Software Survival 3.0

Take Karpathy and Amodei at their word — AI that can build almost anything on demand — and the obvious question is which software is left standing. The essay answers it with a selection argument: inference costs tokens, tokens cost energy, energy costs money, and treated as one constrained resource, a world of resource-limited superintelligence starts behaving like any other ecosystem under scarcity. The rule that falls out is software tends to survive if it saves cognition. The framework is the Survival Ratio — cognitive value over cognitive cost, six levers to pull, ratio clears 1 to survive. Two levers raise savings (compress insight the way Git does; move work to a cheaper substrate the way grep does); two more govern whether agents reach for you at all (awareness, the pre-sales problem; friction, the post-sales one); a fifth is breadth of use; the sixth is the human coefficient, the demand for things specifically because a person made them. Recurring survivors — Git, grep, Temporal, Postgres, Dolt — are tools it would be absurd to re-synthesize, what Brendan Hopper calls crystallized cognition.

Delivered as "Squirrel Math," credited to Marv's chonkulous, ordinance-defying backyard squirrel — a running joke that lets Steve disclaim the rigor while still making the case. It lands hopeful: six levers is a lot of paths forward, desire-path design works without an OpenAI training budget, and our appetite for new software is effectively infinite. Build something it would be crazy to re-synthesize, make it easy to find, make it easy to use.