that JEPA built. LeCun published the original paper from FAIR in 2022. I-JEPA and V-JEPA came out of his lab. The architecture was supposed to be the escape hatch from LLMs, the path to robots that actually learn physics instead of hallucinating about it. Every version shipped fragile. Stop-gradients. Exponential moving averages. Frozen pretrained encoders. Six or seven loss terms that had to be hand-tuned or the model collapsed into garbage representations. Meta kept funding LLMs. Llama shipped. Llama scaled. Llama got beat by Qwen and DeepSeek. Zuck spent $14 billion to buy ScaleAI and install Alexandr Wang. The FAIR robotics group was dissolved. LeCun's research kept winning papers and losing the product roadmap. He left, started AMI Labs, and said publicly that LLMs were a dead end. Now the paper. LeWorldModel. One regularizer replaces the entire pile of heuristics. Project the latent embeddings onto random directions, run a normality test, penalize deviation from Gaussian. The model cannot collapse because collapsed embeddings fail the test by construction. Hyperparameter search went from O(n^6) polynomial to O(log n) logarithmic. Six tunable knobs became one. The downstream numbers are what should scare the robotics capex class. 200 times fewer tokens per observation than DINO-WM. Planning time drops from 47 seconds to 0.98 seconds per cycle. 48x faster at matching or beating foundation-model performance on Push-T and 3D cube control. The latent space probes cleanly for agent position, block velocity, end-effector pose. It correctly flags physically impossible events as surprising. It learned physics without being told physics existed. Figure AI is valued at $39 billion. Tesla Optimus is mass-producing. World Labs raised $230 million to sell generative world models. Everyone in humanoid robotics is burning capital on foundation-model pipelines that plan in 47 seconds per cycle. LeCun's group just showed you can do it with 15 million parameters on a single GPU in a few hours. This is the Xerox PARC pattern running again. Meta had the next architecture. Meta had the scientist. Meta dissolved the robotics team, passed on the productization, and watched the exit. Three months later the lab that was supposed to be Meta's publishes the result that resets the robotics cost structure. The paper is worth more than Alexandr Wang.

π0.7 just proved that more robotics data can make your model worse. what this means for robotics teams: 1. scale without metadata is pointless. without annotation density and quality scores, more data averages together conflicting strategies. the model degrades. 2. the data volume race is a trap. the teams that win won't have the most data. they'll have data their models can actually learn from. 3. PI and Standard Bots disagree on where the data should come from. they agree on what matters: structure over volume 4. the bottleneck is the loop on how fast you can organise, annotate, and feed data back into training.

Our newest model, π0.7, has some interesting emergent capabilities: it can control a new robot to fold shirts for which we had no shirt folding data, figure out how to use an appliance with language-based coaching, and perform a wide range of dexterous tasks all in one model!

Introducing GEN-1. Our latest milestone in scaling robot learning. We believe it to be the first general-purpose AI model to master simple physical tasks. 99% success rates, 3x faster speeds, adapts in real time to unexpected scenarios, w/ only 1 hour of robot data. More🧵👇

this weekend i learned something extremely important about writing software for manufacturing the person writing the code must be deeply entrenched in the factory operations and nuances of the workflows also i really dont know why you would buy any off the shelf manufacturing software anymore when you can custom build your own with all the nuance and with your own data and train your own models

🌐 Bridging Bitchat + MeshCore: Resilient communication when infrastructure fails Bitchat = Bluetooth mesh on phones you already have (~100m range) MeshCore = LoRa long-range mesh (km+ with cheap hardware) The bridge connects them. Your phone talks to the city-wide mesh network. Perfect for disasters, protests, internet shutdowns. Code: github.com/jooray/MeshCor… Releases: github.com/jooray/MeshCor… Read more: juraj.bednar.io/en/blog-en/202…

Beyond AI… China isn’t building towards 20,000 TWh of electricity generation so that they can power their datacenters for AI. Nobody needs 10,000 TWh to train AI models, and they aren’t doing that. So what are they up to? China is very obviously targeting the thing beyond AI datacenters, which is the enormous energy demands of ubiquitous robotics. Datacenters benefit from Moore’s Law, the law of ever improving compute efficiency. But mechanical actuators obey the laws of Newtonian motion which have no scope whatsoever for moving greater quantities of mass with ever smaller quantities of energy. China is skipping a whole paradigm. Their bet being they can backfill AI, once they have total domination over physical work. China is building the god body first (the robot fleet, or rather the infrastructure that allows it), and will build the god brain later. Probably speed running it, aided by espionage. America is building the god brain first and hasn’t really thought much about the god body. The two strategies are quite different and we should acknowledge this. The AI race risks being a strategic cull-de-sac, a pyrrhic victory, because the longest lead part of the future stack is building the energy system you need to operate an automated Newtonian economy at such a scale.

🇨🇳 China is scaling agricultural robots. Autonomous harvest at 24/7 cadence is the new baseline for food security. Vision models pick, arms place, logistics sync, human supervisors handle exceptions. Cheaper fruit, fewer bruises, happier supply chain

@brucefenton @ChrisPacia It sounds like you're just too poor to pay that, peasant

They voted to raise taxes 3 weeks ago. But that’s only half the tax burden. The School district is also voting to raise taxes again. For yet another luxury building. Plus, they are going to raise for next year by at least 8% and they have a proposed school bond which will be like $600 per $100k in accessed value.

