00:00Let's start with the idea of a world model, distinct from an LLM.
00:03I appreciate you and I have discussed this already,
00:05but for our audience, that distinction is important, the basic definition.
00:09Yes, I'm very excited by world models,
00:12because if you look at language models, what they output are words.
00:16Words can be turned into codes, words can be turned into conversations.
00:21But world models, the ones that World Labs has been developing,
00:26what we generate are not words.
00:28The models generate beautiful pixels, very accurate geometric structure of the 3D or 4D world,
00:36and which in turn can allow people, creators, designers, robotics developers,
00:42to have precise control of the cameras or the construction of worlds.
00:47Your latest generation of world model that you've been working on at World Labs is Atlas.
00:52Yes.
00:52You haven't said very much about Atlas.
00:54We just shared a technical blog about Atlas.
00:58And we are working really hard to release Atlas as a product to the world.
01:04And it is the latest, the greatest world model that World Labs has developed.
01:09It's the first of its kind in the ability to allow camera control,
01:15allow that kind of precise pixel consistency with the real world, with the 3D world.
01:22It also tries to solve a very hard problem in the field of computer vision that combines the generation of
01:29beautiful pixels
01:30with the reconstruction of intricate 3D structure.
01:35Does it solve for it?
01:36It does.
01:37How good is the underlying model?
01:38It's better than any state-of-the-art specialized model for these tasks.
01:44There is the underlying model Atlas, and you mentioned product.
01:46What is the killer use case application for deployment in the real world?
01:52As many foundation models, these use cases are pretty horizontal.
01:57We are already engaging with users from creative industry, with design industry, with robotics industry,
02:05such as healthcare lab development, pharmaceutical laboratories.
02:12So it is pretty horizontal.
02:14But one thing that really excites me is by allowing this kind of control for humans,
02:22creators can use this as an augmentative tool to really express their creativity
02:29or develop downstream technology for their use cases.
02:34From a commercial perspective, World Labs has been focused on robotics, right?
02:37You acquired Scenics.
02:41Explain the use of a world model in the development of physical AI,
02:45the deployment of robots in real human-centric settings.
02:50Yes.
02:50First of all, World Labs is a spatial intelligence company.
02:53A world model is very much useful for spatial intelligence applications
02:58from design creativity to robotics, and also which includes healthcare education.
03:04What is important for robotics is the ability to understand and construct the world
03:12that you can in turn train and evaluate your robotics tasks or policies.
03:19There is a lot of still bottleneck in terms of creating the right training environment or evaluation environment.
03:27A robust world model can help unlock some of this bottleneck.
03:33And, for example, for laboratory sciences, when you want to empower a robot to do pharmaceutical manufacturing
03:43or lab science, you don't get that much data, right?
03:46These are very highly specialized industries.
03:50But using world models to simulate and then turn that simulation into very realistic training ground
03:57as well as evaluation ground helps to develop robots.
04:01We are talking about the physical world in many cases.
04:05And the issue that your industry is confronting right now is alignment.
04:10In LLMs in particular, does the model behave as humans intended it to?
04:14Is that the same calculus for a world model?
04:18Is there a risk that the world model misbehaves in the same way that we have seen LLMs misbehave?
04:25I think it's the same calculus for all technology.
04:28I've always been saying technology is a double-edged sword.
04:32We develop it with the intention to be benevolent and helpful to humans.
04:37But we have always to put safety and responsibility in the forefront of the development and deployment of technology.
04:46So from that point of view, whether we're talking about LLMs, we're talking about world models,
04:52or we're talking about classic technology like electricity, steam engines, modern transportation,
04:58all of this, we do have to hold both truths in the hand.
05:03One is human safety, human benevolent.
05:07The other one is use this for incredible, helpful, incredible advances of the technology
05:15to help humans and develop our civilization.
05:19Your industry colleagues and peers at the frontier say we should pace development at the frontier,
05:25the most capable models, in part because safety and alignment has not kept pace
05:30with the improvement in capability.
05:32How have you seen that through your work in world models?
05:36And, you know, the argument that, for example, a Dario Amode would make is that would only work
05:41if everyone agrees to do the same thing.
05:44It's in part why they ask for antitrust waivers, right?
05:46So that they can work collaboratively with the other labs.
05:51So look, AI is a civilizational technology, and with this level of impact it can make for human society,
06:00I believe responsibility is so important.
06:04We cannot leave it to fatalism.
06:06One thing that is always important for me is that not only I'm a CEO of World Labs and an
06:12entrepreneur,
06:13my entire career, I've been a scientist and an educator.
06:17And I want to apply the same kind of sense of responsibility of a scientist and educator
06:23to the development of world models or any technology we develop.
06:27So it's easy to use simple words, but really this is very nuanced.
06:34We need to understand and keep this technology, continue to develop this technology in a robust ecosystem
06:43where there is public sector, there is open source, there is closed source,
06:48there is collaboration with academia, with government, with civil society.
06:53It is very, very important that we come together as an ecosystem and do this right.
07:00Dr. Lee, what is your true assessment of the existential threat to human life from AI?
07:07Your colleague Andrew Ng was on the show last week, and he said,
07:10this is ridiculous, it's near zero in his mind,
07:13because he reflected back when GPT-2 came out and said,
07:16we had those concerns then, look at the development path since.
07:21I think as an educator and a scientist,
07:25that any threat to human society, including existential, is within ourselves.
07:32It's that how we humans wield the power of technology,
07:37how we humans collectively work together and govern it,
07:42and how we humans use these powerful tools to make our lives better.
07:47Are we humans in control of that?
07:50If we, we must be in control.
07:53I've been advocating for years that AI, it's not about AI, it's about humans.
07:59It's about our shared, as well as individual dignity, agency, as well as our responsibility.
08:07And the goal of our civilization is to create a better society for our future generations.
08:14And whether the tool is electricity or AI, we must use it right.
08:19Can we bring this back to world models to end?
08:22I think I'm right in saying you believe that there isn't right now a sufficient benchmarking for world models
08:29to assess lots of the things that we're talking about.
08:32Who should be responsible for sort of setting that benchmark?
08:36That's a great question.
08:37Obviously, as scientists and engineers within World Labs,
08:41we do a lot of internal technical benchmark.
08:44Our benchmark cares about safety.
08:46Our benchmark cares about the model capability.
08:49But the bigger picture is that this problem, this question of evaluation and benchmarking
08:57should be an ecosystem level responsibility.
09:01Is there a format for that to your mind?
09:03I've been for years advocating for public sector participation in this technology.
09:08I understand industry.
09:09Through a regulatory body or just through a...
09:11Through all of the above.
09:13Look, I co-founded Stanford Human Center AI Institute.
09:17Before ChatGPT was released, we formed a center called Language Model Research Center
09:24to put out those benchmarks.
09:27You know, you would say ImageNet 15 years ago was one of the very first benchmark of AI.
09:34I continue to believe importance of benchmark from independent bodies and public sector bodies
09:40like academia as well as the shared responsibility of industry of government together.
Comments