Stanford professor Chris Potts discusses his route from linguistics to language-model research with host Sam Charrington. They consider interpretabilityInterpretability is the ability to understand how a model represents information or produces its behavior., architectural efficiency and the role of modular tools such as DSPy.
Potts describes a research approach that compares tokens spentAI inference economics examines the revenue, cost, utilization, pricing, and margins involved in serving trained models. with a basket of engineering outcomesCost per completed task measures the total AI, tool, infrastructure, retry, and repair expense for each verified useful outcome., adjusted for a measure of code survival. He emphasizes that the findings depend on the chosen measuresAn AI evaluation metric is a defined measurement used to quantify a particular aspect of an AI system's performance. and may miss value from explanation, discovery and other work.
They discuss reasoning-token costs, changing product defaults, user expertise and the difficulty of verifying outputs outside familiar domains. Potts presents alternative architectures and data-focused interpretability as research directions rather than established solutions.
Watch on YouTube
