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
I I understand that yeah, I can I can see your side.
Anyway, so so yeah, reward bench is is kind of like and so we've tried a bunch of off that.
Um the thing is, like, I guess my my maybe meta take on this is any task that is extremely common is gonna end up in like as a specific like part of the training data for the frontier labs.
And LMS judge is
Just something everybody's doing in so many different contexts that you have to assume that all the Frontier Labs have a bunch of like LMS Judge style tasks that they're training their models on.
And I do believe that if something does kind of like make it in in a like more than minor way into their training data, that like they're gonna do at least as good a job as as a dedicated model.
So I don't think there's probably a lot of alpha in dedicated LMS judges just because it's something that like the
Yeah.
Let me caveat that and say like if you've got like a very, very specific task that's like weird and has weird requirements and you have a lot of data on what's good or bad, then like training a reward model for your specific task, I think, could still work.
Um, or you know, fine-tuning an LMS judge on your specific task could work.
I'm pretty bearish on like a hey, this is a model that is trained as an LMS judge, but it's a generic LMS judge that can be used to judge anything.
I I just don't think you're gonna beat the frontier labs on that.
Yeah.
Um so we've been playing around with it a little bit.
It's one of the directions that I'm like fairly optimistic on for solving the environment problem specifically.
Because if you think about it like like a world model, it's it's a simulated environment.
That's like what it its whole purpose, right?
So if you get one that's in in an LM like thing.
not like a Docker.
Yes.
Showing 501–520 of 658 · page 26 of 33
← Previous
Next →