Jason Wei: A Year Obsessed With RL Taught Me an On-Policy Lesson About Life
paigeinsf · x · 2026-10-20
Researcher Jason Wei reflects on spending the past year thinking about RL nearly every waking hour, and the unexpected life lesson it produced.
The core analogy comes from the "on-policy" principle in RL: rather than mimicking other agents' successful trajectories (imitation learning), a model should take its own actions and learn from the environment's rewards. Imitation learning is only useful for bootstrapping a nonzero pass rate; once reasonable trajectories exist, training on the model's own trajectories better leverages its strengths, which differ from humans'. He cites the widely accepted example that RL beats supervised finetuning on human-written chains of thought for teaching LLMs to solve math word problems.
He maps the same logic onto life: copying others' paths can get you to a baseline, but the highest returns come from taking your own actions and learning from your own feedback.
More from AGI Musings
- Carissa Véliz: algorithms are often the hitmen companies send to execute their will to power — CarissaVeliz · 2026-10-20
- UK government advisers warn Britain risks becoming a 'high-value AI customer' — Polymarket · 2026-10-20
- Mathematician says AI proved a combinatorics conjecture, now wonders where to share it — tak3sh8 · 2026-10-20
- Altman reportedly outlined three superintelligence endgames to OpenAI staff, including one lab locking out all rivals — kimmonismus · 2026-10-20
- Moshi CEO Anekha Sokhal: AI should do the quant grind, humans keep the edge — AI Engineer · 2026-10-20
- School trains you to fear mistakes, not to survive being ignored — birchlse · 2026-10-20