Dispatch from Mitch
Did China obtain a key AI chip-making machine?
Can a 180-ton machine the size of a school bus just go missing?
Maybe. This is the grave concern of Commerce Secretary Howard Lutnick, as reported by Bloomberg.
For background, the most fundamental step in computer chip manufacturing is photolithography — writing microscopic patterns onto semiconductor wafers with intense light and then chemically etching away the marked areas (or the reverse, depending on the process), layer by layer.
Generally speaking, the smaller the features you can etch out, the faster, denser, and more efficient you can make the chip. To etch out the smallest features, you need photolithography machines that can generate intense light at very short wavelengths and focus it very precisely with minimal loss. This is incredibly hard.
The most advanced AI chips require the most advanced photolithography machines, which generate and focus Extreme Ultraviolet (EUV) light. These bus-sized machines cost more than $400 million and are only made by a single Dutch company, ASML. They are also very fickle, requiring constant maintenance by ASML employees. There are only 314 of them in the world, 26 of which have been decommissioned.
This makes EUV technology a bottleneck — one we’ve examined before as a factor making it tractable to globally ban the development of artificial superintelligence.
To say that replicating ASML’s technology would be no easy feat is a gross understatement. But it would probably be a lot easier if you had your own EUV machine to play with.
Now we can understand why Commerce Secretary Lutnick might be concerned if he has reason to think one of ASML’s EUV machines has made its way to China. He reportedly expressed this suspicion to ASML executives in a series of meetings that began in April, without providing any evidence.
The company correctly perceived this as a serious allegation, and has been choosing its official words about it very carefully. A company spokesperson said:
ASML has never shipped an EUV machine to China nor have we shipped to China any component, module or equipment specially designed to be used in an EUV machine.
And:
We recognize the national security considerations behind export control regulations in the US and the Netherlands, and we’re fully committed to complying with all applicable regulations.
But I don’t see anyone accusing ASML of outright selling one of these machines to a Chinese company. More likely, it would have been transferred by an existing customer who may have been misled about its final destination. But this should be hard to pull off without ASML knowing, given the limited number of machines and their aforementioned need for constant expert maintenance. ASML says it has no reason to think any EUV systems are currently in China, but senior U.S. officials have claimed to have evidence that ASML shipped, to destinations in China, specialized equipment needed to move EUV machines.
If China has an EUV machine, and ASML is complicit in this technology transfer, this could seriously damage relations between the U.S., ASML, and the Netherlands. The Trump administration could feel compelled to take aggressive punitive and restrictive measures. U.S. officials were already talking about cutting ASML off from all business dealings with China at the start of Trump’s current term.
An alternative theory behind U.S. concerns is that Chinese chip companies have independently made progress that makes it look like they must have obtained an EUV machine when in fact they haven’t. China has, after all, been making a Manhattan Project out of developing world-class chip fabrication capacity.
But to me, the smuggled machine theory seems more likely, because it rhymes with the rampant smuggling of finished chips that has propped up Chinese AI companies. The AI supply chain isn’t nearly as carefully tracked as it should be. This needs to change if the world hopes to retain any optionality over the race to superintelligence.
Dispatches from Alana
Luddite love?
NPR’s word of the week was “Luddite”, a term often used to refer to someone who hates modern technology and is against progress. You’ve probably heard it in the context of AI, with critics of the AI explosion viewed by some as simply “behind the times” or “scared of all technology.”
But according to NPR’s deep dive, “Luddite” actually means something quite different. The article quotes tech journalist Brian Merchant, who identifies as a Luddite himself. He says Luddites aren’t anti-tech; they are instead against tech being used to exploit people. And who could argue with that?
A Luddite asks: What are the implications of this technology? How is it going to impact society? Should we engage with this technology on the grounds that it might make somebody a lot of money or should we engage with it on the grounds that this could have real impacts for the way that people work and live?
Apparently, the term Luddite comes from an actual group of people: textile makers in the 1800s who revolted against companies adopting automated looms. These looms could be operated by anyone, making their artisan skills — which required years of apprenticeships — irrelevant, and resulting in lower wages and lower quality. The revolts turned violent and were quelled, with many Luddites hanged. The term took on a derogatory meaning.
I didn’t know “Luddite” had such a colorful history, and I’m fascinated to learn how the term has been oversimplified. When people raise valid concerns about a new technology (concerns like job loss or worker mistreatment) they’re often labeled “anti-innovation” rather than “person raising concerns about new innovation.” I don’t condone violent rebellions, or smashing looms, and for those reasons, I certainly don’t laud the Luddites. But I think there’s a broader lesson here: we should be able to ask questions like: “How will this new technology affect us?” and “Has anyone really thought through the implications or are we just racing ahead blindly?” without being labeled as anti-technology.
