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 think that like when you you know, you you see the PPO equation or something with all the symbols, like if if that's your first intro to it, then it feels very complicated.
But I think like
If you were to show that exact same equation, just like code, not maybe not PyTorch code, because that you also have to like understand.
But if you just like did the naive implementation in like Python and like showed someone like, hey, this is this is kind of like how we're we're computing the loss here, who was like a strong engineer, like I think it's actually like quite grokkable.
So yeah, I mean like I d I don't think it's like the buried entry is that high.
I think you just have to like believe you can do it and then like spend
Some time staring at it.
That would be what I would recommend.
It's like uh you know, you can read the papers and look at the equation.
I I think actually this is one area where where OMs have been super helpful.
If I'm reading a new paper and I look at one of those equations and I'm like, I don't understand how this new term they introduced intro like corresponds to like the these other terms, then I can like dump like all the context around it into GPD5 and say, like, hey, can you like write this out of Python for me and show me what what they're doing?
differently and that's super helpful for kind of like my background, I guess.
know like Jeremy Howard's been beating this drum for for for years and I mostly agree with him.
Yeah, I mean it's interesting.
There's real pros and cons.
Um, if you're moving from PPO or or something similar to it to GRPO.
There are some big pros.
I mean, one pro is just sort of like operational simplicity, like there's a whole extra model you need for this value model you need for PPO that you can throw away with GRPO.
And that just like makes your life easier.
You don't have to train that model, but also like there's like no hyperparameters around that model that you have to f configure.
Showing 201–220 of 658 · page 11 of 33
← Previous
Next →