What's the value prop of RL environment startups? A Reddit founder's three questions
arnavneedshelp · reddit · 2026-10-27
The poster keeps getting pitch emails from companies selling RL environments to frontier labs like Anthropic and OpenAI, and questions the model: you must prove 1) their models currently fail the task and 2) it's solvable via RLVR.
Core puzzle: open-source models lag the frontier by only 3 months (implying the frontier is mostly new RL environments, or distillation is very good) — so why don't these well-funded startups just raise money with scaling laws plus their environment and train their own model?
He offers three possible answers: 1) his assumption that pretraining/infra is now much easier is wrong; 2) aggregating all past RL environments is still hard and only certain labs have them; 3) these companies will all eventually train their own models.
More from Companies & People
- Manus on rebuilding itself as an AI-native org: expose interfaces, demand evidence, go async — dotey · 2026-10-27
- Dream Labs is hiring robot learning researchers in SF, plus security and ops roles — scott_e_reed · 2026-10-27
- Beff Jezos: Elon always frontloads mass production and wins the ramp asymptotically — beffjezos · 2026-10-27
- rakyll: open source work creates jobs and a mini economy around projects — rakyll · 2026-10-27
- Palantir's 12-week American Tech Fellowship opens applications for industry workers — eliano · 2026-10-27
- 'Software as content': a new creator class whose medium is software, with 9.3M-view proof — yungcontent · 2026-10-27