00:00We've talked quite a bit on this program about the challenge.
00:03The challenge is that in the physical world, models need to be multimodal.
00:08But I think let's start by defining the problem, generalization.
00:12That might be a newer term.
00:14What are you getting at there?
00:15Thanks for having me, first of all.
00:18Currently, pretty much all robots are trained by showing a few examples of specific tasks.
00:23So the entire field basically trains one task at a time.
00:27If you compare that to the world of language models where you train a large model and then it's able
00:32to handle a wide variety of tasks, unseen problems as well.
00:36So that is an example of generalization.
00:38And in robotics, we have this critical gap where we are just stuck in the valley of specific tasks.
00:44In order for robotics to be impactful in the world and for us to be able to just talk to
00:49them and ask them, hey, okay, do this.
00:51Then when you're done with that, go take care of that thing.
00:55And it may be a new scenario being presented to them for the first time.
00:58For the first time.
00:59So generalization is this problem of how do we allow robots to solve generally any task.
01:05Even in a world where you have access to a lot of synthetic or virtual data, having a grounding in
01:11physics, the real physics of the world, is a principal challenge.
01:14So why is it that having an open science physical AI lab is a solution?
01:18It sounds somewhat abstract with respect.
01:20Right.
01:21So, I mean, there's two aspects to this.
01:23One, that it is open.
01:24And the second is the work that we are specifically doing to solve.
01:27Open being open source, you mean?
01:28Open being open science, open source, both of those aspects.
01:32So, first of all, what the solution actually looks like from our perspective.
01:37Currently, the best solution is brute forcing this approach, which is gather data on every single task you can imagine
01:44in combination of tasks.
01:45Everything humans do, from picking up cups to, you know, digging, like, you know, mines, all of these things, and
01:50gather data one piece at a time.
01:53That's a practically impossible solution.
01:55On the other hand, the teams at Luma, the work we have been doing for the past four years is
02:00in building out these general systems out of multimodal data, internet-scale multimodal data, and extracting signals from that that
02:07allows control, that allows simulation of reality, and that allows physical control.
02:11So this lab's job is to leverage that skill into physical AI.
02:17And the second aspect of this is open, and I think we would not be doing it any other way,
02:21and I think this is really, really significant.
02:23So, if you think about what physical AI would mean for the world, it will be everywhere.
02:30It will be in our houses.
02:31There will be these systems, robots, will be manufacturing everything we depend upon, everything we eat.
02:36They will be in our hospitals.
02:37They will be in scientific labs.
02:39They will be on our streets.
02:40Right.
02:41So, it's completely untenable that one or two people control this entire stack.
02:48So, we want to live in a world, and we want to affect a world where a small group of
02:53people can, like, you know, take these technologies and build them into productive systems, and that's why it is an
02:58open initiative.
02:59I mean, this seems like more of a philosophical push of yours, the idea that we shouldn't have so much
03:05control.
03:06I mean, it's almost akin to the encyclical by the Pope, but Meta at one point was pushing open source,
03:12and then it moved away from that, just the sheer amount of money that needs to be made.
03:17How are you going to fund this?
03:18How do you commit to open science when everyone's so worried about China and geopolitical tensions?
03:24Right.
03:24I think, first of all, it has to be philosophical because it is not just a tool.
03:29This is going to be technology.
03:30I mean, AI already is immensely impactful in our world beyond what anybody imagined even two years ago.
03:37Physical AI will be deployed even faster because of the economic impact it's going to have.
03:42So, a physical, philosophical stance is absolutely necessary.
03:46But you're absolutely right, funding these systems.
03:48So, what we believe, actually, is that this level of control over means of production is actually not a tenable
03:56economic situation anyway.
03:58You know, nations would not be happy with one or two companies outside their borders controlling their means of production.
04:06So, we believe, actually, this is not just a philosophical stance.
04:08This is an economically sound stance.
04:10And building an ecosystem where, like, you know, chip partners, the model brain providers like Luma,
04:16and deployment partners work together to build these out into systems of productive work is the right economic path.
04:23And currently, intelligence, especially LLMs, are going on the wrong path here.
04:28I mean, we've got 30 seconds.
04:29But you've raised 900 million, Series C.
04:32We're just seeing the output that Luma creates.
04:34But why are you the right person when you've got world labs, for example, with Fei-Fei Li?
04:38Yeah.
04:39I think the – just like LLMs were not – or language models were not solved by linguists,
04:45we believe to solve physical AI, you need the systems of large-scale multimodal data infrastructure.
04:51This is what Luma does.
04:52This is our bread and butter.
04:53And we have produced some of the best models in this space on 3D, on images, on video, taking raw
05:00Internet data.
05:01This skill is, I think, what is essential to solving this problem.
05:04And that's why we think we are one of the best suited companies in the world to solve it.
Comments