Why OpenAI's Astra crushed specialized robotics models: a scaling-transfer theory
willdepue · x · 2026-11-01
Reacting to OpenAI's result where Astra decimated specialized robotics models, willdepue offers a gut analysis: no flops-based justification exists (specialized models likely spent equal or more compute on robotics-like data); Astra's encoders probably aren't better; chain-of-thought helps robotics a lot even at low reasoning budgets (bigger pretrain → better CoT → better tasks); and transfer increases with scale for both optimization/parameter reasons (hard to probe without GPT-4-scale ablations) and data-scale reasons, since larger models extract mutual information between seemingly unrelated datapoints that smaller models cannot.
More from Embodied
- Naval predicts 4-passenger drones will eventually end commercial air travel, with a long interregnum ahead — naval · 2026-11-01
- Meta Quest game scans your real room and turns it into a haunted house — Scobleizer · 2026-11-01
- RoboAuto: Robot Program Search With Multi-Level Physical Understanding — DJiafei · 2026-11-01
- Developer dresses robot 'Bob' as a ghost to hand out Halloween candy — chrismatthieu · 2026-11-01
- GRASP authors compare with SuperMimic: extracting more optimization signal from sim experience — hankyang94 · 2026-11-01
- Harvard's GRASP refines human demos in physics sim for dexterous hands and Unitree G1 — hankyang94 · 2026-11-01