Software Survival 3.0

A strategy memo for the boards and C-suites now asking whether their software survives the Software 3.0 era, delivered by a man who rates his own evolutionary-biology credentials at exactly the level of his 82-year-old neighbor Marv's illegally overfed squirrel. The costume is load-bearing: it lets Steve hand executives a fitness function while admitting, more than once, that it isn't real math. He also owns up to the model's messy history — Claude talked him out of a stack of claims he couldn't defend, and the whole thing began as a thermodynamics argument before he and the squirrel "failed that exam even worse" and switched to evolution. Two tastes of what's inside: why an LLM doing long multiplication is "as bad as a person using nothing but chickens," and hallucination squatting, a genuine supply-chain attack Steve inverted into the benign design method behind Beads' hundred-subcommand CLI.

Read it knowing the evidence skews homemade. Beads and Gas Town appear as case studies inside their author's own general theory, and the canonical survivors he cites are tools nobody would dispute anyway. What the framework earns is a vocabulary: a name and a knob for each reason an agent does or does not reach for a tool, offered to executives who were having this conversation with no framework at all. And near the end, a warning aimed at the opposite of where AI doom usually points — the software in the worst trouble is anything trying to be "smart" in the space between a human and an AI.