Summary

York Yang of Dyna reflects on 10 months of shipping robots into real customer environments. The core argument is that robotics will not follow an LLM-shaped or AV-shaped growth curve because it lacks both software-only distribution and a pre-existing installed base. The real “bubble” is in the assumption that pre-training scale alone will translate to commercial value without a deployment flywheel that closes the pre-training / post-training loop in production.

Dyna 的 York Yang 回顧了 10 個月將機器人部署至真實客戶環境的經驗。核心論點是機器人不會遵循 LLM 或自動駕駛的成長曲線,因為它既缺乏純軟體發行管道,也沒有既有的安裝基礎。真正的「泡沫」在於假設預訓練規模可以在沒有部署飛輪的情況下自然轉化為商業價值。

Key Points

  • Robotics bubble definition: gap between capability and expectations multiplied by time — not “will it work eventually” but “will it pay off within capital’s time horizon”
  • Unitree’s humanoid revenue: only ~9% true industrial deployment; most is research/education
  • Tesla Optimus: zero robots doing useful factory work as of Jan 2026
  • Hardware is not a distribution channel — a channel requires scene assessment, data routing, remote diagnostics, and continuous updates that compound across deployments
  • Post-training feedback loop (equivalent to RLHF for LLMs) barely exists in robotics; most teams still optimize pre-training as if scale alone transfers
  • Three camps: model-first, hardware-first, integration — Dyna argues only integration closes the loop today because layer interfaces are not yet stable

Insights

The deployment acceleration thesis (each deployment faster than the last) has not been demonstrated by any team in the industry, including Dyna’s own. This is the most honest admission in the piece and the clearest signal that the “channel as compounding asset” model is theoretical rather than proven. The field’s reliance on benchmark metrics and demos actively hides this because demos do not require the deployment flywheel to close.

Connections

Raw Excerpt

“The test of a channel is whether the next deployment is faster than the last one. If it is not, you have not built a channel. You have built inventory and PR.”