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
You have to have a reward, you have to know it's doing well or
poorly.
Sometimes that's easy to do.
If you're solving like a math problem or something, you can come up with a data set of math problems and the known solution and check if it's the same.
The on the coding side, there's been a lot of like innovative work around, I mean, there's first of all like a lot of open data and and a lot of, you know, the there's like I think the approach a lot of companies take is you you find existing test cases and then you break them.
But there's sort of like a way to figure out if you know you you can run the test case, right?
And see if if your code fixes
fixes it or not.
In a lot of other domains, it's like much more murky.
It's like what is a good job versus a bad job?
How do I know if I did a good job?
And you really need that information.
So we've tried a bunch of different things.
Ruler is a library that we released.
Um which let me let me relative universal LM illustrat
Thank you.
Yes.
And the way it works is basically this depends on the sort of GRPO insight, which was mentioning earlier, that you actually don't in or with GRPO, it has this nice property where you don't have to have like an absolute judge of the truth.
You just have to judge relatively.
And so simplifying it a lot is basically just LMS judge on a whole group.
Showing 441–460 of 658 · page 23 of 33
← Previous
Next →