Columbia CS Professor: Why LLMs Can’t Discover New Science

In this episode of the a16z Podcast, Vishal Misra and a16z's Martin Casado discuss large language models (LLMs), retrieval augmentation generation (RAG), and formal models that explain the capabilities and limitations of LLMs. Vishal shares his background in networking and how his attempt to fix a cricket stats page led to a breakthrough in AI. They explore the concept of LLMs creating distributions for the next token, reducing the world into Bayesian manifolds, and the implications of information and prediction entropy. The conversation covers the pace of LLM development, the potential plateauing of progress, and the need for new architectures to achieve artificial general intelligence (AGI). They also discuss Vishal's matrix abstraction model, in-context learning, and the possibility of recursive self-improvement in LLMs, and the importance of formal models in understanding AI systems.