LTX made a deliberate strategic bet on an extremely compressive latent space with a variable token rate to compensate for having far less compute than big labs, even though this created serious technical problems other companies didn't have to solve.
Zeve describes LTX's core architectural bet — a highly compressed latent space with variable token allocation (more tokens for hard physics, fewer for easy scenes) — as a necessity-driven strategy to offset their compute disadvantage versus big tech. ✦ AI generated
Zeve Farman · The Cognitive Revolution · 2026-07-08 · original ↗
starts at this moment · 72:31
“How do you make what is your process for assessing research bets that you have to make?”
we place like this like really big bat like compressive latent space a variable token rate. It creates like a whole host of technical problems that people who are having less compressive latent spaces do not have right like it creates like some diffusibility problems etc. But again it was a constraint like we had to do it so we did it.
verbatim transcript · starts at 72:31
72:31cracked a lot of technical problems there. Pink Ltx VA is still the most compressive latent space that people actually use. Nvidia now when they presented their like a world model on top of SA if they use our latent space and that's again like strategic constraints they're coming and say listen I just like can't uh spend the same amount of compute as a big guy so I need to figure out something
72:56algorithmically that's going to give me some kind of edge. So that's like strategic bets, necessity. In terms of like tactical bets, well, you're always kind of trying to figure out uh kind of most reward from the least amount of effort, right? So for example, there are like a ton of cool inference time ideas, right? That do not require, you know, like training or kind of post training. So you obviously
73:23prioritize that. There are like a ton of cool hacks that you can squeeze from that. and uh you are most careful around changing something in pre-training that's like the the most expensive part right so you are before you're like touching pre-training you're trying to show something on a small scale models right so constantly maintaining like models of different size you're trying to show okay it works model of this size
73:49okay works on the model of the bigger size but I guess that also is like a part of the story of necessity we just like can't be as a cavalier with our computers, the big guys. So, we need to, you know, think more before we're placing these bets. >> How close do you think we are to models working on mobile? I I just spoke not long ago to Raine, who you have probably
74:16crossed path with at some point, the CEO of Liquid AI, and I was really struck. I mean, this this is sort of an obvious fact, but it's also like good calibration that the global smartphone and laptop market is $800 billion a year and has been for a number of years. And so, you know, we're only now crossing that level with the global data center buildout, right? So, it's like there's a