The potential dangers of A.I. are weighing on many minds this week. They’re being discussed by world leaders at the U.N. General Assembly; they will almost certainly be discussed when Donald Trump and Xi Jinping sit down for talks tomorrow in Washington.
Calls for global cooperation to slow the development of what is known as frontier A.I. — the most advanced, most capable A.I. models — have grown louder. So far there isn’t a whole lot of optimism about the world’s two A.I. superpowers reaching an agreement any time soon.
But there’s a precedent for this moment. Today, I write about what global efforts to control nuclear weapons can teach us about controlling A.I.
We’ve been here before
Two rival superpowers. Technology that has the capacity to be enormously destructive. A moment of awakening that reveals just how much danger the world could be in.
There’s a lot that’s unprecedented about this moment when it comes to A.I. But in some respects, we’ve been here before.
During the Cold War, the U.S. and the Soviet Union were sworn enemies, locked in a race to build ever more, and ever more powerful, nuclear weapons. Then came the Cuban Missile Crisis. Thirteen tense days focused minds, and in the years that followed, the two nations went on to contain what was then the gravest known threat to humanity.
It’s easy to forget that a lot of people didn’t think this was possible. Today, commentators and officials are skeptical about an A.I. treaty between the U.S. and China, too.
The nuclear analogy goes only so far. There are important differences between artificial intelligence and nuclear weapons, and some of those might make reaching an agreement to constrain A.I. more difficult. But some could arguably make it easier.
Vast differences
Artificial intelligence and nuclear weapons are very different technologies.
They’re used differently. A.I. is developed by private companies, and is effectively in the hands of anyone who has access to a computer or smartphone. Nuclear weapons are … not. They’re controlled by governments. Even their civilian equivalent, nuclear power plants, require expensive, highly specialized physical infrastructure. (Though the differences are not as stark as they might seem — more on that in a bit.)
But perhaps the biggest difference is that nuclear technology was well understood for decades before the world moved to control the spread of nuclear weapons.
A.I., by contrast, remains something of a black box. Even America’s frontier labs cannot always understand their own A.I. models; indeed, that is one of the worries driving calls to slow down the technology. But it’s not easy to regulate something you don’t understand.
Trust but verify 2.0
Still, some principles of nuclear arms control can be applied directly to A.I. For instance, the first step toward any agreement between two deeply distrustful parties is to build, well, trust. And the way to do that is to start sharing information.
The first thing the U.S. and the Soviet Union did, only months after the Cuban Missile Crisis, was set up a hotline, a 24-hour telecommunications link from Washington to Moscow. The idea was to give leaders a chance to communicate instantaneously to prevent accidental escalation.
That’s the kind of thing A.I. policy experts say could be adapted readily. Connecting not only Washington and Beijing but also the frontier labs would allow all sides to share information about worrying incidents. (In fact, my colleagues have reported that U.S. and Chinese officials have already discussed implementing such a system.)
Another is a nuclear-style inspection and monitoring regime. The International Atomic Energy Agency conducts intrusive in-person inspections of nuclear power plants in many countries to try to ensure nuclear material is used for peaceful purposes.
Instead of uranium enrichment, A.I. inspectors would monitor, say, caps on compute, the raw processing power that governs how powerful large language models are, said Anthony Aguirre, chief executive of the Future of Life Institute and an expert on A.I. safety and governance.
And A.I. might be a widely accessible tool, but the data centers that host it are as large-scale and specialized as enrichment facilities; satellites could help with monitoring them too.
In one respect, Aguirre said, it might even be easier to keep tabs on some A.I. The most advanced A.I. chips are, for now at least, typically manufactured by one Taiwanese company with machines produced by one Dutch company. In other words, they should be easy to trace. And they already have capabilities that can monitor A.I. computing power and be deactivated from afar if that compute is tampered with. “There’s already a kill switch, if we choose to use it,” Aguirre said. “But we need the political will.”
The Hiroshima factor
That’s of course where the rubber meets the road. China and the U.S. have major trust issues. Commitments that involve private companies, not just governments, add a complicating factor. And Beijing may balk at the idea of freezing A.I. development at a time when it’s still behind in the race. (The Soviets only agreed to cap their arsenal once they’d pulled even.)
Perhaps the biggest problem is that humans still don’t agree on the actual threat they’re facing.
That was not the case with nuclear weapons. Images of Hiroshima and Nagasaki are seared into our collective consciousness. The recent spate of rogue A.I. incidents has people concerned, but not at anything close to the same scale.
Still, alarm is growing. Aguirre’s institute organized the Pro-Human Assembly last week, a summit where a wide variety of people debated A.I. regulation. Steve Bannon and Bernie Sanders appeared on the stage, one following the other, to demand tighter restrictions.
Bannon and Sanders are not exactly political allies, but the potential dangers of A.I. brought them together. It’s not the U.S. and China, but it’s a start.
Trump’s threat at the U.N.
President Trump told world leaders at the United Nations yesterday that he had a decision to make on Iran: strike a deal with Tehran or “annihilate the Islamic republic.”
In a combative speech at the U.N. General Assembly, Trump claimed he had prevented Iran from obtaining a nuclear weapon, and asserted that he could still destroy the entire country. “Do I drive them into hell with no chance of survival and no hope of future greatness or generations?” Trump said. It was startling to hear such threats delivered from the U.N. podium, where world leaders typically champion peace and global cooperation. We fact-checked his speech.
Iran’s president, Masoud Pezeshkian, will deliver his address today, a first for a country at war with the U.S.
“Swan Song: Diana, My Sister”
In a new memoir, Earl Spencer revisits the days surrounding the death of Princess Diana, his sister. The most incendiary claims: Spencer’s descriptions of the telephone conversations he says he had with Prince Charles, now king, after Diana’s death. “They read like grenades lobbed lazily over a fence and intended to destroy everything in a neighbor’s yard,” my colleague Sarah Lyall writes. Read about the claims here.
MORNING READ
Ties between Japan and Russia were tested when Vladimir Putin made a surprise visit in August to islands seized by the Soviet Union from Japan in World War II. Nowhere are the tensions more deeply felt than in Nemuro, a Japanese fishing town known for the bright-red king crabs harvested on its shores.
Inside tech party houses
Tech group homes have proliferated in Silicon Valley during the A.I. boom. Residents are often start-up founders or engineers working at OpenAI or Anthropic. Rent can hit $10,000 a month for a room.
One of these is an eight-bedroom mansion next to a golf course. It has gone through a lot in recent years. At one party, a car flew through the air. At another, guests arrived in horse-drawn carriages to a replica of the Eiffel Tower. But what has unfolded in many of these homes has sometimes gone far beyond networking. Read about what the police reports reveal.
Hiyashi chuka ramen, a cold noodle dish popular in Japan, highlights seasonal vegetables. This version uses a soy-based broth with cucumbers, ham, corn and tomato, but feel free to change the toppings.
