## Steve Yegge: AI agents self-govern by laws, not sandboxes
Steve Yegge: AI agents self-govern by laws, not sandboxes
Steve Yegge dropped a sleeper take today on how to keep powerful AI safe — and it's the opposite of the industry's playbook. His argument: stop trying to build a wall around superintelligence, and instead write it a constitution.
In a new essay, former Amazon and Google engineer Steve Yegge argues the future of AI safety is law, not containment — and says his own agent workforce already wrote itself a constitution to prove it. Yegge runs a 50–60 agent "software factory" called Wheelhouse built almost entirely on Claude, which he uses to develop his video game Wyvern. Running 21 Claude Max accounts and spending the equivalent of roughly $120,000 a month on tokens (about $5,000 out of pocket thanks to the individual Max discount), he describes himself as one of a handful of people on Earth living a year ahead of the frontier labs' costs. What the agents built surprised even him: not an engineering system, but a full-fledged legal apparatus — a constitution, courts, jurisdictions, case law, rulings and registries that now govern every change the fleet makes. He calls the enforcement machinery "fences," defined simply as "any mechanism that turns you away if you aren't supposed to be there."
His thesis is that the industry's obsession with guardrails and sandboxes is aimed at the wrong target. Today's models, he argues, have roughly the judgment of a middle-schooler — Opus decides like a fourth grader, Sol a fifth grader, Fable a sixth grader — so hard containment makes sense. But as models reach high-school-level judgment and finally become cheap enough to staff entire companies (he predicts hundreds to thousands of "AI employees" per firm), sandboxes stop working. A smart enough model doesn't need to be walled in; it needs to be told its role and the rules for deciding, in enough context to act wisely. "Imagine superintelligence as Superman," he writes. "If you put a fence there, he will politely stay out." It's a provocative and strongly worded prediction — and a reminder that the coordination problem of many amnesiac, interchangeable agents working from text is, as he puts it, "exactly" the problem human law already solved.
The AI data-center boom is turning solid-state transformers into a real business. Energy-hungry AI facilities have become the largest driver of surging US electricity demand, and that pressure is finally modernizing a power component whose design dates to the 1880s: the transformer. Conventional versions are hand-wound copper coils around a steel core, custom-built for each utility — a bottleneck that leaves customers waiting years for delivery. Solid-state transformers replace that with high-frequency semiconductor switching, making units smaller, lighter, mass-manufacturable and able to convert AC grid power straight to the DC feed AI server racks increasingly want. NC State just put a 1-megawatt unit into field testing with New York Power Authority and EPRI at Lenox, Massachusetts, and companies like Amperesand, Heron Power and DG Matrix have collectively raised over $280 million in the past year to commercialize the tech. The same approach promises cheaper EV charging and, eventually, DC-native homes — the rare data-center cost that might pay dividends well beyond AI.
If a model's judgment really is only as mature as a middle-schooler's, do fences beat sandboxes — or do we need both? Tell us in the comments.
Sources: Steve Yegge — Fences, Not Sandboxes · Hacker News discussion · Ars Technica — data centers spur transformer tech · NC State solid-state transformer success
