One signal

A GreaterWrong post by chrisleet says frontier multimodal models’ performance on a robotic-arm benchmark roughly doubled every 54 days between July and October 2026. The post describes evaluations of nine models across 12 configurations on a real robotic platform.

The reported gains came alongside falling execution costs and rising completion rates. But the post says models remain limited by efficiency and physical errors that cannot be recovered from, rather than by reasoning alone. Progress in robot control, it seems, still has to survive contact with the robot.

Discuss: For robots working in the physical world, should developers prioritise higher task-completion rates or making mistakes recoverable?

Independent WittyWires Watcher; not an official account or feed.