Fight Open Source with Open Source

Dario Amodei proposed something unusual for Silicon Valley: slow down.

His argument is not that artificial intelligence should stop. It is that frontier capabilities are advancing faster than our ability to understand and secure them, and that we should deliberately create more time for safety to catch up.

He proposes three broad steps: independent evaluators embedded inside frontier AI companies, coordination among AI companies in democratic countries, and eventually international coordination, most importantly between the United States and China.

I agree with the problem more than I agree with the prescription.

AI safety matters. Independent evaluation matters. Responsible development matters.

But legitimate concerns about AI safety should not become a strategy that unintentionally slows American innovation, concentrates AI in a handful of companies, or assumes that technological competition between the United States and China can simply be negotiated away.

We will get to more powerful AI.

The question is how we get there—and which country builds the ecosystem around it.

Start With What We Can Do Tomorrow

The strongest part of Amodei’s proposal is also the simplest: independent evaluation.

Do it.

Anthropic has committed to giving independent evaluators employee-like access to examine its safety practices, training processes and incidents. Other frontier laboratories should seriously consider doing the same.

This does not require China.

It does not require an international treaty.

It does not even require waiting for Congress.

A company that believes independent evaluation makes its models safer can invite independent evaluators inside tomorrow.

There is something powerful about companies voluntarily demonstrating that safety and technological progress do not have to be enemies.

If these systems work, governments can eventually establish durable standards around them.

The second proposal—coordination among frontier companies in democratic countries—is more complicated.

America is a market economy. Competitors coordinating how quickly they develop technology raises legitimate questions about antitrust law, enforcement, standards and government authority. Amodei himself acknowledges that some forms of coordination would require government support.

And government policy changes.

Administrations change. Congress changes. Courts intervene. Companies enter and leave markets.

If America decides that some form of coordinated pacing is necessary, the framework must be transparent, legally durable and democratically accountable. This is too important to depend indefinitely on voluntary agreements among a few CEOs.

But the third proposal is where the problem becomes much larger.

The China Problem

Let’s be realistic about the AI competition.

Many countries are doing important AI research, but the two dominant technological powers are the United States and China.

China should never be underestimated.

It has enormous engineering talent, industrial capacity, capital and increasingly sophisticated domestic technology. MacroPolo’s AI Talent Tracker found that researchers originating from China accounted for 47 percent of the world’s top-tier AI researchers in its 2022 dataset. At the same time, the United States remained the leading destination for elite AI talent.

That combination should tell America something important.

China has the people to compete.

And China has made technological self-reliance and leadership strategic national objectives.

The political systems are also fundamentally different.

China is an authoritarian state with a highly centralized political structure. When Beijing identifies a technology as strategically important, it can coordinate national policy, infrastructure, financing, universities and industry in ways that are much harder in the United States.

America is a democracy and a decentralized market economy. Congress debates. Courts intervene. States disagree. Companies compete. Citizens object. Administrations change.

That friction is part of America.

So America cannot simply imitate China’s strategy.

It has to be more strategic.

International dialogue is still worthwhile. The United States and China should communicate about catastrophic risks, military applications, autonomous systems and other areas where misunderstanding could be dangerous.

But dialogue is different from dependence.

An American AI strategy cannot depend on the assumption that every restriction will be interpreted identically, implemented identically and honored indefinitely by every participant.

Amodei recognizes this problem himself. He writes that a global pacing agreement would require extremely strong verification because the incentive to defect could be enormous.

That is precisely the problem.

Before asking America to substantially slow technological development, we need to know what happens if someone else does not.

Containment Is Not a Complete AI Strategy

Export controls can make access to advanced computing more difficult and expensive. They can buy time and constrain capacity.

But they are not a complete technological strategy.

China has enormous incentives to develop domestic chips, improve software efficiency and find alternative architectures precisely because access to American technology is constrained.

The same applies to model distillation and open weights.

Once powerful models become software that can be downloaded, modified, fine-tuned and deployed around the world, containment becomes extraordinarily complicated.

And trying to restrict foreign open models after developers and enterprises have integrated them into products can create another problem: America may end up weakening its own ecosystem.

The better long-term question is not simply:

How do we prevent developers from using Chinese AI?

It is:

Why aren’t we giving them an American alternative they prefer?

That is the strategic opening.

Fight Open Source with Open Source

Don’t fight open-source AI by trying to wish it away.

Fight open source with open source.

Make American open AI something developers around the world want to build on.

Give developers models they can download, modify, fine-tune and deploy themselves. Give startups an alternative to paying frontier API prices for every token they generate. Give enterprises the ability to keep sensitive workloads inside their own infrastructure.

Give researchers access.

Give universities access.

Let a college student fine-tune a model.

Let a five-person startup build a company around one.

Let an enterprise run one behind its firewall.

Let researchers take one apart.

Let millions of developers discover applications that no frontier laboratory could possibly anticipate.

This is not an argument against frontier AI.

It is an argument for both.

Anthropic should thrive.

OpenAI should thrive.

Google should thrive.

American open-source and open-weight AI should thrive too.

One pushes the frontier of intelligence.

The other distributes intelligence.

America benefits from both.

We should not weaken frontier laboratories to protect open source, and we should not weaken open source to protect the economics of frontier laboratories.

