George Hotz at AMD Advancing AI 2026

George Hotz's latest talk at AMD Advancing AI 2026 is a goldmine for anyone trying to understand where the intersection of hardware and LLM agents is heading. He doesn't just talk about specs; he digs into the actual efficiency of how models interact with silicon, which is critical for anyone currently building a real-world AI workflow.

The core takeaway is the shift toward tighter integration between the model's reasoning and the hardware's execution. If you're using tools like Claude Code or Cursor, you're seeing the surface level of this—the "agentic" behavior. Hotz argues that the next leap isn't just bigger models, but smarter utilization of the compute we already have.

For those of us focused on prompt engineering and deployment, this highlights why latency is still the biggest enemy. When an agent has to "think" through a loop of tool-calling, every millisecond of overhead at the hardware level compounds.

If you're looking for the video, I'd suggest paying close attention to the sections where he discusses the bottlenecks of current tensor processing. It explains exactly why some of our more complex agentic loops feel sluggish despite having "fast" GPUs. It's a great deep dive into the plumbing that makes our high-level AI coding tools actually function.