Self-driving cars have advanced further than AI agents in the real world not because digital tasks are technically easier, but because roads, stoplights, and traffic laws already give self-driving cars standardized scaffolding to operate within — scaffolding AI agents dropped into arbitrary digital environments don't have.

24:02remember in the early days of self-driving, a  lot of the researchers around self-driving were saying that the roads themselves will have to  change to accommodate self-driving. There might be sensors everywhere so that the self-driving  cars can interact with it, which I think is like, retrospect overkill. But I actually do think  self-driving cars have a good amount of scaffolding in the world for them to operate in.  Like not completely unlimited. You have roads,

24:24roads exist, they're pretty standardized. You have  stoplights. People generally operate in pretty normal ways. And there are all these traffic  laws that you can learn. Whereas AI agents are just kind of dropped in the middle of nowhere,  and they kind of have to feel around for them. And I actually think going off of what Olivier  just said too, my hunch is some of the enterprise deployments that don't actually work out likely  don't have the scaffolding or infrastructure for

24:50these agents to interact with as well. A lot of  the really successful deployments that we've made, a lot of what our FDEs end up doing with some  of these customers is to create almost like a platform or some type of scaffolding, connectors,  organizing the data so that the models have something that they can interact with in a more  standardized way. And so my sense of self-driving cars actually have had this in some degree with  roads over the course of their deployment. But

25:14I actually think it's still very early in the  AI agents space. And I would not be surprised if a lot of these, a lot of enterprises, a  lot of companies just don't really have the scaffolding ready. So if you drop an AI agent in  there, it kind of doesn't really know what to do, and its impact will be limited. And so I think  once this scaffolding gets built out across some

25:30of these companies, I think the deployment will  also speed up. But again, to our point earlier, I think there's no slowdown. Things are still moving  very fast. That's great. Well, you know, I've thought about autonomy as a three-part structure.  You've got perception. You've got the reasoning, the brain. And then you've got the scaffolding,  the last mile of making things work. Maybe we can dive into the second part, which is the reasoning,  which is the juice that you guys are building with

25:58GPT-5, most recently. Huge endeavor, congrats. The  first time you guys have launched a full system, not a model or a set of models, but a full  system. Talk about that. I mean, the full arc of that development, what was your focus? I mean,  honestly, the benchmarks all seem so saturated. Like clearly it was more than just benchmarks  that you were focused on. And so what is a