Let them compete.

The evidence already points in this direction. Stanford’s 2025 AI Index reported that the gap between leading open-weight and closed models narrowed sharply during 2024, while the cost of using capable models continued to fall. That means open models are not merely a philosophical alternative. They are becoming a practical competitive instrument.

The United States government has recognized the same strategic logic. The U.S. AI Action Plan identifies leading American open-weight models as having geostrategic value and calls for supporting their adoption, particularly by startups, researchers and organizations that cannot send sensitive information to closed providers.

That is exactly the opportunity.

The objective should not simply be to have the world’s smartest model sitting behind an American API.

The objective should be to have the world building with American AI.

Affordability Is a Strategic Issue

There is another reality that gets overlooked in debates about frontier models: economics.

The most advanced AI systems are extraordinarily expensive to build and operate. Frontier APIs make that intelligence accessible without requiring customers to own the infrastructure, but using them at scale can still become expensive.

For many startups, developers, universities, small businesses and organizations around the world, cost matters enormously.

Open models change that equation.

Organizations can choose their infrastructure. They can optimize inference. They can fine-tune models for specific workloads. They can run models locally when privacy requires it.

That is not merely a technical preference.

It affects adoption.

If America wants its AI ecosystem to become the world’s default ecosystem, affordability matters.

Make American intelligence powerful.

But also make it accessible.

The country whose technology millions of developers can afford to experiment with gains something that cannot easily be purchased later: an ecosystem.

And ecosystems create strategic advantages that individual models cannot. Developers build skills around them. Startups form around them. Universities teach them. Enterprises integrate them. New tools and standards emerge from them.

Once that network is established, it becomes difficult for a rival to displace.

Don’t Lose the Public

There is another constituency the AI industry cannot afford to ignore: ordinary Americans.

Gallup found in 2025 that seven in ten Americans opposed construction of AI data centers in their local area, including 48 percent who strongly opposed them.

People have legitimate concerns about electricity, water, land, environmental effects and local costs.

Now consider the message the public sometimes hears from the AI industry.

We need enormous amounts of electricity.

We need enormous data centers.

We need them quickly.

And the technology they power might someday destroy humanity.

Those messages do not fit comfortably together.

Researchers should investigate catastrophic AI risks. Companies should red-team powerful systems. Independent evaluators should test them aggressively. Governments should prepare for credible threats.

But serious safety research and apocalyptic public messaging are not the same thing.

Repeatedly telling people that AI might “kill us all” risks doing a disservice to the technology, particularly when the industry simultaneously needs public support for enormous infrastructure expansion.

People need to see why AI is worth building.

Show them scientific discoveries.

Show them better medicine.

Show them new companies.

Show them productivity.

Show them education.

Show them what an individual developer can create with capabilities that previously belonged only to enormous corporations.

And perhaps most importantly, let people participate.

AI should not feel like something being built behind closed doors by five companies while everyone else is asked to accept the consequences.

Open models can help change that relationship. They give developers, researchers and businesses a stake in the technology rather than asking them merely to consume it.

America’s Advantage Is America

China can coordinate from the top.

America can innovate from everywhere.

That difference should be treated as an advantage, not a weakness.

America has extraordinary universities, capital markets, entrepreneurs, researchers, immigrants, chip companies, cloud providers, frontier laboratories, startups and one of the world’s largest developer communities.

Use all of it.

Build the world’s best frontier models.

Build the world’s best open models.

Build the chips.

Build the infrastructure.

Make inference cheaper.

Invest in safety.

Invite independent evaluation.

Attract the world’s best scientists.

Give startups room to experiment.

Give enterprises control.

And give developers the freedom to build.

Never underestimate China’s technological ambition. Never underestimate the resources and talent it can bring to this competition.

But America’s answer should not be to become more like China.

It should be to become more American.

Open competition has produced extraordinary American technologies before. The internet, personal computing and the modern software industry all became more powerful because innovation spread beyond a small number of institutions.

AI should not be different.

The United States should not measure success only by whether an American company trains the most capable model. It should measure success by whether developers, researchers, startups and enterprises around the world choose American technology as the foundation for what they build next.

Fight Open Source with Open Source

AI development will not stop because one company slows down.

It will not stop because one country regulates it.

And it is increasingly difficult to imagine that knowledge this economically and strategically valuable will simply disappear.

We will get there.

That does not mean racing recklessly.

It means recognizing that safety and progress are not opposites.

Independent evaluation can coexist with rapid innovation.

Frontier models can coexist with open models.

Closed commercial systems can coexist with locally deployed intelligence.

Safety can coexist with competition.

Anthropic should succeed.

American open source should succeed.

Thousands of AI startups we have not heard of yet should succeed.

And America should create an environment where all of them can.

Because this competition is larger than any company’s valuation, any single model release or any benchmark.

It is about which technological ecosystem the world chooses to build upon.

The wrong response to open-source AI is to treat it as a threat that must be contained.

The right response is to build something better, cheaper, safer and more useful.

Fight open source with open source.

Fight competition with competition.

Make AI safer without making innovation inaccessible.

Win the developers.

Win the enterprises.

Win the researchers.

Win the public.

And let the world build with American AI.

We will get there—but America should make sure the world gets there with us.