GPT-6 moves a camera to insert a flower, yet burns an egg: reflections on the Astra report

liuziwei7 · x · 2026-10-27

Liu Ziwei shares Baiqiao Yin's long essay reflecting on the GPT-6 Astra tech report. Two contrasting scenes: with a two-armed robot inserting a flower, GPT-6 swings its free arm over so that camera can see the vase's mouth — a move rarely present in teleoperated demos. Yet in a live kitchen, it leaves the fridge door open for minutes and unhurriedly considers turning off the burner only after the egg is done, burning it.

The essay's thesis: everything it can do is in the books; what it lacks is known only to a body that has touched the world — and its training world never demanded touch. GPT-6 makes the questions "must representations be hand-designed?" and "must physical understanding come from vision?" obsolete (both no). The question worth asking: an agent grows the capabilities its world demands, so design the pressure, not the representation — what world would force it to grow what's missing?

Original post →

More from Embodied

Embodied channel →