Our Town Council keeps raising taxes. So I showed up with an English Lord’s wig to speak in *favor* of taxes. “If you can’t pay your taxes…don’t be poor.”

@brucefenton You're a fucking legend

Something big is happening in robotics - and it’s hiding in plain sight. This post is not about dancing robots but in the data that powers them. Open robotics datasets have exploded this year, turning the field into a more scalable and collaborative ecosystem. In just two years, @huggingface datasets grew from 11k to over 600k - and robotics is by far the fastest-growing segment. We went from 1k robotics datasets in 2024 to 27k in 2025! For comparison, text generation, the second-largest category, has only around 5k datasets in 2025. That gap is massive. Open datasets are important because robotics lives and dies by real-world robot data - video, actions, sensors, failures. By making this data easy to upload, reuse, and benchmark, researchers, startups, and large players are now releasing real-robot datasets that would have stayed locked inside labs just a few years ago. Major contributors include @nvidia, LeRobot initiative, and a rapidly growing maker community. This surge is also enabled by cheaper video storage, better tooling, and an open-source AI culture now spilling into the physical world. And it really matters: open robotics data dramatically lowers entry barriers, accelerates learning-by-doing, and speeds up progress toward generalist and humanoid robots. Robotics won’t scale through hardware alone - but to a large extent through shared data. Viz below from @aiworld_eu - link to the story and more viz/filters in comment.

Holy recession indicator

My CISO called me at 3 AM last Tuesday. "We caught someone." I asked, "Caught them doing what?" He said, "Typing." Let me explain. We have an employee in IT. Great worker. Always online. Never complained. Perfect Slack etiquette. One problem. His keystrokes were arriving 110 milliseconds late. One hundred and ten milliseconds. That's 0.11 seconds. The average American remote worker has 20-40ms of latency. This guy? 110ms. Every. Single. Keystroke. My security team ran the numbers. That latency doesn't come from a bad router in Ohio. That latency comes from Pyongyang. Our "Senior DevOps Engineer" was a North Korean operative. Running his work laptop through a laptop farm. In America. While he worked from a government building. In North Korea. He passed the interview. He passed the background check. He passed the vibe check. He did not pass the speed of light. Here's what people don't understand about physics: Light travels 186,000 miles per second. But it still has to go through China. And China adds latency. Since April, Amazon has caught 1,800 of these attempts. Eighteen hundred. I called an emergency meeting with my board. I said, "We need to implement Keystroke Velocity Auditing across all remote employees." They said, "That sounds invasive." I said, "You know what else is invasive? The Democratic People's Republic of Korea in your Jira tickets." They approved the budget. We now monitor keystroke timing to the microsecond. If your latency exceeds 60ms, you get a call from HR. If it exceeds 100ms, you get a call from the FBI. We've already flagged 47 employees. Turns out 44 of them just have bad Wi-Fi. 3 of them are "still under investigation." The lesson? You can fake a resume. You can fake a background check. You can fake an American accent on Zoom. But you cannot fake the speed of light. Physics is the ultimate background check. Hire accordingly.

@Defipeniel DrEureka is promising but in lieu of great physics simulators, isn't a complete answer yet

@chjango Looks like a deep tech lineup. Curious which breakthrough really shifts real-world robotics first

4 AI Breakthroughs Changing Robotics Forever: DrEureka, Voyager, Darwin-Gödel Machines & AlphaEvolve Explained x.com/i/broadcasts/1…

China now has farms that plant, grows, and harvests automatically. This could scale really fast in a manufacturing powerhouse like China.

GITAI robots cooperatively assemble a 5-meter tower, a building block for future off-world habitats, on its own. The combination of AI and robotics creates the necessary technological breakthrough and acceleration that we have always hoped for.

Always important to remember that a lot of these robots are "faking" the humanlike motions -- its a property of how they're trained not an inherent property of the hardware. They're actually capable of way weirder stuff and way faster motions.

Combat robot from Northeastern University in China.

San Francisco could use these lmao

A simple idea. Let robots collect the data that current foundation models are missing. A robot that gets better by doing real work in the real world. For two weeks in the Stanford East Asia Library, Scanford scanned shelves, helped librarians, and improved the vision language model it depends on. The idea is very simple: Robots do useful work. They gather the real world data foundation models never see online. They fine tune their own model They go back out stronger A full loop. What they found in deployment: ✅ 2103 shelves scanned with multilingual, faded, occluded book spines ✅ 18.7 hours of librarian time saved ✅ Book ID accuracy jumped from 32.0 percent to 71.8 percent ✅ English OCR improved from 24.8 percent to 46.6 percent ✅ Chinese OCR improved from 30.8 percent to 38.0 percent The most interesting part is the shift. Robots do not only consume foundation models. They create the data these models are missing. A clean robot powered data flywheel. Work. Collect. Fine tune. Repeat. Thanks for sharing, @jenngrannen! If you want the full write up: 📍Website: scanford-robot.github.io Paper: arxiv.org/abs/2511.19647 —- Weekly robotics and AI insights. Subscribe free: scalingdeep.tech