00:18:18 · Finn Meeks, Amit Jain
[00:18:18] S1: With this unified model, do you do you think that this eventually replaces language models as well? Yeah, I'm curious, like, is LMs just going to be a thing of the past a few years from now and everything will be structured more as unified models? [00:18:34] S0: It depends on what you mean by that. If you mean that empirically, I think no. If you mean that theoretically, I think yes. So empirically, the considerations are different. When I say empirically, I mean in the field, in deployment, in real usage. In real world usage, you want to strike for efficiency a lot more. So I think LM scaling is now at a point where you just don't see like, oh, okay. So for instance, the recent Qwen 27B model actually is outperforming a Qwen 289B model basically, which is ridiculous at this point. I think we will still continue to use language models quite a bit, but theoretically, a world model which has understanding of language and physics and the world model will strictly be better. It will be more general, it will be able to use in more scenarios, But does that mean you should be using that role model for writing front end code? Maybe not. But if you're making a front end that is really involved and has a lot of visual states and a lot of interactions on it, then yes, you do want to have the world model that can understand the full layout and flow of it and can like, you know, make makes and watch the video of the thing play out and say like, no, no, this is not very good. Like, you know, let me edit on it. So use case.
00:18:34 · Amit Jain, Finn Meeks
[00:18:34] S0: It depends on what you mean by that. If you mean that empirically, I think no. If you mean that theoretically, I think yes. So empirically, the considerations are different. When I say empirically, I mean in the field, in deployment, in real usage. In real world usage, you want to strike for efficiency a lot more. So I think LM scaling is now at a point where you just don't see like, oh, okay. So for instance, the recent Qwen 27B model actually is outperforming a Qwen 289B model basically, which is ridiculous at this point. I think we will still continue to use language models quite a bit, but theoretically, a world model which has understanding of language and physics and the world model will strictly be better. It will be more general, it will be able to use in more scenarios, But does that mean you should be using that role model for writing front end code? Maybe not. But if you're making a front end that is really involved and has a lot of visual states and a lot of interactions on it, then yes, you do want to have the world model that can understand the full layout and flow of it and can like, you know, make makes and watch the video of the thing play out and say like, no, no, this is not very good. Like, you know, let me edit on it. So use case. [00:19:47] S1: So you might see kind of just systems of models, with a unified model guiding guiding them tailored to a specific application.
00:19:47 · Finn Meeks, Amit Jain
[00:19:47] S1: So you might see kind of just systems of models, with a unified model guiding guiding them tailored to a specific application. [00:19:54] S0: Correct. The unified model will replace LLMs as the core, but you might do a tool call to an LLM and ask it like, you know, hey, can you just code this out because you're more efficient than me? [00:20:03] S1: Tell us a little bit more about Luma Agents. This is a recent release of yours, and I I think it kinda goes back to to what you were saying earlier is that there's a lot of value too in just the the orchestration and the ability to actually, like, solve, you know, that end to end problem for our customer beyond just having a good model. And there's more than just a model that goes into it. Can you tell us a little bit more about Luma Agents and just how you actually think about making these useful for customers?
00:20:27 · Amit Jain
[00:20:27] S0: Right. So Luma now has a very fast growing and sprawling business in entertainment, media, advertising. Basically the use cases for visual models when the accuracy is good enough, where humans think it looks real, but not good enough yet where it's metrically accurate. So for these kind of works, it's actually really useful. Just to give you a sense of that market, by the way, on average, companies spend about 10% of their revenue, nonprofits, revenue on telling their own stories. Every company, if you're a consumer, you're going spend about 15% to 20% of your revenue on telling the world about it. If you are a services company, you're going be somewhere in the 10% to 12% range. And if you are a government or defense or those kind of companies, you're going to spend less, not zero, less. How many of you know about Palantir, right? Like how many of know about Northrop Grumman? You have no business knowing about Northrop Grumman because none of us are going buying jets like on Sundays at IKEA. But we know about Northrop Grumman because well, at the end of the day, the dollars come from our funds and then from our like tax, right? So every company needs to be known in the world. So that's a very big business. So if you have a stock market that is trading at 30,000,000,000,000, right? About 3,000,000,000,000 of that value is actually going towards telling that story, media advertising. But the thing is like, and this doesn't apply just to our business, it applies to everything. Promise of AI is not doing a little bit of spot work, all right? Hey, I can make copy. It's like, all right, good for you, man. Can you actually make the book for me fully? Can you actually just go talk to the customer? Hey, I can make a little bit of image for you or video for you. It's like, who cares? Right? Can you actually do the full end to end movie or can you do a campaign? In 2024, 2025, we struggled with this a lot because agents weren't really there and the lens weren't really good enough for that. Our models were also really not good enough to do long horizon tasks. But we just actually announced a full production on Prime Video with Sir Ben Kingsley. It's a TV show where the production budget is $4,000,000 an episode and it has Sir Ben Kingsley, it's called Moses, right? And now we are at a place where it is about like 90% produced with Luma Agents.
00:22:36 · Amit Jain
And it was never possible to do before that. So what is that? So it's a product that allows you to do large scale end to end production in visual world. So if you are a company that is selling a product online, it can actually help you produce very large scale assets in every language, for every geography, for every location, and it can interpret a huge amount of things. We use LLMs as well and those kind of things. But the big gap was in order to make agents, you need two things. The generation part, the thing that does the task and the understanding part. Coding agents didn't get good enough until models got good enough to understand the code they were writing at reading large volumes of code and running it. Until you do that, you don't know you made a mistake and you can't fix it. And then the person who is at the other end is now left to hold the bag and understand like what you did and where you made a mistake. So we had to build really elaborate systems to be able to understand the work that these models are doing. Right? So understanding and then generation. So Luma Agents now is like the flagship product for Luma. And it's also actually really, really interesting for other reasons. So first of all, it's really interesting because it's really valuable to our customers and to a large population in the world. There's 120,000,000, by the way, you just... If you don't know, visual artists in the world, people who are doing creative work. That is about two times to three times the number of coders in the world, massive part of the economy. Second, it's really valuable because now we have a end to end optimizable loop. Every interaction and task that someone is doing in Luma Agents is then used to train the agents, is then used to train the models. This is the only way. So we can bootstrap on the internet, but that's the best we can do, right? We can't actually do the final end to end problem because nothing on the internet has the data for how you got to that artifact. You can see the movie, but you don't know how the movie was made. And here we come into the same problems as robotics, same problems as biology, same problems as literally everything in coding and whatever have you, we're like, how did you get to it? And if you want to have agents that can do end to end work, they need to know how you get to What actions do you do? How do you iterate on it?
00:24:46 · Amit Jain, Finn Meeks
All these kind of problems. And now there's this really incredible flywheel where we get to see some of the best creatives in the world using them and making like really, really great things, and then that improves our models, That improves the agents further. Now the people take it even further, that gives us even more intelligent data. And then that's how you get to a system that is like Claude Code, where just every iteration is able to do many, many, many, many more things. And the goal of Luma Agents is end to end multimodal or creative work. [00:25:17] S1: I think maybe a week or two ago we hosted a demo night at SPC and one of your forward deployed creatives came by and gave a demo here. I'm curious, tell us a little bit more about that program and why you spun it up and maybe how that also plays back into a similar loop of creating real end customer value and making sure these models are actually solving their problems.