Import AI 466: Jack Clark Argues the Bitter Lesson Has Arrived for Robotics — Anthropic's Opus 4.7 Cuts a Robot Task From 181 Minutes to 9
Clark's July 27 issue pairs two results into a single thesis: Anthropic showed Opus 4.7 completing a robot task in 9 minutes against 181 minutes with earlier technology, and Sunday Robotics' ACT-2 hit a 99.1% ±0.3% success rate across 785 autonomous laundry-folding attempts spanning 9 garment types in homes the robot had never seen, with zero home- or garment-specific tuning at deployment. Sunday's recipe is 'scale pretraining, then hill-climb with minimal in-house data,' and their key observation is that as the pretrained model strengthens, gains from small in-house datasets become increasingly transferable. Clark's read: stronger foundation models unlock robot capability without robotics-specific training, so if generalization keeps improving, robotics accelerates on someone else's scaling curve.
↳ Follow the thread