MindLecture @MindLecture

Everyone is obsessed with which AI model is smartest. But one of the people building the infrastructure behind modern AI says the real bottleneck is somewhere else. Ion Stoica — UC Berkeley professor and co-founder of Databricks — gave a lecture at Oxford explaining what actually has to happen behind the scenes before an LLM becomes useful at scale. A model is only one layer. You still need systems that can handle thousands of GPUs. Serve millions of requests. Reduce inference costs. Keep latency low. And somehow figure out whether the model is actually better than the one you had yesterday. Stoica has worked on some of the tools behind that stack: Ray for scaling AI workloads. vLLM and SGLang for high-throughput inference. LMArena for comparing models based on real user preferences. And this is the part most people miss: The AI race isn’t only about building smarter models. It’s also about building the infrastructure that makes those models fast, cheap and reliable enough for everyone to actually use. The model gets the headlines. The stack underneath it decides whether it works. Ion Stoica — Oxford University, 2026 55 minutes worth watching. save & watch ↓

ChatGPT might be the least important thing generative AI ever produces. The real breakthrough starts when AI stops generating words — and starts understanding the physical world. Oxford professor Ingmar Posner explains why robotics has struggled with the same problem for decades: You can program a machine for one room. One object. One situation. Change the environment — and everything breaks. Generative AI changes the approach. Instead of programming every possible scenario, machines can learn patterns, build representations of the world, predict what happens next — and eventually decide how to act. Text was the beginning. Images came next. Now the race is moving into robots, autonomous systems and machines that can operate in environments nobody explicitly programmed for them. That’s the part of the AI boom most people are still missing. The future may not be AI that talks better. It may be AI that can see, understand and act. Ingmar Posner — Oxford University 30 minutes worth watching. save & watch ↓