What SI Force must do.

Six companies are racing to build superintelligence.

Every one of them knows that checking its work slows it down.

Researchers modeled this race in 2016. More teams and more distrust meant more corners cut. Even knowing more about each other's progress made it worse.¹

So each lab would be safer if all of them slowed down to check. And each one loses if it slows down alone.

The whole problem fits in four boxes.

Two labs, two moves: be careful, or race. The numbers are points, for illustration. Amber is Lab A's score, blue is Lab B's.

Look at it from either side. Whatever the other lab does, racing scores more. So both race, and both land in the box they'd never have chosen together.

Game theorists call this a prisoner's dilemma. It has one cruel rule:

You can't fix it from inside. A promise to be careful doesn't change a single number.

This year the players tried anyway.

In July, more than a thousand people at the major labs asked for a way to slow down together. Their statement admitted the "intense competitive pressure not to unilaterally slow."^3,5

In September, OpenAI asked members of Congress whether a shared slowdown would even be legal.³ The next day Anthropic's chief executive called for pacing the frontier together. According to a lawsuit, the heads of OpenAI, xAI and Google DeepMind endorsed it.⁴

Six days later, four labs were sued for it.

The suit calls coordinated pacing an illegal cartel under antitrust law.⁴ A legal analysis had warned of exactly this: the things that make a pause work, like commitments, checks and penalties, are the same things that make it a cartel.²

Without a referee, the careful move isn't just unrewarded. It's legally dangerous.

Then came the White House Accord on Super Intelligence. Six companies promised internal controls, outside auditors and board review.

It's voluntary, it has no penalties, and it gives the government no role in enforcing it.⁶ It changes none of the numbers.⁷

No one inside the race can change this. One player outside it can.

Racing pays. Pausing together can get you sued. Promises are free. That's the game. Only a player outside it can change what each move costs.

The Super Intelligence Force carries four weights no company has:

• The buyerWashington spent about $755 billion on contracts in fiscal 2024, $445 billion of it through the Pentagon.¹¹
• The eyesIts chair runs the intelligence community and can see what rivals abroad are building.⁸
• The refereeThe chair of the Federal Trade Commission sits on it, and antitrust is the FTC's job.⁸
• The tableIts mandate is to bring in consumers, public interest groups, religious organizations, critical infrastructure and the companies themselves.⁹

Move the weight.

Each point is a cost the Force attaches to racing: a lost contract, a failed audit, a liability. Past two, careful wins no matter what the other lab does. The trap opens.

It doesn't have to slow anyone down. The administration has rejected calls for a slowdown.⁸

It only has to make the careful move legal and the reckless one expensive. A rule every lab must follow isn't a cartel. It's the law.²

Govern means to steer.

It comes from the Greek kybernan, to steer a ship.¹⁴ A sailboat can head into the wind only because of its keel: heavy weight, carried low, out of sight.

Too little weight, and the wind decides where the boat goes.

That's the race.

Too much weight in the wrong place, and the boat goes under.

That's a single hand on the wheel: one founder, one company, or the government itself taking over the labs.

America has built a keel before.

"Ambition must be made to counteract ambition."

James Madison, 1788.¹² Checks and balances are weight placed so that no single hand can capsize the ship. Try it: move the weight.

Weight low, where it changes the game. Never on the wheel.

Seven moves, one for each question on our map.

Judge systems one person at a time.

Before the government uses a system, require evidence that it doesn't shrink any group's options toward zero, not just that it helps on average.

Match testing to freedom.

A system allowed to act on its own for a month must have been tested over a month. Pay for public measures of foresight, so power can't outrun it unseen.

Keep more than one hand on the release.

Above a capability line, releasing a system needs sign-off from several independent parties, and every lab says who holds its off-switch.

Make careful legal.

Ask Congress for one binding rule: above a capability line, every lab passes the same tests before release. Then the labs can pace together without being sued. Buy only from labs that pass, and publish the audits.

Tell the referee, not the rivals.

Labs report capabilities and incidents to the Force in confidence, because racers who watch each other cut more corners.¹ Build the counting and checking now, so a deal with China can be verified when one becomes possible.

Keep the table open.

Use the mandate to seat public interest and religious groups as a standing council, and publish what they say.

Decide in daylight.

Publish the 120-day findings as a draft for public comment before they become policy. And define Super Intelligence in law by what systems can do, not by what they're called.¹⁰

• Grab the wheel. Taking over the labs swaps one controller for another.
• Race without brakes. Being first to something no one can steer is not winning.
• Let the players referee themselves. A pledge with no penalties changes nothing.

Its weight is greatest now.

Early on, a technology is easy to change but hard to understand. Later it's easy to understand and hard to change.¹³ The Force sits in the short stretch where it can still do both.

Two dates. On 28 November the proposed legal definition of Super Intelligence is due to the President.¹⁰ Around 1 February the Force's 120 days run out.⁸

After that, what it decides gets set in law and habit.