World models should not be judged by what they can generate, but by what remains invariant when they are asked to render, simulate, and act on the same world.
Dr. Fei-Fei Li’s article is clearly walking in the correct direction (nice work!)
With passive, IR reflective markers too! That’s actually pretty hard considering occlusions.
I would have thought active LED markers would be required here.
100 people. Real-time motion capture.
A Chinese company just demonstrated simultaneous mocap of 100 performers in a single capture stage.
Imagine 100 people teleoperating 100 robots at the same time. The potential for large-scale embodied AI data collection is enormous.
Why LLMs are a dead end for human-level intelligence, and especially for Physical AI / Robotics.
The next leap isn’t bigger language models.
It’s World Models.
I just dropped a full 1-hour presentation from Shanghai:
“World Models: the ChatGPT moment for robotics?”
→ Why LLMs hit a wall
→ Why action-conditioned world models planning in latent space are the real path
→ Live World Forge demo with LeWorldModel + Hugging Face LeRobot
Watch here.
The future of intelligence is embodied, not just chatty.
“Skillify it” is the way you should write most tasks
Write markdown that makes code
Don’t make elaborate Foxconn factories to call agents
Let agents make their own tools, kaizen style
WTF is this maze cast into your automatic transmission?
It’s a mechanical computer that picks the right gear for you. Runs on oil, not electricity. Worked out decades before the microchip.
Here’s how it works:
- The inputs. Two oil pressures. One rises with road speed. One rises with how hard you press the gas. That’s all it knows: how fast am I going, and how fast do I want to go.
- The job. Route oil to the right clutch so the right gear engages for what you’re asking of the car.
- How it works. A spool, a small grooved cylinder, slides in a bore. Oil pushes each end; harder push wins. Speed pressure on one end, throttle plus a biasing spring on the other.
Slide one way it blocks a passage, slide the other it opens one: oil floods a clutch, the clutch locks, that’s your gear. Floor it and throttle pressure shoves the spool over, drops you a gear, the car surges.
Ease off and speed pressure wins, sliding it back to grab the next gear up. Two pressures, fighting it out.
- Why a maze? No valve decides alone; each one’s output is another’s input. So oil has to run from every valve to the others that depend on it, and oil can’t jump gaps. Every connection needs its own channel carved in metal. Dozens of valves, hundreds of passages, none crossing where they shouldn’t
- It’s a routing problem. Same one a chip designer faces: connect many things to many things without the paths colliding. They even stack plates with a drilled separator between them, like vias on a circuit board.
The result: a real-time computer in aluminum and oil. Only in old-school automatics. In a manual, you’re the computer.
I am trying to write a robotics sensor library that decouples the sensor data timing from the controller's sampling needs.
https://t.co/D4oOyEkzuM
I am open to PRs or suggestions.