Tian Keyu's world-model startup raises ~$30M at ~$200M valuation, teaching AI its own 200k-symbol language

teortaxesTex · x · 2026-10-08

Tian Keyu—known for the ByteDance intern controversy and later a NeurIPS 2024 Best Paper first author—has founded a world-model startup, per a Bloomberg interview: 10 people, no name or product yet, $30M raised from 5Y Capital, IDG and others at a $200M valuation.

The core idea:

Human language is fundamentally ill-suited to describe the world—text can't capture object positions, motion, lighting and physics in video.
So don't use human language at all: the team built a 200,000-symbol lexicon, incomprehensible to humans, that AI uses to represent, compress and predict video.
They plan to feed in 100 million hours of video so the model learns the physical world through its own language.

The goal isn't video generation but a foundation model that understands space, motion and physics—for robotics, autonomous driving and interactive video. Tian claims the approach already improves video generation efficiency by at least an order of magnitude. A public demo or livestream is planned first, full model targeted for 2027, no commercialization yet. Commenters see it as a spiritual evolution of VAR and a world-modeling paradigm superior to diffusion, inheriting LLM strengths.

Original post →

Tian Keyu's world-model startup raises ~$30M at ~$200M valuation, teaching AI its own 200k-symbol language

More from Venture