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
And it's it's all about this kind of like reliability issue where it's like, okay, like it it works in controlled circumstances.
There's other areas where it doesn't work.
And so if we can solve this problem, there's that that like 90% of the like inference market, like addressable market today that's just gonna like come on.
online because we've solved that problem.
So um that's what we want to do.
Um I'm super excited about it.
And like I think we have very concrete ideas on like the specific pieces we need to make that work.
And we just have to execute against them.
I'm not that worried about it.
Uh and the reason why is because it is reward hacking is quite easy to detect once it starts happening, because once the model's found some hack, it just starts like doing it all the time.
It's like, oh yes, this worked great.
I'm just gonna keep doing it.
And so you you you like notice very quickly, whoa, it's doing this thing.
And assuming you're using at least in part an LM as judge to like determine which ones are good and bad, it's so easy to just throw in an extra term and be like, hey, that like
weird thing that you keep doing, like if it if it does that, like that's bad.
Give it a low reward.
So we've we've done this with a bunch of customers and like reward hacking does happen, but like you'd you just see it and you'd like adjust your, you know, reward prompt and it just goes away.
Oh, that's a good question.
One thing that I've that I I really identify with and I've tried to do a good job is kind of like, you know, sort of uh I I I think they say like hold your um problem tight and your solution loosely, right?
Where it's like
Showing 621–640 of 658 · page 32 of 33
← Previous
Next →