The Risk Is Not China Copying. It Is America Refusing to Use the Tools

Cognitive Revolution host Nathan Labenz introduces a session recorded at the Hill & Valley Forum, a private bipartisan gathering of lawmakers and innovators convened around America’s national-security problems. The moderator is Senator Cory Booker. Across from him sit three people who actually ship the systems: Alex Wang, founder and CEO of Scale AI; Jack Clark, co-founder of Anthropic; and Andrew Ng, managing partner of AI Fund and co-author of more than two hundred AI papers.

Labenz says the three men agreed on a point he shares: China is already producing outstanding research, will not be easily held back by chip or intellectual-property export controls, and should not be underestimated. He also wants the United States government to take the boring daily uses of AI seriously — clinical tools, faster government processes, even the Jetsons fantasy of self-driving cars and household robots. The argument on stage is not that the technology is imaginary.

What left him uneasy, he says candidly, is the frame. Too much of the analysis treats AI strategy as a contest between the United States and other nations. He would not want to live in Xi’s China, and he values the freedom to disagree with a sitting senator, but he worries that a habit of us-versus-them thinking produces bad decisions. Technology revolutions are a chance to rethink how people live and relate. In his view the true others of the AI era are not the Chinese. They are the AIs themselves.

Booker is tired of dystopian AGI talk and asks whether the fear is overblown. Wang takes the jump ball. There is some truth, he says, that very powerful systems may arrive, and also a lot of truth that today’s models are already beneficial. The job is a grounded, scientific conversation about what the algorithms can do, how those capabilities are developing, and when controls actually become necessary — not an emotional argument about science-fiction endings.

He points to a quieter institutional fact. The United States has stood up an AI Safety Institute inside NIST; the United Kingdom has one; Japan and Korea are starting their own. A global effort is forming around a precise question: how do you adjudicate whether a model is worthy of concern? Clark builds on that with an electricity analogy. If the room sat around regulating everything that plugs into an outlet, it would drive everyone crazy. What is needed is product-safety testing, a shared global framework, and a refusal to crush a young ecosystem in the name of caution.

Ng is more pointed about the last year of lobbying. Parties that spent billions training foundation models, he says, pushed with surprising intensity to trash open source and shut down innovation, while the stated safety worry kept shifting — takeover, then bioweapons, then cyber defense. AI has real risks. The regulatory mistake, in the United States, abroad, and in Europe, is trying to make home appliances safe by regulating the solar-panel makers and demanding that every generated electron be guaranteed harmless. He wants safe self-driving cars, reasonable underwriting software, and safe medical devices. Regulate the application layer.

Booker answers with an FAA story from his Commerce Committee days: Europe was already flying drones over mines and power lines while the United States issued almost no licenses except for film. If the FAA had existed in the Wright brothers’ time, he told its head, the airplane would never have left the ground. Clark’s carve-out is national security. You may found an AI company with no interest in that world, but that world is interested in you; tests should catch a system that could unwittingly help a non-state actor do harm. Beyond that, he would lean on the FDA, the FAA, and existing law rather than invent a new infrastructure. Ng then rejects the word dual-use. Nuclear technology and rockets can be contained without trashing civilian work. AI is general-purpose, with far more civilian uses than military ones, and should be treated as multi-use.

Pressed for the regulation they actually want from Congress, the panelists turn to defense. Wang says national security has the opposite problem: not enough AI, and not enough modernization. The Department of Defense is hamstrung by process built for an era when technology moved slowly, while near-peer adversaries move fast. Look at China or Russia, he says, and their willingness to fold new frontier technologies into every component of how they operate vastly outpaces what the United States can do inside the same bureaucratic confines.

People inside the Pentagon know this, and some are working it through legislation, appropriations, and internal practice. Wang’s greatest fear is not that Washington will fail to write the right rule preventing AI from being used in national security. It is that the Defense Department and the intelligence community will not use the technology enough to stay competitive. Booker jokes about swapping jobs with “Senator Clark.” Clark’s bill, he says, would aggressively field experimental technology and get departments to actually touch it.

Inside Anthropic, Clark reports, the more they find ways to use the systems, the more ways they find the systems can help. He wants a testing and measurement regime that asks whether the technology is working, how to fix it when it is not, and only then whether extra regulation is required. The greatest risk, he repeats, is not using it. Private industry is making itself faster and smarter by experimenting. If the nation fails to do the same, some other entrepreneurial country will.

