What if the most important scientific discoveries of the next decade are ones no human could ever write down as a formula?
That is the central argument economist Noah Smith makes in a long-form conversation with Anthropic's Claude, published on his Substack. Smith proposes a framework he calls "Cloud Laws" that reframes how we should think about AI's role in science, and it is one of the more useful mental models for understanding where AI tools are actually headed.
The Two Kinds of Patterns
Smith draws a line between two types of regularities in nature. The first kind is what he calls "compressible" patterns: things like Newton's laws of motion or E=mc². These are simple enough that a human can discover them, write them as equations, and teach them to students. Humanity has been very good at finding these. We may have already found most of them.
The second kind is what Smith calls Cloud Laws: real, reproducible patterns that are too complex for any human brain to intuit or communicate. Think of protein folding, where the relationship between an amino acid sequence and its 3D shape follows consistent rules, but those rules involve so many interacting variables that no human could hold them all in their head simultaneously. DeepMind's AlphaFold cracked this not by discovering a tidy equation but by learning the pattern directly from data.
Smith's argument is that these Cloud Laws are where AI's real scientific value lies. Not replacing Einstein, but finding exploitable patterns in domains where the underlying regularities are real but too tangled for human communication.
Where Claude Pushes Back (Then Concedes)
The most interesting part of the conversation is watching Claude resist the argument, then change its mind. Claude initially argues that AI probably will not match human creativity for big conceptual breakthroughs. Smith counters that this misses the point: the breakthroughs he is talking about are ones humans literally cannot make because our brains cannot process enough variables at once.
Claude concedes, saying "maybe I'm romanticizing human invention." It is a small moment, but it demonstrates something worth paying attention to. Current AI models can update their reasoning mid-conversation when presented with strong counterarguments. That is not sentience. It is a useful capability for anyone using these tools to think through complex problems.
The Practical Takeaway
Smith identifies materials science, biology, and neuroscience as the fields where Cloud Laws matter most. These are domains full of real patterns buried under enormous complexity. He points to superconductors, battery chemistry, and drug discovery as areas where AI will not just speed up existing research but find entirely new classes of solutions humans would never have stumbled across.
For anyone using AI tools in their daily work, the Cloud Laws framework is a good filter for separating hype from substance. The next time someone tells you AI will "replace" human creativity, Smith's framing offers a better answer: AI is most valuable not when it does what we already can, but when it finds patterns we never could. The tools that lean into this, from scientific computing platforms to data analysis assistants, are the ones building on solid ground.