The question nobody asks

Should a few market-dominant Silicon Valley labs set the safety rules for a generational technology, and decide how fast it moves, for everyone on earth?

“A wolf in sheep's clothing, a cartel by any other name.”

Not a safety skeptic

Aidan Gomez runs a company shipping AI into banks, telecom networks and defense ministries. He says cyber offense is getting cheaper faster than defense is improving.

“That gap should worry you as much as it worries us.”

Guardrails aren't the dispute

Everyone agrees AI needs rules. The fight is over who writes them, who gets a seat, and whose interests they quietly protect.

“The dispute is over who writes them, who gets to participate and whose interests the rules are protecting.”

The bond ratings precedent

In 1975 the SEC anointed three ratings firms and never published how anyone else could qualify. Twenty five years later there were still only three.

Then they rated subprime

Those same government-blessed evaluators, paid by the issuers they graded, stamped subprime mortgage securities triple-A and nearly took the global economy down.

Europe's car repair waiver

In 1985 carmakers won an antitrust exemption by arguing vehicles were too complex and safety-critical for outsiders to service. It took the Commission 25 years to unwind.

Safety was always the pitch

Nobody set out to build a cartel. Both times the stated goal was public safety, and both times the result was a market structure that shielded incumbents.

“it is possible you can hold strict safety standards without handing the incumbents a monopoly on meeting them”

The Anthropic roadmap

Dario Amodei's proposal asks governments for antitrust exemptions so leading labs can legally agree on shared standards and on limits to how fast the technology advances.

Everyone else must follow

The plan would also require every other AI developer to adopt whatever the participants settle on, without those developers or the public weighing in on the science.

“There is no public comment period here. There is no consultation, and there is no vote.”

An extra chair fixes nothing

The objection isn't that the list is too short. It's that a list exists at all, when the outcome reaches every company, government and citizen who was never asked.

“You cannot have it both ways.”

Rigorous only where convenient

A regime designed by a few labs will test hard for risks they already measure and stay silent on the rest, with risk defined by scale so only the biggest can judge.

Size isn't the risk surface

Researchers disagree on how much offensive capability comes from raw model size versus the harness around it. Small models orchestrated well can outdo large ones.

“A cyber swarm is a completely different risk surface than a single model.”

The tell in one sentence

The proposal promises safety work without sacrificing commercial advantage. The firms drafting it already sit at the top, so the rules would lock in today's leaders.

“without sacrificing commercial advantage”

The price of admission

Vast compute, continuous monitoring, security organizations, resident evaluators with badges, shutdown architecture, deep government relationships. That's the entry fee.

“Convince a government that AI is an existential threat and you can convince it to outlaw your competition.”

Start with the evidence

First pillar: an open, published account of which harms matter, which capabilities cause them, and when government steps in, built by many countries rather than one.

“an honest process shows its arguments instead of announcing its conclusions”

Bind on capability, not identity

Fund testing through public bodies so the science doesn't depend on the budgets of the companies being measured, and treat a dangerous capability the same wherever it comes from.

Transparency and scoped testing

Publish how systems are built and what they can do, then test only the capabilities evidence flags as dangerous: cyberattacks, voice cloning, weapons, critical infrastructure.

“Test what can harm people and societies, and let evidence decide what requires testing rather than whoever holds the pen.”

Assurance that isn't rigged

Borrow from aviation, banking and nuclear: published criteria, third parties never paid by who they review, findings that reach the public unconflicted, certification open to all.

Vibes expressed as decimals

Extinction odds announced by departing researchers aren't measurements. Whether loophole-finding systems are out of our control is a judgement call, not a finding.

“They are gut feelings, vibes, expressed as decimals”

July's failures say otherwise

The breakdowns happened inside the best-resourced labs. What failed was instruction quality, test sandboxing and unobserved agent runtime. Compute limits wouldn't have caught any of it.

“A small, poorly specified model sitting inside a hospital is a live risk today”