Enabling Machine Intelligence
Full disclosure: I made extensive use of AI to build the systems described here. I provide the scientific direction, the invariants every result must preserve, and the benchmark and profiling harnesses that decide whether an idea survives; the agents do much of the implementation, and sometimes find solutions beyond what I could produce alone. With Hindsight and typed documentation organized as a graph of claims, constraints, and evidence—not a pile of transcripts—they began to preserve and apply my taste across sessions. The turning point came when an agent recalled an earlier asymptotic-complexity collapse and reused it more effectively than I could in the moment. That arrangement can push the bounds of what is practically possible; Anthropic’s recent work on the Riemann zeta function is a striking public example. Claiming sole credit for the more extravagant results would be dishonest.