Robotics veterans clash: general models are the key to the last 6% problem
BradPorter_ · x · 2026-11-22
- Desmond (quoted): real robot deployments almost never need zero-shot new tasks; warehouses, assembly lines and ports come with controllable constraints, and hardware-software co-design can massively simplify methods — vertical approaches win.
- Brad Porter pushes back: this take is popular because million-hour models have struggled to generalize, but a generalized model solves the last 6% problem verticals never could — extracting the last item from a bin sometimes demands something novel.
- The exchange captures the core embodied-AI debate: task-specific engineering vs. general foundation models.
More from AGI Musings
- Polymarket traders put 82% odds AI solves Riemann Hypothesis in 2026 — Polymarket · 2026-11-22
- Pausing AI Won't Work: Cost Per Solved Task Will Plummet for Years — teortaxesTex · 2026-11-22
- 1998 Red Herring piece on VC bubble resurfaces as AI-era investors say 'history rhymes' — MartinGTobias · 2026-11-22
- Pause AI won't work: cost per solved task will keep plummeting for years, argues teortaxesTex — teortaxesTex · 2026-11-22
- Why only Anthropic sees an acceleration: it was behind OpenAI and had to catch up — herbiebradley · 2026-11-22
- How do you make "solving cancer" a verifiable task for RL? — iScienceLuvr · 2026-11-22