Foundation

Better materials have always been the quiet engine of innovation: bronze made tools, silicon made computing, and whatever comes next will be designed, not stumbled into. Material science builds useful substances from the atomic level up, and AI has begun automating the trial-and-error that used to take careers. Once a material can be represented in language, a language model can do more than recall what is known about it; it can begin to reason with it. Foundation is the question I built to test that: an agent that fuses Aristotle's First Principles Thinking with LLM material knowledge, scoring every candidate on a single convenience metric, how practical a substance really is to manufacture, sustain, and design with. If an agent starts from the true root of a problem, can it reach a fundamentally better material, faster?

The honest finding is a no. Wired into an automated pipeline, First Principles Thinking is overwhelmingly likely to hurt the convenience calculation rather than sharpen it: the agent falls into fatal thought-loops, reasoning in circles until accuracy collapses. Six months, one clean lesson. The way humans think at our best is not yet the way agents think at theirs. The philosophy gave the study its shape, then broke the machine that tried to run it literally, and that collision is exactly where the next generation of computational material discovery has to evolve.