The internal model they used to solve the Navier-Stoke's problem was significantly better than the public Astra model, and they also used 10,000 agents.
Astra wasn't even released 3mo ago. It would not be surprising in the slightest that a public model from 3mo would not be capable of solving this problem, but an internal one from current day would be.
my speculation is that they have math-specialized model retrain, so it doesn't need to have all world info in weights, but can focus on math RL training.
Nah.
The model is probably comically big and inefficient but big enough
Finally, size really does matter!
The internal model they used to solve the Navier-Stoke's problem was significantly better than the public Astra model, and they also used 10,000 agents.
Astra wasn't even released 3mo ago. It would not be surprising in the slightest that a public model from 3mo would not be capable of solving this problem, but an internal one from current day would be.
my speculation is that they have math-specialized model retrain, so it doesn't need to have all world info in weights, but can focus on math RL training.