Shunyu Yao: RL finally works, marking the start of AI's 'Second Half'
dejavucoder · x · 2026-08-28
Shunyu Yao published a blog post 'The Second Half', arguing AI has entered a new era. Past progress relied on training innovations (search, deep RL, scaling), but the turning point is that Reinforcement Learning (RL) finally generalizes. A single recipe now tackles software engineering, creative writing, IMO math, and GUI control. In the 'Second Half', the focus shifts from 'solving problems' (training models) to 'defining problems' (evaluation metrics). Evaluation becomes more important than training, requiring a mindset shift closer to a product manager.
More from AGI Musings
- AI regulation will likely follow major disasters — Afinetheorem · 2026-08-28
- Legal Expert: AI Catastrophes Will Trigger Bad Laws — Afinetheorem · 2026-08-28
- Open Source Forced OpenAI to Become an App Layer Company — yacineMTB · 2026-08-28
- Are Agents Getting Smarter or Just Better Tooling? — sophiamia1346 · 2026-08-28
- Compute Will Move to Deserts, Oceans, and Orbit to Evade Local Blocks — Dr_Singularity · 2026-08-28
- Upcoming Interview: Worthy Successor & AGI Governance — danfaggella · 2026-08-28