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
Yeah, yeah.
So it's it's like, you know, whatever, hallucinating, generating, imagining the responses you'll get from the world.
So you can imagine, right, if you had like a really, really great world model that you were training on, yeah, it's like your your agent that you're using, it would go out and make some tool call, and then this world model would generate, hey, this is like probably what the tool called and if if you have a smart enough, strong enough one, then it could keep its own effective
internal state of like the changes you made so far and how that affects.
So we've played around with it some.
You know, I think if we can get it to work really well, then that could be a solution for the environment problem where you just
Take a bunch of production traces and use those to condition your world model so it understands your specific system and what its failure modes are, and then train against that world model.
Um and uh and worse and you know, and and the resultant, the you know, agent that you train with at would would then be able to perform in your real environment.
So I I do think it's like a like a really interesting area of research.
Yeah, I think the way they envision that model being used is a little different.
Like I think they're they're they're try it's actually I'm curious, I'll have to see the talk.
Um but my understanding from that paper is like the goal they're imagining is this is almost sort of like a pre-training step, and then now that this model understands code really, really well, we can then use it as basically like a code generation um or a coding agent of some kind.
Okay.
Yeah.
Which which I think makes sense.
That's almost more like a different kind of
Pre-training, I would say.
Um, the way I'm interested in applying world models is as not it's basically as its own end, right?
Where it's like actually the goal is to come out of this with something that simulates the world, which is not something you really need in code at all because it's so easy to like run code and you don't need to model what will happen if you execute this code typically, because you can just execute the code and and see what happens for for training purposes.
Yeah, so that was driven by actually mostly the weights and biases founding team.
Showing 521–540 of 658 · page 27 of 33
← Previous
Next →