Towards the end of this deeper exploration into the Luddites, YouTube historian “The History Guy” summarizes the view of Kevin Binfield, who edited a collection of Luddite writing: “It wasn’t technology that made them angry. It was unscrupulous businessmen who were threatening their very lives by driving down prices.”
The term “Luddite” is apparently being reclaimed, with Gen Zers starting Luddite clubs at colleges and a large backlash against social media. I’m not about to give up my smartphone. But the next time “Luddite” is used to describe those raising concerns about AI… maybe it’s time for a quick history lesson?
A mad, mad, mad, mad world
The world is absurd. And people are greedy. These are two throughlines from the 1963 film It’s a Mad, Mad, Mad, Mad World, in which various factions compete to find a suitcase full of money only to end up in full body casts by the time the credits roll.
They were also my two takeaways after learning about the project World ID, covered in a Forbes piece discussing the need for infrastructure to make digital actions more accountable, whether by distinguishing humans from machines, or knowing which model is behind an AI agent. That’s an interesting topic in itself, but I’ll focus today’s dispatch on World ID, which offers a way for humans to prove they are human. (Yes, we need such a thing right now. As I said, the world is absurd. I’ll get to greedy in a second.)
World ID is intended to be used for things like social media, dating apps, and gaming, and promises the ability to “prove you are a unique human, without revealing anything else about you.” To do this, you download an app and then visit the “Orb”, a shiny, sphere-shaped device which scans your irises. You can find the Orb at World store locations around the globe or other locations. (Just in case you want to celebrate your proof of humanity with a meal, Everytable restaurants are among the options.) After verification with the Orb, you get a “proof of human” ID, which is stored on your phone. The company promises to delete your data from everywhere else.
World.org, which houses World ID, greets viewers with a full screen slogan proclaiming: “Let’s make the internet more human.” And the company behind it is named “Tools for Humanity.” This, to me, sounded like one of the many orgs criticizing big tech and espousing “humans first” messaging. So imagine my surprise when I learned the company was actually co-founded by Sam Altman of OpenAI, the company behind ChatGPT.
The world is absurd. People are greedy. Tools for Humanity is valued at around $2.5 billion, though it may struggle to generate revenue. Sure, there’s a charitable interpretation here where Sam is trying to address the problems his tech is creating. But, given what we know about the guy, I think the cynical interpretation is more likely. Why not pander to the AI-optimizing crowd and the “let’s make the internet more human” crowd? You gotta look for that suitcase of money in every possible direction.
Dispatch from Donald
A call to move fast and regulate things
“Move fast and break things” has long been the strategy of the tech industry: Deploy what you can, when you can, and entrench yourself before the government can say whether there should be some ground rules. In this respect, the application of this principle within the AI industry is nothing new, except that it involves radically dangerous technology. Today’s models push the frontier before regulators understand the import of yesterday’s models.
For the latest and most severe example of AI development outpacing government regulation, Forbes’s Craig S. Smith points to the three-day lifespan of Fable 5. The government is now working with Anthropic to build a framework for identifying, grading, and responding to security flaws in frontier models. But this only happened after Fable 5 was pulled due to fears that it could be jailbroken and used for cyberattacks. In the two months since Fable’s more capable and threatening counterpart Mythos was announced, this policy framework had not been built.
Smith also refers to a recently published Google DeepMind paper, “From AGI to ASI.” It lays out four possible routes to artificial superintelligence, including recursive self-improvement. Recursive self-improvement is the capacity for an AI to rapidly increase the pace of its own development in a feedback loop; powerful AI begets even more powerful AI. Of the four routes examined by Google DeepMind, recursive self-improvement poses the greatest threat to regulation because it promises the most rapid pace of change. The government must regulate AI, but to do so it must move rapidly: “There’s little reason,” Smith writes, “to expect the labs to slow down so the rule-writing can catch up.”
I would add that governments need to not just react to, but also anticipate, what is coming. We need to build frameworks to regulate models before those models are released or even built. If the pace of development is too fast to build adequate frameworks then we need to halt development entirely.
The analyses and opinions expressed on AI StopWatch reflect the views of the individual contributors and the sources they cover, and should not be taken as official positions of the Machine Intelligence Research Institute.