Business Data

We’re still validating these assumptions, but I don’t think the limits of Transformer have been fully explored yet. There’s a lot of research going on right now about alternatives to the Transformer architecture, which I’m very interested in. There’s a recent architecture called training-while-testing, and there are a few other alternatives that have some very promising ideas. While these alternative architectures haven’t quite caught up with the performance of state-of-the-art models yet, I’m excited to see new alternatives to Transformer emerge. Stephanie Zhan Are there any models that particularly caught your eye, and why? Jim Fan Yes, I mentioned the “member” jobs and “train-test” models that work more efficiently over different time periods. Unlike the Transformer model, which needs to process all past tokens, these models have more efficient internal mechanisms, so I think they are very promising. However, we need to scale them up to the scale of leading models to really see how they compare directly to the Transformer.



the field of embodied intelligence, what interests you most hungary phone numbers in AI? Jim Fan I’m particularly excited about video generation because I think of video generation as a kind of world simulator. We can learn physics and rendering from data. We’ve seen models like OpenAI’s Sora, and many new models have followed Sora since then, so this is an ongoing topic of research. Sonya Huang What can world simulators bring us? Jim Fan I think it would be amazing to give us a data-driven simulation environment where we can train embodied intelligence. Stephanie Zhan What interests you most in AI in the long term? Ten years or more from now? Jim Fan There are a couple of aspects. First, in terms of inference, I'm very interested in programmable models. I think programming is a very basic reasoning task that also has huge economic value.



Maybe in ten years we'll have human-level programmable agents, which will greatly accelerate the development process, using these large models alone. The other aspect is of course robotics. I think in ten years we'll have humanoid robots that are as reliable and agile as humans, and maybe even better than humans. I hope that by then the Groot project will be successful and we'll be able to have humanoid robots that will help us in our daily lives. I just wish a robot could do my laundry. This is my dream. Sonya Huang What year do you think robots will do our laundry? Jim Fan The sooner the better, I can't wait.Sonia Huang Who inspires you most in the field of AI? You’ve had the opportunity to work with many great AI minds, since your internship. Who is your biggest influence right now? Jim Fan I have too many heroes in the field of AI.