Descriptions:
Machine Learning Street Talk host Tim Scarfe sits down with Thomas Ahle — a researcher whose background spans locality-sensitive hashing, probabilistic machine learning, and chip design — to explore two provocative ideas reshaping how AI hardware is built and verified.
The first thread covers Normal Computing’s thermodynamic chip, the CN-101, which inverts the usual engineering goal of eliminating noise. Instead of fighting randomness, the chip harnesses it: its physical behavior naturally solves stochastic differential equations, making it natively suited to probabilistic workloads that would otherwise demand enormous conventional compute. The CN-101 is already at silicon, though Ahle is candid that it targets a narrow band of workloads and the real test will come at scale.
The second thread may be even more immediately consequential. Ahle describes building a production-grade Verilog simulator almost entirely through a swarm of collaborating AI agents — generating over half a million lines of code in 43 days. The conversation digs into why that matters: commercial EDA tooling costs tens of thousands of dollars per CPU core, and the hardware industry has none of the open-source ecosystem that software engineers take for granted. Ahle and Scarfe then wrestle with the deeper epistemological problem this creates: if AI writes a chip design or a formal proof, how do you actually know it is correct? The discussion of autoformalization, functional coverage, and orthogonal verification teams makes this one of the more technically rigorous AI-hardware conversations available in podcast form.
📺 Source: Machine Learning Street Talk · Published June 28, 2026
🏷️ Format: Interview
