Coding Is Not All You Need: CMU author argues GPT-6's robot tasks hit a world-model wall
ceciletamura · x · 2026-09-12
Hokin Deng (CMU) pushes back on the viral "Coding is All You Need" narrative sparked by GPT-6 Astra scoring 95% on a robot-arm block-stacking task (vs 40% for Fable 5.1), with 6.2x fewer output tokens at 2.3x lower cost.
- The same report shows a fine peg-insertion task with millimetre tolerance where both new and old models score just 2/20 — the authors admit they "hit the same wall."
- Deng argues that wall is not about scale or code but the oldest problem in intelligence research: intelligence began as bodies sensing and moving, before language.
- Citing a century of single-cell learning experiments (including Gershman et al. 2021 on conditionable Paramecia), the piece contends the key organ is a world model that can run offline for memory, imagination and planning — language came last, and machines acting in the world need that organ, not more code.
More from AGI Musings
- Fruit Fly Brain Simulation Solves Rubik's Cube, Igniting Consciousness Debate — sebkrier · 2026-09-12
- jd_pressman: the AGI bottleneck is institutional, not alignment progress — jd_pressman · 2026-09-12
- Roko argues human expected utility hinges on P(Banksian Culture), not alignment — RokoMijic · 2026-09-12
- Riley Ralmuto Posts the Connor Leahy Interview Video Behind His Viral Critique — RileyRalmuto · 2026-09-12
- Riley Ralmuto Rips Into Connor Leahy's Doom Claims: Assumptions Stated as Facts — RileyRalmuto · 2026-09-12
- What do we actually gain after the Singularity? A Reddit thread seeks the non-doomer case — Assistant-Klutzy · 2026-09-12