Ng’s race metaphor is the cleanest line of the afternoon. You can try to trip the other runner, and maybe do a little of that, or you can run faster: spend on training, education, and upskilling, on compute infrastructure, and on a less sclerotic appropriations process. He adds a second-order risk. Many countries already buy surveillance software from America’s adversaries, which limits Washington’s ability to press for human rights and privacy. Stifling rules that shut American firms out of the supply chain would hand those customers to someone else. Someday, when a large language model is asked what it thinks of democracy, he would like the answer to reflect “our values rather than someone else’s.” That, he says, would be an own goal.

Booker puts the competition on the table. RAND, he notes, recently suggested China is unlikely to produce major new AI advances because of America’s private-sector edge. Wang’s answer is immediate: strong disagree. Some of the best open-source models in the world have been built by Chinese companies, which is one measure of how fast they can catch up. He then offers a “triple whammy” for why the moment is urgent. The technology is moving faster than anything in recent decades, in both money and raw scientific progress. Global conflict is rising, with no obvious off-ramp. And leaders such as Xi Jinping and Vladimir Putin are in late stages of rule, which in Wang’s telling makes them more aggressive, more risk-seeking, and more willing to make bolder moves. Those three together are a powder keg.

Clark has heard too much at the same event about China as a fast follower that only copies. In AI, he says, it is different. China has great inventors and great researchers. In response to export controls they have done “really fantastic work” on distributed training and on the low-level tricks that make training more efficient. “It would be a chronically stupid thing to underestimate their capacity for inventiveness.” The United States is not facing an uncreative rival. It is facing a competitor that will be just as creative, and even more motivated.

Ng wants democratic values to spread with the technology, and he thinks Washington is undermining its own ability to do that. Scaling giant models is one front; a major second front is agentic AI. Using ChatGPT or Claude to write an essay in one pass, he says, is like demanding the model never hit backspace. Agents iterate: outline, first draft, self-critique, web search, revise. The result is much better. Looking at the research literature, he sees probably as many cutting-edge innovations coming out of institutions such as Tsinghua University as out of many American universities. “The cat’s out of the bag.” When another country downloads an open-source package to implement its own agentic writing or journalism stack, the open question is whether the source is primarily a foreign supplier, or whether America still has a huge voice.

Clark offers a proposal that, he stresses, does not favor Anthropic, which funds its own training. He wants more competition for the large foundation-model companies. Startups and academic researchers still have to cross the chasm from papers into production. One way to use America’s inventiveness is to put the national supercomputing base under them — a public compute ladder rather than a closed club. Booker, a supporter of a national AI research effort, thanks him for arguing against his own firm’s narrow interest.

The senator then asks about Saudi Arabia and the United Arab Emirates, whose moves in this space he calls mind-boggling, and whose leaders he knows. Is that hopeful, or a values problem? Clark’s answer is about leverage. Infrastructure lets you build the systems and define the norms. Money has to go into doing this well at home. Picking up another panelist’s line, he jokes that who knew the path to AGI ran through American permit reform — but data centers, power, and domestic infrastructure are how you lead abroad.

Wang is unsentimental. Regardless of what the United States does, the Middle East and other countries will build gigantic data centers and huge amounts of AI compute. The American task is the same one it faced with the internet and telecom: make as much of the world’s incoming capacity line up with its values and its system. Ng reaches for Edison. After the electric grid was modernized, other countries got electricity too, and that was good — growth, prosperity, a better grid in more places. He thinks it would also be good if many countries had more intelligence, including artificial intelligence.

Booker closes by asking for one thing about 2055 that most people have not yet wrapped their minds around. Clark’s answer is education. Anyone will be able to learn anything: a customized plan and tutor that lets a curious person become an expert, which he thinks is already within reach of the technology. Wang’s answer is scientific breakthrough. Decades of using these systems, he says, will make the sci-fi futures people have been imagining slowly become ordinary life. He reaches for the same cartoon Labenz used in the intro: we will live in the Jetsons.

Ng talks about price. Today intelligence is one of the most expensive things in the world, which is why only the wealthy can hire a specialist doctor to look hard at a condition, or a patient tutor for a child. Artificial intelligence is making intelligence cheap. A few decades out, he thinks any of us will be able to hire an army of well-trained specialists for advice that only the richest can buy today. Booker prays the result is not a further concentration of wealth and power. Genius, he says, is equally distributed: as many geniuses are born per capita in Burkina Faso as in the most affluent parts of New Jersey. What societies do badly is cultivate it.

He adds a civics postscript. High-school history over-weighted Washingtons, Jeffersons, and Reagans; the unsung work was done by innovators and scientists who built systems later generations take for granted. Sitting with three front-line builders for twenty minutes, he says, leaves him humble. Ng does not take the compliment. “Don’t give us too much credit,” he says. “We really need the politicians to set the framework.” Booker lets that be the last word.