For the last sixty years, "machine verification" — mathematically proving that a piece of software or hardware does exactly what it's supposed to do — has been a luxury reserved for the most critical systems: aircraft controllers, cryptography, nuclear reactors. It's just too expensive and slow for everyday engineering. This paper argues that generative AI flips that economic reality on its head.

The premise is simple but striking: when an AI can crank out code, chip designs, or system configurations faster than any human could review them, you can't rely on human eyeballs to catch mistakes anymore. You need a referee that can't be bribed or fooled — and that's exactly what a formal verifier is. It doesn't get tired, doesn't hand-wave, and doesn't accept plausible-looking nonsense. If the AI produces something wrong, the verifier rejects it. If it accepts it, you have a mathematical guarantee.

To demonstrate this, the author ran a five-week solo experiment. Working alone, on consumer AI subscriptions (think ChatGPT Pro tier, not a datacenter), he directed a fleet of AI agents to build software spanning the entire computing stack — from application-level code all the way down to actual silicon (chip) designs. At every layer, the AI's output was funneled through a verifier that acted as the gatekeeper.

The key insight is a role reversal:

• Old world: humans write code, verification is an optional afterthought that costs too much.
• New world: AI writes code at superhuman speed, and verification becomes the only affordable way to keep up — because a human simply cannot audit that volume of output.

Put another way: verification transforms from a cost center into the very thing that makes one-person-plus-AI teams possible. Without it, you'd drown in unreviewable AI output and eventually ship something broken. With it, a single engineer can safely orchestrate work that used to require a whole team, because the verifier catches the AI's inevitable hallucinations and errors before they matter.

The paper positions this as the emergence of "AI with authority" — AI that isn't just suggesting, but actually doing, with a mathematical safety net rather than a human one. It's a claim about how software (and hardware) engineering itself is about to be restructured: fewer humans writing code, more humans specifying what they want and letting verified AI agents build it.