Matt Wolfe discusses a near-future shift from large language models improving via self-feedback toward embodied AI — LLMs being integrated into robots and drones (e.g., Boston Dynamics, Figure, Tesla Optimus). He emphasizes the need to train these models on physical interaction and domain-specific knowledge for the vehicles or robots that will embody them.
I think we're going to see that kind of thing get more and more prominent with large language models where it gets to a point where it's giving itself its own feedback and getting better and better and better. But then I also think there's a phase after large language models where we start talking more about the embodied AI. Right? Like we were looking at the Boston Dynamics robot, we've talked about the Figure one robot, we've got Tesla, Optimus, all of these are robots. They're going to have AI injected into them and now they have more stuff they need to learn on. They need to learn on how to interact with the physical world. And when my arm moves like this, what's the result? There's going to be this shift to now we need to train these LLMs to work with a specific use case, like embodied robots going into drones or whatever we use it on next. And they need to sort of train on the domain knowledge to operate the vehicle that's now embodying that AI.
AI News: Dots, GPT-6.1 Sol, Sonnet 5.5, Gemini 4, and everything you need to know
