Duan posts uncut Astra footage to show how slow it is
A real-speed clip of GPT-6 Astra driving YAM arms puts latency, not capability, at the centre of the argument.
Jiafei Duan posted a real-time rollout of GPT-6 Astra controlling MolmoAct2 YAM arms through a task, uncut and at real speed. His point is simple. People watching Astra demos on X may not realise how slow the system actually is once you stop trimming the dead time.
I applaud your patience if you make it through the whole video.
His verdict on the model is not a dismissal. Duan says there is potential, but he does not think it is practical for real-world robotics deployment yet. The clip itself is the argument. He signs off by applauding anyone who sits through the whole thing.
Why real speed matters
Most robot footage that circulates is cut, sped up, or both, and the edit quietly removes the part that decides whether a system can be shipped. A robot that completes a task is a different claim from a robot that completes a task fast enough for anyone to pay for it. Playing the rollout at real speed separates those two claims and lets viewers judge the second one for themselves.
Federico Martelli, replying to the post, put the same idea in a line. He said demo cuts hide exactly this gap and that latency is the unglamorous constraint that decides deployment readiness more than raw capability does. Duan replied with one word, "Truth".
latency is the unglamorous constraint that decides deployment readiness more than raw capability does
Where this sits
Astra has been the model everyone in the field has been poking at, and the tests have been piling up with researchers split on what it is actually good for. This post adds a different axis to that argument. It is not about whether Astra can reason about a scene or pick the right object. It is about how long the whole loop takes when nobody is editing.
What the clip does not tell you is where the time goes. Nobody in the sources breaks the delay down into model inference, planning, or arm motion, so there is no number to point at yet. What is on the table is one researcher's read, backed by footage he did not cut.