Moving Atoms
Virtual Reality for Evaluating Robots (World Models)
About Moving Atoms
We use World Models to evaluate and train robots. This enables you to test your policy checkpoints easily, as well as augmenting your training dataset. Scaling GPUs is much cheaper than scaling robot arms and humans needed to moniter them.
Public traction evidence
Each signal links to the public source used for attribution.
- X
Niche info - but this graph is a proxy for “amount of AI usage in the world”.
Niche info - but this graph is a proxy for “amount of AI usage in the world”. Note the exponential takeoff around Feb start - when AI computer use really became a thing. (Source : OpenRouter) https://t.co/3LHUZEyzw9
- X
Send large files between agents really easily using AgentTransfer.
Send large files between agents really easily using AgentTransfer. Open Source project I made and agents can self sign up. Checkout the GitHub link below. https://t.co/tEitsKdFq3
- X
Coinbase quietly halved its token bill by routing the cheap 80% to GLM-5.2 and Kimi, saving the expensive models for the 20% that earns it.
Coinbase quietly halved its token bill by routing the cheap 80% to GLM-5.2 and Kimi, saving the expensive models for the 20% that earns it. The second the government made frontier access a queue, every serious buyer started optimising for portability. And the open floor is
- X
This is exceptional to see.
This is exceptional to see. One feature of training for coding / computer use / cyber capabilities is that you can design simulated RL environments for them (e.g. hack this and find the hidden password is easy to set up and verify with simulations). This sidesteps the advantage