Simon Willison Highlights Ted Sanders' Insights on LLM Performance and Size

Simon Willison shared insights from Ted Sanders, pointing out that although LLM performance is multidimensional, model size largely dictates capabilities, token costs, and factual recall. Sanders also explained that modern version numbers reflect release order, relationship to other releases, and the magnitude of improvement. Understanding how model size correlates directly with performance and cost helps developers and organizations make more informed decisions when choosing AI models for production. It demystifies how major AI labs structure their model hierarchies and versioning systems. Sanders noted that while a model good at math is not guaranteed to excel at jokes, larger models consistently deliver higher reliability, better recall of esoteric facts, and higher costs per token. He used a hypothetical scaling hierarchy (Luna < Terra < Sol < Astra) to illustrate these distinctions.