Embodied AI field debate: why VLA work is a 'influencer field' and pretraining is the steepest slope
ChongZzZhang · x · 2026-11-17
- A Chinese AI researcher lays out the "steepest slope theory" arguing against doing VLA/world-model research in embodied AI: the field lacks production practice and objective standards, demos are flashy hacks, and judgment of what's "good work" is monopolized by senior figures and their protégés.
- Cites Yao Shunyu's interview saying that without practice, multiple self-consistent frameworks emerge and authority decides who wins.
- Conclusion: anyone serious about robotics should join general LLM pretraining (eai in llm pretraining) — the highest-slope, most production-grounded route — framing the upcoming next transformer as a once-in-a-lifetime opportunity.
- The quoted tweet jokingly catalogs Chinese AI PhD discourse as "Big Bench theory," "Big Intern theory," "Big Fan theory," "Big Slope theory," etc.
More from AGI Musings
- Open weights don't democratize power: frontier compute gates decide who can build — tallmetommy · 2026-11-17
- Anthropic's amended S-1 says it hit its AI R&D capability threshold on expected future models — thione · 2026-11-17
- The technical value now is reinventing, not just implementing — neal_lathia · 2026-11-17
- Alignment researchers should raise the alarm publicly, not marginally improve model safety — zetalyrae · 2026-11-17
- Gary Marcus mocks AI bubble denial: Nvidia could fall to $10 and bros would still say no bubble — GaryMarcus · 2026-11-17
- Fields Medalist Maynard stunned by OpenAI's quasi-Riemann hypothesis proof — emax · 2026-11-17