Hacker News Reader

More RLVR. Give it verifiable problems, if it doesn't find a solution move on, if it does, use that as a reward signal.

Can’t this be extended quite far? Use a cerebras-served model, use verification techniques to generate and solve millions of problems and then use that as training?

That’s the whole point, just cost and compute limitations in your way (mostly).

This isn't latency bound, it is trivially parallelize. So you want to run it on the most efficient compute you have, not the fastest.

OpenAI was bragging the other day about how HLE on Cerebras only took 11 hours. But you're saying if they just spun up a separate cheapo VM for each task, it would have finished faster and cheaper?

Models are trained in highly distributed and parallel ways these days.

It is entirely possible to run multiple concurrent post-training runs. When a frontier lap deploys a 1M RL gym rollout, these 1M environments are absolutely not talking to each other or interconnected. They individually generate traces and movements that can be then combined for post training.

Reply on news.ycombinator.com