Kyle Corbitt
speaker
658 appearances
1 recordings
1 series
first heard Oct 2025
last heard 16 Oct
Kyle Corbitt’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 1 in all, peaking in Oct 2025 with 1.
Appearances
So that that's nice.
Another thing is the benefit that you're talking about, which we've observed.
So the way GRPO works is you have to do like, you know, a set of different uh trajectories or set of different rollouts, um, all in parallel with the exact same environment, the exact same conditions, and then you score each of them.
And GRPO uses the differences in those scores to promote the trajectories that did better and sort of like decrease the probability of the ones that did worse.
Because they do it in sort of a group relative way, the only it lets you be a little bit looser with how you score them potentially.
Like you don't have to necessarily have a globally aware scoring function.
You just need some scoring function that is able to distinguish between this small set of things you have in front of you.
And then that's easier.
That's easier for a human, you know.
If you if you tell a human which of these to choose which of these is better, it's easier for them to do than say, like, is this one good or bad in in absolute terms?
Yeah.
So that's nice.
The big downside, the huge downside of GRPO, and I think actually the reason why GRPO actually is likely to be a dead end, and we probably will not be continue using it indefinitely, the fact that you need to have these parallel rollouts in order to train on it is actually the like that makes the data generation much more complicated.
Because you need a fully reproducible environment to be able to do these sort of parallel.
parallel rollouts.
And it turns out in practice, that's like
Getting that set up is the hardest challenge today with getting RL working is like actually designing this robust, uh reusable, you know, environment that you can run all of this training in.
Most companies and and that's not true.
Like sometimes that's easy to do.
Like like there's certain situations where you where you can do that.
Showing 221–240 of 658 · page 12 of 33
← Previous
Next →