Let's bet on this shit (how to act rationally if you believe in RL scaling and envs)
Summary
Charlie O'Neill argues that believers in RL scaling must act rationally by building long, real, and domain-specific environments rather than chasing quick riches through synthetic task slurry.
Belief requires rational betting.
O'Neill notes that scaling-pilled or AGI-pilled individuals often fail to invest in compute, data, or land, mirroring the current inconsistency in RL environment conviction.
I see very few people acting rationally on this belief to its logical conclusion.
RSI makes environments exponential.
If Recursive Self-Improvement is imminent, the scarcity of high-utility environments and the rising human cost of production drive their value up exponentially.
the value of one that actually works should increase exponentially too.
Horizon length is the differentiator.
Since RL capabilities do not generalize well between domains, builders should focus on increasing the horizon length of tasks through irreducible complexity.
Real data beats synthetic difficulty.
Chinese labs win by distilling from real user distributions via router farms, whereas frontier labs rely on expensive, synthetic, frontier environments.
The cheap side is winning.
Build environments like cathedrals.
Instead of creating task slurry for quick profit, builders should focus on the 'Long, Real, Yours' framework to capture domain expertise lost in pretraining.
If you believed everything was going to be RL environments, you would build them the way you'd build a cathedral.
Highlights
Let's bet on this shit (how to act rationally if you believe in RL scaling and envs)
So one side is doing cheap distillation on the real distribution, and the other is doing expensive distillation on synthetic, hard, frontier environments.
Read the original ↗Save to your Sponge
Summarized with Sponge. Quotes are the author’s own words.