Post by Vishal Misra

Pi is maximally complex by one measure. Trivially simple by another. Shannon entropy says pi is incompressible - the digits look random, no pattern, infinite surprise. Kolmogorov complexity says pi is trivial - the generating program is four lines. Deep learning optimizes Shannon entropy. That is why it is extraordinary at what it does. It is also why it will never find G_μν = 8πT_μν. Einstein didn’t fit a better curve through Mercury’s data. He asked what kind of universe would make the anomalies disappear. That question - abductive, causal, generative - is the question (current) AI cannot yet ask. Two days ago Don Knuth published a note about Claude solving an open combinatorics problem. Claude found the pattern across 31 explorations. Knuth wrote the proof over a weekend. That division of labor is not anecdotal. It is structurally predicted. This is a preview of a conversation I recorded today with Martin Casado for the Andreessen Horowitz AI podcast, releasing next week. Full post: https://lnkd.in/e-GZE8Nm