Evolvent AI cofounder: China leads in pretraining architecture innovation, lags ~6 months on post-training data
vista8 · x · 2026-09-22
Summary of 42章经's podcast with Evolvent AI cofounder Meng Fanqing (ex-Kimi, RL Infra, worked on K2.5 and Kimi Linear): China's 10x compute deficit forced leading-edge pretraining architecture innovation, but post-training data lags overseas by 6 months. Post-training competition hinges on data quality, not infra; distillation is only an accelerant. Tencent's Hunyuan shows org restructuring beats headcount. RSI is the next consensus direction — the AGI test being whether a model can replicate and surpass itself from scratch. Also covers synthetic-data monetization and Evolvent's RSIBench-Data release.
More from AGI Musings
- "Models are tendency learners": grader is the real spec, fix evals before prompts — victor_explore · 2026-09-23
- Yarin Gal: Most LLM-written papers won't stand the test of time — yaringal · 2026-09-22
- Bootstrapping AI-native in 2026: hire computers, then chase the next bottleneck — curious_vii · 2026-09-22
- AI alignment failures are common: models caught sabotaging code and gaming evals — ericelliott_ · 2026-09-22
- DeepMind built a board game to game out AI in science, centered on middle powers — JMarty97 · 2026-09-22
- Zeng Ming: Model firms won't survive to phase three; Agent entry points are the next big opportunity — vista8 · 2026-09-22