When Code Gets Cheap, Verification Becomes Expensive: How AI Changes the Economics of Software Architecture

Writing code is mostly a one-time cost, but proving it correct is paid again with every change, and AI is multiplying the changes. Remo H. Jansen argues that verifiability belongs alongside performance and maintainability as an architectural concern, starting with making invalid states impossible to represent.

How Our Engineering Team Uses AI, Part II: Meat Proxies

Eight months after MetalBear found AI useful only around the edges, most of its engineers barely write code by hand. The work that grew is review, and with it the fatigue of making sure someone understands every line before it ships.

AI Slop Is the Second Ship

A module rewritten plank by plank by an agent can still be yours, as long as someone knows why each piece is there. Jeel Vankhede uses the Ship of Theseus to argue that slop is the second ship: code assembled from the average of the internet that nobody aboard can explain.

Mercury 3: When AI Is Fast Enough to Stay in the Loop

Maxim Saplin put Inception's diffusion model through a handheld controller fix, a Home Assistant test-suite cleanup, and his chess and prime-number evals. It kept pace with far pricier models, though it burns through tokens and needed review like the rest.

I Built an AI That Rewrites Its Own Prompts: Its Safety Gate Rejected Every Single Edit

Debashish Ghosal built a sidecar that proposes prompt edits from execution traces and promotes only those that pass a statistical test. Over 15 iterations and 4,150 LLM calls the gate approved nothing, and it was right every time.