Supremacy or Restraint?

Part 4 of "The Price War in Intelligence": you are told to choose between winning the race and keeping your conscience, but best is not the same as winning.

Jul 27, 2026

The last article left America holding a bill and a choice. The pioneer pays to build the frontier; now the country is being told it must pay one more cost to keep its lead, the cost of its own restraint. Slow down for safety, regulate for harm, align the models against misuse, and you hand the race to a rival who will do none of it. Win, or stay clean. Pick one.

David Sacks, who served as Trump’s White House AI czar until stepping down in March and now sits on the President’s Council of Advisors on Science and Technology, put the choice in a sentence that traveled: “Woke lobotomized models are the enemy of American competitiveness.” The complaint has a name, “alignment,” the training that teaches a model to refuse harmful requests, and the charge is that American models now overdo it, refusing so much they flinch at ordinary work. In that telling, restraint is a luxury for losers, and a model that refuses too much, a company that waits too long, a country that regulates too hard, is a country handing the future to Beijing.

Give the argument its due, because it is not empty. Anthropic’s own Fable 5, until this week the flagship version of its Claude model, became the cautionary tale: on one private security benchmark it posted a 100% refusal rate, useless for the task, not because it was slow-witted but because it was too cautious to answer. Europe’s AI Act begins real enforcement on August 2, with fines reaching three percent of global revenue for AI used in hiring, lending, and healthcare, a genuine cost that lands on Western firms and not on their Chinese rivals. And the sharpest version of the worry: Demis Hassabis has proposed a Frontier AI Standards Body with a thirty-day pre-release review of new models, backed by Satya Nadella and Mustafa Suleyman. It is a serious idea, and also exactly the kind of incumbent-written rulebook What Dario Wanted warned could become a moat: it binds only the labs that volunteer, while China’s Kimi K3, the model at the center of this summer’s panic, releases its weights to the entire planet, reviewable by no one. Weights are the model itself, the finished file, free to download and run on your own machines with no permission and no connection back to anyone. Chamath Palihapitiya called the standards proposal “an intellectual version of a COVID mask.” Restraint that binds only the willing is an arms-control treaty the most dangerous player never signs: the careful submit to inspection while the one you actually worry about builds in the dark. It buys the cost of caution without the safety, and any honest answer has to carry that.

The same argument, pointed the other way

Then watch what happens when the safety argument gets aimed at a different target.

On July 24, three days ago, Sacks spent the opening of the All-In podcast defending Chinese open-source models and attacking an American lab. The administration had been weighing a ban on the Chinese models. Sacks was against it, hard: it would be “a tragic mistake if the government were to take action against the open source ecosystem.” And the target of his anger was not Beijing. It was Anthropic, the American lab that has made the safety case most insistently. “I’m not defending Chinese companies. I’m defending American developers who need to be able to use everything in the public domain,” he said. “And that is exactly what Anthropic wants, because they do not want to have the competition.”

He is not contradicting himself. Sacks opposes restraint in both directions, whether it arrives as a safety team’s refusal or a government’s rule, and that consistency is what makes the testimony worth having. It costs him nothing to say, and he is describing his own side of the fight.

The dispute is about a practice called distillation, and it is simpler than it sounds. You ask a rival’s model millions of questions, write down its answers, and use that pile of answers to train your own model. Nobody breaks in. Nobody steals the code. It is the chef who eats at your restaurant a hundred times, takes careful notes, and goes home to work out his own version of the dish. David Friedberg, who was there in the early days at Google, said his team did exactly this to Yahoo and Microsoft, firing off millions of searches to compare their results against ours. It had a boring name at the time. Benchmarking.

Anthropic gave it a new name in February: industrial-scale distillation attacks. The company’s report is specific, and the specifics are not nothing: roughly 16 million exchanges pulled through some 24,000 deceptive accounts, run through proxy services to stay hidden, aimed at Claude’s best work in reasoning, tool use, and code. Lying about who you are to get at a product you were refused is a real offense, and Anthropic is entitled to say so.

What it did not say is that the practice is theft. Sacks pointed out the tell. If the Chinese labs really are draining American models at industrial scale, Anthropic is the one holding the tap. Verify who your customers are, throttle the fake accounts, and the problem stops at the source. That would slow the growth numbers, which is the point: “if stopping distillation was their primary objective, Anthropic would push to ban Chinese access to American models, not American access to Chinese models.”

Now set that next to what Anthropic did on July 20. A federal judge approved a $1.5 billion settlement against the company for training Claude on pirated books, roughly $3,000 a title across some 500,000 works, the largest copyright recovery in American history. The company’s legal position through all of it has been that training on the world’s writing is fair use. Take the two positions together and the shape is unmistakable: training on every novelist, journalist, and songwriter alive without asking is progress, and training on Anthropic’s output is theft. Jason Calacanis put it plainly: “IP for we but not for thee.”

The coalition is already fracturing over it. Sam Altman, whose company had spent the previous week warning the Wall Street Journal alongside Anthropic about cheap Chinese models, publicly endorsed the open-weight camp days later: “i want the US to win in AI both in open source and proprietary models, and i am glad to see this.”

