The narrative currently being spun by Treasury Secretary Scott Bessent and the D.C. establishment is dangerously seductive. They claim we are talking to China about AI safety because we are in the lead. They suggest that from a position of dominance, we can dictate the ethical guardrails of the most transformative technology in human history.
It is a lie. Or at the very least, a catastrophic misunderstanding of how power and technology actually interact. Learn more on a connected subject: this related article.
Talking to a competitor about "safety" when you believe you are winning is not leadership. It is an invitation for them to catch up while you tie your own hands with red tape. If the United States thinks it can slow-walk the development of Artificial General Intelligence (AGI) through bilateral talks without handing the keys to the kingdom to the CCP, it hasn't been paying attention to the last thirty years of industrial history.
The Myth of the Magnanimous Leader
The "we are in the lead" argument assumes that technology is a race with a finish line. It isn't. It is an ecosystem of relentless iteration. Bessent’s rhetoric mirrors the flawed logic of the 1990s when Western leaders thought bringing China into the World Trade Organization would "democratize" their economy. We saw how that worked out for the American manufacturing base. Additional reporting by Engadget explores related views on the subject.
By engaging in high-level dialogues about AI risks, we are essentially sharing our threat models. When you tell an adversary what you are afraid of, you aren't "aligning" with them. You are giving them a roadmap of your vulnerabilities.
I have watched boards of directors make this same mistake for a decade. They get a head start, get terrified of the "risks," spend millions on consultants to tell them how to be "ethical," and then watch as a leaner, meaner competitor eats their market share by ignoring the very rules the leader tried to establish.
Safety is the New Protectionism
Let’s be precise about what "AI Safety" actually means in a geopolitical context. To the U.S. government, it means preventing a rogue model from crashing the power grid or generating a bio-weapon. To China, "safety" means ensuring the AI never questions the central party.
These are not the same thing. They are diametrically opposed.
When we sit at a table to discuss "safety standards," we are engaging in a process of mutual slowing. But China’s version of slowing is selective. They will happily agree to Western-style "ethics" boards for export-grade models while their internal military applications remain completely unencumbered.
Imagine a scenario where the U.S. pauses a $100 billion training run because of a perceived "hallucination risk" in social nuances, while a lab in Shenzhen continues to train a model on stolen intellectual property and unrestricted data sets. That isn't a safety win. That is a strategic surrender.
The Compute Fallacy
The "lead" Bessent refers to is largely based on hardware—specifically Nvidia’s H100s and B200s and the export controls preventing them from reaching Chinese shores. This is a fragile, temporary advantage.
- The Smuggling Economy: History proves that hardware embargoes are sieves, not walls.
- Algorithmic Efficiency: When you have less compute, you innovate on the math. China is becoming world-class at doing more with less.
- The Energy Wall: The U.S. is currently hitting a massive bottleneck in power generation. Building data centers in Northern Virginia is becoming a decade-long regulatory nightmare. China builds nuclear plants like we build Starbucks.
If we stop to talk, we are giving them time to solve the compute gap through sheer engineering grit and infrastructure speed. Our "lead" in LLMs (Large Language Models) is currently measured in months, not years.
The Sovereignty of the Weights
People often ask: "Shouldn't we have global standards for something as dangerous as AI?"
The premise is flawed. Global standards only work when there is a shared definition of the good life. There is no such shared definition between the Silicon Valley techno-optimist and the Beijing central planner.
By pushing for "global alignment," we risk creating a "Cartel of AI" where a few large players (OpenAI, Google, Anthropic) and a few large governments (US, China, EU) decide what ideas are allowed to be processed by silicon. This doesn't make the world safer; it makes it more brittle. It creates a single point of failure.
If an AI model in 2027 becomes capable of discovering new materials or curing diseases, the nation that owns those weights owns the future. You don't "discuss" the terms of that ownership with your primary rival. You build. You deploy. You win.
The Cost of the "Seat at the Table"
The diplomatic class loves the phrase "a seat at the table." But in the world of deep tech, if you are at the table and you aren't the one setting the price, you are likely the product being sold.
The U.S. government’s obsession with bilateral talks is a symptom of a deeper malaise: the belief that regulation can substitute for raw capability. We are attempting to use the tools of the 20th-century statecraft—treaties, summits, communiqués—to manage a 21st-century intelligence explosion. It is like trying to contain a hurricane with a picket fence.
I've seen venture-backed startups die because they spent too much time worrying about what the incumbent would do if they succeeded. The U.S. is acting like a nervous incumbent. We are so afraid of what happens if we win that we are making it easier for us to lose.
Stop Asking the Wrong Questions
The media asks: "How can we ensure China uses AI responsibly?"
The real question is: "How do we ensure the first AGI is built in a jurisdiction that respects individual liberty?"
You don't get there through a dialogue with the CCP. You get there by:
- Deregulating Power: Let companies build small modular reactors (SMRs) next to data centers tomorrow, not in 2035.
- Immigration Reform: Grant a green card to every PhD in physics, math, or CS who wants to leave China and work here.
- Embracing Open Source: The best defense against a centralized, authoritarian AI is a decentralized, robust open-source ecosystem that the government can’t switch off.
The Transparency Trap
Bessent and others suggest that transparency between the two superpowers will lower the "existential risk." This is a fundamental misunderstanding of game theory. In a zero-sum race for a dominant technology, transparency is just another word for industrial espionage.
Every time we share a "safety benchmark" with Chinese researchers, we are telling them exactly how we test our models. We are giving them the "answer key" to our most advanced systems. They can then build models that pass our safety tests while retaining the capabilities we are trying to restrict.
We are currently in a period of "Pre-AGI Phoney War." The stakes are everything. The idea that we can manage this through polite conversation is not just naive—it is a dereliction of duty.
The "lead" we have is not a cushion to rest on. It is a narrow window of opportunity. Every minute spent in a conference room in Geneva or Beijing discussing "AI ethics" with people who see the world as a zero-sum struggle for dominance is a minute we aren't spending on the engineering problems that will actually decide the future.
The only way to ensure AI safety is to ensure that the most powerful models are controlled by those who value human agency over state control. You don't negotiate for that. You out-innovate for it.
Stop talking. Start building.
The dialogue is a distraction. The lead is an illusion. The race is the only thing that is real.
