Walden talk: base models improving fast, mid-training data commoditizing
ZeYanjie · x · 2026-11-14
A slide from a Walden talk at CoRL 2026 (referencing Astra, Fable, etc.) lays out core beliefs held since the company's founding: base models will get better fast, so plan for it; mid-training data is being commoditized and stronger base models need less of it; deployment data remains highly valuable—even a PhD-level manipulator still requires on-the-job training in the factory; and stronger base models actually make deployment data more valuable. A notable strategic take for robotics and embodied AI data strategy.
More from Embodied
- XREAL AURA Founder Edition ships Nov 17; Scoble compares pricing to China's drone-industry playbook — Scobleizer · 2026-11-14
- Panasonic builds 250-strong team to teach Physical AI veteran craftsmen's expertise — CyberRobooo · 2026-11-14
- Runway open-sources Praxis-1, a video-trained AI model that controls robots — The Decoder · 2026-11-14
- Scoble: Snap Specs are the best AR glasses on the field, if you ignore price — Scobleizer · 2026-11-14
- Europe's robotics opportunity is retrofit: cognitive robotics for existing plants and ageing grids — ingliguori · 2026-11-14
- Tesla's no-wheel vehicle called overpriced; KOL: limited supply means high pricing is just Business 101 — AIFlow_ML · 2026-11-14