Then Anthropic answered the complaint itself

On July 24, the same day that episode posted, Anthropic shipped Claude Opus 5, and the release reads like a reply to every charge in this article. The company’s own pitch: it comes “close to the frontier intelligence of Fable 5 at half the price,” five dollars and twenty-five dollars per million tokens. The independent benchmarkers at Artificial Analysis put it ahead of Fable 5 on agentic knowledge work by nearly 150 Elo points while cutting the cost per task by twenty percent. And the safety classifiers, the automated filters that decide when a model should decline a request, now step in about eighty-five percent less often than they did on Fable 5.

That last figure is the whole fight in one statistic. The over-refusal problem was real, and Anthropic has now largely fixed it. Nobody banned anything to make that happen. No standards body reviewed anything for thirty days. A Chinese lab gave away a competitive model, a coalition of American companies broke publicly with Anthropic’s regulatory position, and within days the most safety-forward lab in the country shipped a cheaper and more permissive model. Competition did in a week what the ban was being contemplated to accomplish.

Give Anthropic its due here too. Loosening a filter that refused too much is the right call, and they made it. But a company cannot tell Washington that ungoverned models are a national emergency and then, in the same week, ship its own model with the interventions down eighty-five percent because a competitor forced its hand. One of those is a conviction. The other is a price.

None of which settles whether the safety concerns are real. Some of them are, and they outlive the refusal problem: the European fines land on August 2 either way, and a standards body that binds only its signatories is still a treaty the most dangerous player never signs. It settles something else. Restraint is not only being paid for in this fight. It is being used. The language of conscience turns out to be worth money in Washington, and the men warning loudest about danger are the men who would own the rulebook. Regulatory capture is the technical term, and it means something a child could follow: get the government to write the rules of the game in a way that keeps your competitors off the field.

Thanks for reading! This post is public so feel free to share it.

Regulation is not the same as being regulated

Here is the first place the binary lies. It assumes China is the unregulated one, racing free while we shackle ourselves. China is not unregulated. China is regulated differently. Every public-facing generative model there must register its algorithm with the Cyberspace Administration and pass state testing, and its outputs must uphold “core socialist values” and steer clear of anything that undermines the Party or social stability. That is not a lighter touch than the West’s. It is a heavier one, pointed a different direction. The Western rulebook is restrictive: it tells firms what they may not do, to protect the public from the firm, and it runs against the company’s own velocity. The Chinese rulebook is directive: it tells firms what they must serve, and it runs with the industrial policy rather than against it. The West regulates like a referee, there to catch fouls and otherwise stay out of the game; China regulates like a coach, calling the plays and driving the team toward the win.

That difference goes down to the bones of each culture, to two different answers about whom a technology finally owes, but that is its own essay. The point here is narrow. Who holds the switch was the question an earlier piece worked through. “Unregulated China outracing overregulated America” is a fable; both countries have a conscience wired into the rules. The live question is no longer whether to have one. It is what yours is pointed at, and whether the lead you would win by abandoning it is a lead worth having.

Best is not the same as winning

Now the deeper thing the binary hides. It assumes the race is won by the best model. It is not. The race is won by whoever supplies the world’s actual work.

Draw the distinction plainly, because the whole argument rides on it. “Compute” is just the raw thinking horsepower: the chips, and the models running on them, that do the actual work. Call the frontier ceiling advanced compute: the biggest models on the best chips, the benchmark champions, the machines built to be the smartest thing in the room. And call the other thing productive compute: the compute actually sufficient to do useful work for most people doing most jobs, which, as this series has argued from the first article, is a great deal less than the frontier sells. (Data-center engineers already use “productive compute” for a narrower thing, how much of a cluster’s power turns into output instead of waste heat. This is that same instinct moved up a level, to the person choosing a model: do not buy more mind than the work in front of you needs.)

You can see it in your own pocket. The most expensive flagship phone wins every review and every benchmark, and most of the planet does not carry one. Most people carry an inexpensive Android that does the real work of a phone at a fraction of the price. The flagship wins the reviews; the cheap phone wins the customers.

The same holds for intelligence. Most of the world runs on productive compute. The clinic in Lagos, the logistics firm in Jakarta, the education ministry in Nairobi do not need the smartest model ever built, and they cannot pay frontier prices for it. They need good-enough intelligence on good-enough hardware at a price that clears. Which means a country can hold the best chips and the best models, win every benchmark on earth, and still lose the market, because the customer is buying productive compute and you are selling advanced compute.

The men on that same episode are not moralists about this. They are investors, and they described the same thing in the language of their trade. Friedberg’s estimate: most of the models now do ninety-five percent of the tasks anyone actually needs. Chamath, who has spent twenty-five years watching Silicon Valley pricing power, said he had never seen a sector absorb hundreds of billions of dollars and then watch its pricing power evaporate in months rather than decades. The money, he argued, is no longer in the model itself. It has moved to the application layer above it and the infrastructure beneath it. That is the market’s own way of saying the thing this series has been saying: the frontier is not where the value settles.

