What actually is "scaling post-training"?
replyMore RLVR.
Give it verifiable problems, if it doesn't find a solution move on, if it does, use that as a reward signal.
replyCan’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?
replyThis isn't latency bound, it is trivially parallelize. So you want to run it on the most efficient compute you have, not the fastest.
replyOpenAI 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?
replyThat’s the whole point, just cost and compute limitations in your way (mostly).
replydoes this suggest 5.3 is the same # of parameters as 5.2?
replywhich is the bigger headline that people don't realize. this is 744b and its head to head with Kimi K3 (2.8T), smashes DS v4 pro (1.5T). even Opus and Sol are rumored to be 1.5T+ this is half the size!
replyin terms of performance the formula seems to be : dense parameters = sqrt(total*active)
replyYou do need to compare active parameter too though. The total size isn’t a reliable indicator anymore
replyIt's the same pre-training, they are just adding more (+ better) SFT, RL, etc. (post-training). Model internal knowledge cut-off is still the same.
replyIt seems we are doing pre-training every 6 months, and post-training every 4-8 weeks now.
“Today we are releasing GLM-5.3. It uses the same base model as GLM-5.2 — every gain comes from post-training.“
replyI was asking if this implies that 5.3 has the same number of parameters as 5.2. I can, in fact, read. What I didn't do is understand the implication of that statement. Thank you for your copy/paste service.
reply