Short answer: Jordan Metzner compares model training to building hotels: if a model takes three to nine months to train, labs look quiet while training at maximum, and they must decide when to start the next model.
Watch this answer from Episode 40 (starts at 16:37) · Episode 40 notes and transcript
From Episode 40: Claude Design Changes Everything (Figma in Trouble?), starting at 16:37.
Sam Nadler [16:36]: Yeah. I feel like about six months ago, it really felt like Google was making huge strides. They were really catching up. But it's been, it's I could be wrong here, but it's felt like they've been a little quiet over the last quarter or so, where Anthropic is obviously, like seems like it's really, like, taking the lead. And then OpenAI is, still contending. Do you expect anything do you know of anything on the horizon for from Google?
Jordan Metzner [17:03]: No. I mean, I've tried their new real time chat. It's pretty good. Some of the new Gemini tools, it definitely seems like they're a little bit as far as, like, competing in the code race against, Anthropic and OpenAI. And, some news on Twitter has come out about how, some teams at Google have adopted AI quite well. Some teams have not. It seems like the teams that have been using Claude Code. So, maybe even at Google, they're not using, Gemini as its tool set. But, I think what you have to take a step back on. It's kinda like building you know, every time these models launch, it's almost like building a hotel. You have a city. You don't have enough, hotels for the tourists to come to your city. So, like, great. Let's get a hotel. But from the idea that, like, hey. We don't have enough hotel rooms to, like, the hotel opens, takes, two or three years. And by that time, you hope the demand is then increased. And then by that time, you need more hotels. And so you're kind of, like, always on this thing where you're behind on the amount of, like, availability that you have based on the demand when you're in this, like, growing market. Now if you apply that type of model here to, like, how long it takes to do the training set of some of these high end models.
Jordan Metzner [18:04]: I don't know how long 4.7 took to train or 5.4 from OpenAI. But, if it takes three months, six months, nine months to train some of these models, then it seems obvious that you're gonna go through these quiet periods where it looks like some of these major laboratories are doing absolutely nothing. But in the background, they're at maximum training. Now, the problem you have is that the longer this iteration is, the more risky it is to throw, approximate iterations in so that they then get built later. So, you start the building of the first hotel today. Do you start the next hotel next month, or do you wait until your hotel is built? Or do you wait halfway through knowing that you don't know whether your hotel is gonna get built on time or not? Right? So you start to get into this space where it's like, Okay, well, how much money do we have if we already have a model that's going to be done in, let's just say, ninety days from now, do we start another model today with some of the things that we didn't get to put in the last model, or do we wait in this one's fill? So I think you have this cycle of understanding, kinda when do we start our next model because we know it's gonna take, like, nth length of time, and then what significant improvements can we make because, like, it does seem that, like, once that model train starts, it's not easy to stop it, start over again, or, like, implement new datasets.