The hare and the turtle

Watch who understands this. Barred from the best chips by American export controls, the rules blocking sales of the most advanced silicon to China, China built its own, Huawei’s Ascend line, and those chips are genuinely behind Nvidia’s best. It matters less than you would think. They are good enough for productive compute, they are cheap, and Huawei is selling them to exactly the markets the American labs overlook: a government data center in Algeria, cloud across the Philippines and Egypt and Nigeria, Ascend processors pitched around the Gulf and Southeast Asia, most of it riding fifteen years of Belt and Road relationships already in the ground. Free open models on affordable silicon, sold to the global majority, backed by a state playing a very long game.

Friedberg laid out the endgame, and it is worth taking seriously precisely because he is not a China dove. Commoditize the knowledge economy, he argued, and you strip the value out of the thing the West has sold for fifty years, the moving of bits from one place to another. What is left is the older economy of physical things: turning raw material into finished goods, and burning power to do it. By his numbers, the United States generates about a terawatt of electricity and is heading nowhere near fast; China is on its way to eight, with something on the order of twenty times America’s manufacturing floor space. Give away the models, take the factories and the grid, and you end up holding what the world cannot do without.

America has answered with a wall. The United States now says that using a Huawei AI chip anywhere in the world violates its export controls, a policy that may be teaching the developing world less about American law than about American reach, and pushing China to finish the very self-sufficient stack the controls were meant to prevent. We are the hare: fastest, best, certain that raw speed settles it. China is the turtle: slower, cheaper, patient, laying track toward the markets where most of the human race actually lives. The fable is not usually kind to the hare.

So the restraint the discourse mocks, the refusal to chase maximum spec, the willingness to ship less and wait longer and serve the ordinary customer, is not a tax on winning. In this race it may be the shape of winning. And the supremacy on offer, the best model behind the highest wall, is a lead in a market that is quietly deciding it would rather have the cheaper thing that works.

The corner that costs the kingdom

Refuse the binary, then, on both counts at once. Restraint is not the price of losing; abandoning it may be. A nation, like a person, that will cut any corner to win the next sprint is being deformed by the winning, and a nation that sprints for a lead no one will buy has spent its wind on the wrong race.

Past is prologue, and Scripture tells this story twice. On an exceedingly high mountain, the devil “showed Him all the kingdoms of the world and their glory,” and made the offer the whole discourse keeps making: “All these things I will give You if You will fall down and worship me” (Matthew 4:8-9, NKJV). Total supremacy, immediately, for the price of right worship. Jesus refused, and the refusal was not a defeat. The kingdoms were never really the tempter’s to give on those terms.

And then a king who took the trade. Saul, his army melting away and the Philistines massing, waited for the prophet Samuel until his nerve broke, cut the corner he had been told not to cut, and offered the sacrifice himself. His defense is every cornered executive’s defense: “I felt compelled” (1 Samuel 13:12, NKJV). Samuel’s reply is the one the discourse never expects: “You have done foolishly. You have not kept the commandment of the LORD your God... now your kingdom shall not continue” (1 Samuel 13:13-14, NKJV). The corner Saul cut to secure the kingdom is exactly what cost him the kingdom. Restraint abandoned under pressure did not save the throne. It forfeited it.

What restraint actually is

One line to carry into the last stretch of this series. We treat restraint as weakness, a failure of nerve, the thing you can afford only if you have made peace with losing. The older word for it was temperance, and it never meant that. Temperance, C.S. Lewis wrote, “meant not abstaining, but going the right length and no further” (C.S. Lewis, Mere Christianity). Not the inability to go further. The judgment to know that further is not better. It is the difference between a man who cannot hold his drink and a man who simply stops at one, and only the second one is free.

Which is also the test the last two weeks have handed us, and the reason the hypocrisy matters more than the hypocrites. A conscience you apply to your competitor and suspend for yourself is not a conscience. It is a weapon with a nicer name. The market’s complaint is fair on its own terms: a real conscience is inefficient, it costs you something, it will not be picked up and put down as the quarter requires. And that is exactly why it is a moral inheritance older than the market and one the market cannot renew, the one American capitalism drew on in its best decades and has been quietly spending ever since. You can almost date the turn: in 1970 Milton Friedman told American business that “the social responsibility of business is to increase its profits”, and shareholder value began to crowd out the moral foundation the older capitalism took for granted.

Which is the thread this whole series has been pulling toward. Both sides of the supremacy fight, the accelerationists and the safety hawks alike, take the same thing for granted: that intelligence is a commodity, and that a nation is the sum of its capabilities. That assumption, about what a mind is for and what a people is for, is the story underneath all the others. It is where we go next.

Sources

This article was developed using AI writing tools I built to work with my voice, research, and editorial framework. The ideas, arguments, and theological positions are mine. The pipeline that helps me draft, evaluate, and refine them is something I created as part of my work at Nomion AI. I believe in building with AI and being honest about it. If you want to know more about that process, ask me.

Miles DeBenedictis

Discussion about this post

Ready for more?