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
Do we know if they used it internally?
Cloud
Okay.
know.
Like I yeah.
I just my experience, you know, knowing a lot of people at these labs is like they launch a lot of products because like some team is super excited about this product, but that I I wouldn't put that much weight on it just because they launched it.
Maybe maybe our baseline was was.
I'll I'll take it.
Yeah.
much more bullish on.
And and we can make the analogy, like we can we can pull in kind of like RL intuition here, which is if you're doing JEPA on a sort of static data set of like, oh, this is the input, this is like what makes a bad good or bad output.
Then like as you're updating your prompt, like your information, the data you're training on becomes less useful, right?
Because it's generated by, you know, because it's based on kind of like the problems you're running into before.
And that's the same problem you have with with RL where
Where where you have this concept of being off policy, where it's like as you're doing training, you really want to be training on rollouts that came from the latest version of your model.
Because if you train on some that came from further back, then it's like it's sort of stale data and it's like not it's no longer representing the current issues with your model.
And so if you try and correct for the issues that existed back then, it it may not actually be helping you that much.
And I think, you know, for either RL or prompt optimization, that's definitely true.
I think that like one way to apply that in practice is exactly what you're saying, where you're using the actual data from your your real emails.
You have some way of saying, like, hey, either people flagging these or Nolan flagging these, or some way of saying like this was a good or bad output.
Showing 381–400 of 658 · page 20 of 33
← Previous
Next →