But the pipeline we've built, including simulation tools and real-world robotic tools, is general enough that it can be adapted to other platforms in the future. So, we're building these tools to be broadly applicable. Sonia Huang You mentioned the word "universal" several times.I think some people in the robotics field believe that a generic approach won’t work and that it has to be domain and environment specific. Why did you choose a generic approach? Our program often mentions Richard Sutton’s “bitter lesson.” Do you think this applies to robotics? Jim Fan Absolutely. I want to start by talking about the success stories we’ve seen in the field of natural language processing (NLP). Before the advent of ChaggbT and GPT, there were many models and pipelines in the field of NLP dedicated to different applications, such as translation, programming, mathematical operations, and creative writing, all of which used different training models and pipelines.
all of these applications into one model. We call vietnam phone numbers these models “general.” When we have generalist models, we can use queries, refinements, etc. to specialize them for specific tasks and form a “specialist” model. Based on historical trends, the specialist-generalist model is almost always much more powerful than the original specialist model and is easier to maintain because there is only one API, which inputs and outputs text. Therefore, I think we can go the way of success in NLP, and this will be true for robotics. Until . most robotic applications will still be in the specialist phase, that is, specialized models for specific tasks, specific hardware and specific data pipelines. But the goal of the Groot project is to build a common core model, first for humanoid robots, and then generalized to different robotic or embodied forms.
This is the general moment that we are striving for. Once we have a generalist model, we can adapt it to specific robotic tasks. And these will become manifestations of the specialist-generalist model. But until we have a generalist model, that will not happen. So in the short term, it's easier to look for a specialist model because you only have to focus on a very narrow area of tasks. But we at Nvidia believe that the future belongs to the generalist model, and although it takes more time to develop and there are more research problems to solve, that's the goal we're aiming for first. Stephanie Zhan Another interesting thing about Nvidia's build of Project Groot is that, as you mentioned earlier, Nvidia owns both the chip and the model. What do you think Nvidia can do to optimize the Groot model on its own chips? Jim Fan Yes, at the GTC conference in March of this year, Jensen also announced a next-generation chip for edge computing called the Jensen Source chip, which was actually announced alongside Project Groot.