Porter says OpenAI's Astra loses to his lab's own model
A reply to Freda Duan's thread calling Astra robotics' GPT-3 moment turned into a ranking claim nobody outside his lab can check.
Brad Porter says OpenAI's Astra is the best generalizing robot model he has seen publicly, and that it still loses to the general model running inside his own lab.
He said it in a reply to Freda Duan, who had posted ten takes after talking to 50 people at the frontier of robotics. Her first take was that Astra is robotics' GPT-3 moment, and that some in robotics seem reluctant to admit it because it did not come from a robotics company. Porter agreed Astra is promising and may offer a path. Then he added the part people latched onto, that it is not performing better than his lab's general model and that he suspects they are ahead of everyone right now.
it’s not performing better than our general model in our lab, but we suspect we’re ahead of everyone right now
The System 1 argument
Porter's second point is about speed. Distilling Astra into something that runs at production performance may not be easy, which he says is why people still argue System 1 performance, the fast motor skill layer, matters. Duan made a version of the same split in her thread, framing the startup case as Astra handling System 2 reasoning while specialized models keep System 1 skills and System 0 motor control. Her question was whether that boundary holds.
Either way, Porter expects the field to get there. He says the industry will have general models in 2 to 3 quarters, whether the frontier labs fix their latency bottlenecks or model builders like his solve System 1. When @thelcjones pushed on the timeline, Porter answered that it is a couple of quarters, not a couple of years. That matches the window he gave earlier for general dexterous manipulation.
Shown versus claimed
Nothing here was demonstrated. No outside party has run Astra against Porter's in-lab model, so the ranking is his word. He also hedged it himself, saying they suspect they are ahead rather than stating it flatly.
The sharper claim in his post is the one about data. If Astra really does offer a path, Porter says it kind of invalidates most all of the data collection going on, which cuts against take six in Duan's thread about million-hour procurement programs and talk of budgets approaching $1B per frontier lab. Duan's own closing take was that the next bottleneck is not intelligence but deployment. Porter agreed with that part, saying the bottleneck was always going to be hardware and deployment, and to always build to where the puck is going.
Astra IS robotics' GPT-3 moment. Ironically, now that it's finally here, some in robotics seem reluctant to acknowledge it.
Faster than a couple of years… a couple of quarters
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.