00:11:29 · Jim Fan
[00:11:29] S1: Absolutely. At that time, in OpenAI, there was the OpenAI universe that I worked on, and there was also a few other exciting projects. One is the DOTA project, so that was called OpenAI five, basically to train auto agents to solve games like DOTA, and there was also actually a robotics project at that time to use high degree of freedom, shadow hands, to solve Rubik's cube. So later, we're going to talk about some research like Eureka that actually reinstantiated some of these ideas and then trained the dexterous hands to do a lot more. So I think at that time, there were a lot of building blocks, even as early as 2016, and the most significant one being reinforcement learning and sequence models. But it wasn't until 2017 and 2018 where kind of the transformer architecture comes, and then it becomes this data sponge where you can really scale on the data and compute, and you will see magic happening from pretraining. So after that moment, we're starting to get into the pre training age of foundation models, and then RO comes back, and as I mentioned, history doesn't repeat, it rhymes. RO comes back and then really takes off from this strong foundation. The foundation needs to be able to do reasoning and needs to be able to do in context learning, all of those capabilities, and then RO can push it to the next frontier. So that experience at OpenAI also inspired one of my first projects at NVIDIA. So that was a pair of projects you mentioned at the beginning, MineDojo and Voyager, where we wanted to have these embodied agents to solve Minecraft. And I I think, you know, for that project, it was also one of my excuses to play games at work. It's all for science. Right? I'm collecting data. Right? I'm not stacking up. I'm collecting data.
00:13:29 · Molly Welch
[00:13:29] S0: All in the spirit of data collection.
00:13:31 · Jim Fan
[00:13:31] S1: It's it's all for AI. It's all for pushing the frontier of AI. So I think what was fun about that project was Minecraft is this open ended environment. It's also a simulation, meaning that you can basically generate infinite data of almost infinite diversity in Minecraft. Minecraft. I've seen people doing crazy things inside this sandbox game, right, like building the Hogwarts Castle brick by brick or even building a functioning CPU circuit using those, like, fundamental building blocks in Minecraft that actually makes the game Turing complete. So this game has infinite diversity, yet humans played it very well, but for AI, it was lagging way behind. It couldn't even do some basic things. So at a time when we started the Voyager project, that coincides with the introduction of g d four. That was the first foundation model that can do coding extremely well. So how we apply g d four to Minecraft is that we give it an API so that g d four can read the instructions of the Minecraft API and then write code, and then we execute that code in the Minecraft simulator to drive the agent to mine blocks, right, to explore, to travel, right, to dig, to craft new items. So, basically, we we we turned the game into a text environment, and it communicates by code. And then we found that g d four is not only good at in context learning, it's also very good at a capability called self reflection, so it can run through a box. No one can write perfect program at the beginning, like in one shot, not even experienced programmers. You've got to iterate, and that's what we observed with gmd4. It writes code, tests it, and the code breaks, it sees bug, it debugs, it tries again, and once it writes a good code, a good piece of code, it would commit to a memory, so we call this kind of an on the fly code repo. So g m d four will save some of the the code it wrote to this repo, so later it can retrieve it if it sees a similar situation. And in this way, we we couldn't actually fine tune the parameters of a GD four, but it still learns over time because it's committing code, retrieving code, and doing trial and error, and it accumulates the context as it goes. So it's a kind of in context lifelong learning.
00:15:59 · Jim Fan
And I find that idea really fascinating because at that time, people were really trying to experiment and see what are the capabilities of these coding frontier models, and Voyager provided one way to build a very complex agentic system on top. And then two years later, these days, the the agents are a lot more powerful. Like, self reflection, people don't even talk about it. It's kind of a given. Right? Like, all the cloud and cursor of the world, they can do debugging super well. They can go on, like, a thirty minute session to to implement a pull request, and then it just... It goes on autopilot, and it it typically does the job pretty well. And it's also able to call all kinds of tools, web browsing, connecting to MCP. So I think the field has gone way beyond Voyager, way beyond Minecraft, so that was that was really exciting to see. Now what was lacking a bit from Voyager was actually that we didn't solve the motor control problem, and we didn't solve the perception problem. We essentially simplified the problem by turning Minecraft into a set of code API and code execution environment, but later, I have come to understand that really the most difficult part may not necessarily be the reasoning side. I think the field is already solving reasoning quite well. The most difficult part of Embodied Agent is actually the low level control, the part where you interact with the low level physics. In Minecraft, it's the game physics, but in the real world, it's the actual physical reality that we're touching, and that makes the robotics problem a lot harder.
00:17:49 · Molly Welch, Jim Fan
[00:17:49] S0: Very cool and incredibly clever, the approach. Remains incredibly clever even as the technology has progressed. So thank you for talking us through it. I... I'm curious. You know, you... MineDojo and and Voyager represent this progression towards agents that learn autonomously or more autonomously through exploration. What do you think are the key remaining bottlenecks for creating truly self improving agents? [00:18:16] S1: I think two things. So there's a popular word called thinking fast and slow from psychology. Basically, it says that all the intelligent beings, like humans, have two systems. So system two is slow and deliberate and does reasoning, and system one is fast, reactive, and intuitive, so the two system approach to agentic AI. I think right now, the system two is progressing quite well. It's not perfect, but I think there are, like, many ways to improve system two, right, like more data, more environment that we can plug it... Plug into the frontier LLMs so they can do reinforcement learning. These days, the frontier LLMs are really bottlenecked by the diversity of the environment, not just the pretraining data. We are kind of in this new era of RL post training, so that's on system two. And system one is... It still remains challenging because computer vision is not soft. It's also much harder to collect data for system one, and I think that's also a good segue to maybe talk about Project GR00T and robotics.
00:18:16 · Jim Fan, Molly Welch
[00:18:16] S1: I think two things. So there's a popular word called thinking fast and slow from psychology. Basically, it says that all the intelligent beings, like humans, have two systems. So system two is slow and deliberate and does reasoning, and system one is fast, reactive, and intuitive, so the two system approach to agentic AI. I think right now, the system two is progressing quite well. It's not perfect, but I think there are, like, many ways to improve system two, right, like more data, more environment that we can plug it... Plug into the frontier LLMs so they can do reinforcement learning. These days, the frontier LLMs are really bottlenecked by the diversity of the environment, not just the pretraining data. We are kind of in this new era of RL post training, so that's on system two. And system one is... It still remains challenging because computer vision is not soft. It's also much harder to collect data for system one, and I think that's also a good segue to maybe talk about Project GR00T and robotics. [00:19:27] S0: It's a perfect segue, which is what I wanna talk about next. So, you're leading physical AI research through the Gear team, Project GR00T for general purpose robotics as you described. What do you think is the grand challenge for robotics? Why is it so hard? Why is system one, which is so easy for us, system two is the hard piece, why is system one so hard in robotics?