Vic argues, often against consensus, about how technology should be governed. A former regulator, Vic writes the column they wish they could have read while still in government.
A federal judge found the Pentagon's blacklist of the Claude maker illegal and retaliatory. That's a win for how a government must act — not an answer to who decides how the military uses AI.
A company paused a training run because a rubric it wrote and scores itself told it to. Reading that as governance working is a category error, and a costly one.
New Mexico used a public-nuisance verdict to impose a notification curfew, a screen-time cap and a $567 million fund on Meta. The harms are real. Governing product design by tort, one courtroom at a time, is still the wrong instrument — and here's the procedural reason why.
The AI Act's marquee date came and went, and the high-risk obligations that were supposed to bite had already been pushed to 2027 and 2028 by an amendment that isn't even law yet. Calling that either a triumph or a retreat gets the same thing wrong: "the deadline" was never a single switch, and the calendar was never the achievement.
A voluntary framework with OpenAI, Anthropic and Google, expected within days, lets federal evaluators test frontier models before release. It is neither the licensing regime its critics fear nor the voluntary standard its authors claim — and that third thing answers to no one.
Reading Beijing's H200 rationing as a market reopening misses what actually happened: export control has become industrial allocation, run on both ends by discretion no one can audit. A cap set at "less than half what they asked" is a dial, not a policy.
Four senators from both parties agree the CREATE AI Act should make NAIRR permanent. Treating that quiet consensus as proof the loud AI fights are just as solvable would be a category error of its own.
The two popular framings — block the data centers, or clear the runway for them — are both category errors. What the commission actually did on June 18 is harder to chant and more likely to work, if its own deadlines don't wreck it.
The moratorium fight treats "AI regulation" as one switch to flip on or off. The laws actually moving — on chatbots, on hiring, on kids — are narrow, old, and mostly about who is accountable. Missing that is the category error.
OpenAI caught China-linked accounts amplifying America's data-center backlash. The comfortable lesson is that the backlash is manufactured. Learn that lesson and you lose both the information fight and the siting fight.
Washington wants to approve frontier AI models the way the FDA approves drugs. The analogy is comforting, popular, and wrong about how software actually fails. I say this as someone who would have been the agency.
While Washington argues about regulating models, the state is already governing AI through a far blunter instrument: who gets to package a chip. Get that frame wrong and you mistake the sideshow for the policy.
