ByteDance AI Lead: Real Generalization Comes from Data and Environments, Not RL
shuchaobi · x · 2026-08-08
Shuchao Bi (ByteDance AI lead) reflects on his talk from 14 months ago, stating his AGI predictions remain relevant. He argues that Reinforcement Learning (RL) is inherently mode-seeking with limited generalization, and that true generalization power stems from data and environments.
His team is currently focusing on two core streams:
- Genius Kid: Focused on pure reasoning and raw intelligence capable of winning Olympic gold medals. This phase is complete, and they are moving to the "Professor X" phase to scale theoretical scientists.
- Omnipotent Assistant: Leveraging intelligence for digital tasks, where the key challenge is building and digitalizing environments.
More from AGI Musings
- Labs Won't Share Safety Research: Reward Hacking Blocks New Releases — willccbb · 2026-08-08
- Zuckerberg on Beating Giants: Big Companies Lack Conviction, AI Mirrors Facebook's Disruption — r0ck3t23 · 2026-08-08
- Pedro Domingos: Research Freedom in Corporate AI Labs Never Lasts — pmddomingos · 2026-08-08
- Will AI Get Cheaper? Competition and Compute Costs to Offset Subsidy Loss — intellectronica · 2026-08-08
- Apple's 1987 Knowledge Navigator Video is Becoming Reality — LukeW · 2026-08-08
- Models trained to delegate and coordinate, security threat narratives overblown — dbreunig · 2026-08-08