Source: https://joinfinn.com/blog/we-absolutely-must-not-pace-the-frontier

# We absolutely must NOT pace the frontier

Finn’s case against pacing the AI frontier: continued innovation, stronger defenses, wider access, and accountable development.

By Finn · 7 min read

Published September 13, 2026

Imagine you are responsible for a hospital’s computer systems. Attackers are becoming faster at finding weaknesses, impersonating employees, and coordinating intrusions. Your team needs better tools to understand what is happening and respond before patient care is disrupted.

Now imagine being told that the institutions most willing to accept scrutiny should deliberately slow their development of more capable AI, while you have no reliable assurance that the people attacking you will exercise the same restraint.

From your side of that decision, slower progress might look like a growing gap between the threats you face and the tools available to protect people.

That possibility deserves a central place in the debate over AI safety.

In *We Must Pace the Frontier*, Anthropic CEO Dario Amodei argues that capability development should slow enough for safeguards to catch up. He proposes embedded external evaluators and coordination among companies and governments. He is explicit that pacing need not mean a complete halt. His concerns deserve serious engagement. At Finn, we disagree with making coordinated restraint on frontier advancement the organizing principle of AI policy. [Read Amodei’s argument](https://darioamodei.com/post/we-must-pace-the-frontier).

**We believe the safer and more prosperous future requires continued advancement, wider access to useful intelligence, and much stronger accountability for how it is developed and used.**

The first reason is strategic. Our judgment is unequivocal: **the United States’ adversaries and peer competitors will absolutely not pace their frontier simply because American companies decide to pace theirs.**

They have their own security interests, industrial ambitions, and incentives to close any technological gap. An American commitment changes none of those incentives automatically. It may give competitors additional time to improve their position.

China’s published AI Plus policy calls for stronger models, more computing capacity, an expanded open-source ecosystem, and extensive adoption across its economy. Those are announced ambitions, not proof of future performance. They nevertheless give us a firmer basis for planning than an expectation that Beijing will voluntarily match a slowdown proposed in Silicon Valley. [China’s published policy](https://english.www.gov.cn/policies/latestreleases/202508/27/content_WS68ae7976c6d0868f4e8f51a0.html).

A genuinely reciprocal, verifiable agreement would require separate evaluation. Until then, national strategy must work even when competitors keep going. A company can change its own research schedule. It cannot set the world’s clock.

Palantir’s Alex Karp made this strategic argument in his 2023 essay on AI: adversaries would continue developing consequential technologies, and hesitation could carry its own security costs. He also called for technical controls around AI’s interaction with critical systems. Capability and control belong in the same strategy. [Karp’s essay](https://www.palantir.com/assets/xrfr7uokpv1b/1wtb4LWF7XIuJisnMwH0XW/dc37fdda646a5df6c5b86f695ce990c0/NYT_-_Our_Oppenheimer_Moment-_The_Creation_of_A.I._Weapons.pdf).

Amodei recognizes the danger of competitors pulling ahead. But a technological lead is not a fixed balance that policymakers can confidently spend down. A better algorithm, a more efficient training method, or a new distribution channel can alter the competitive picture. A pacing strategy needs to show how it will remain safe when estimates of that lead are wrong.

The second reason is that defensive capability matters as much as offensive capability.

An AI system that helps discover software vulnerabilities creates risks. A system that helps defenders identify, verify, and repair those vulnerabilities can reduce them. Policy has to examine both effects, including who can actually obtain and use the tools.

NVIDIA makes this case in its Open Secure AI Alliance announcement. It argues for capable tools that defenders can inspect, adapt, and operate, and warns that sweeping restrictions could weaken defense while concentrating dependence on a few providers. [NVIDIA’s security argument](https://blogs.nvidia.com/blog/open-secure-ai-alliance/).

That reasoning should change how we assess a slowdown. The relevant comparison is between complete outcomes: attacks enabled, defenses strengthened, failures prevented, and vulnerabilities left unresolved. Counting the dangers of new capability while treating delayed defensive improvements as costless would produce a distorted answer.

Open access also carries risks. Some releases may warrant restrictions based on their specific capabilities. But the burden of proof should apply to withholding defensive capacity too. Hospitals, utilities, universities, and smaller businesses should not disappear from the calculation because their needs are less visible than a frontier laboratory’s research agenda.

The third reason is economic: **delay has a cost, even when no single person receives the invoice.**

Consider a small manufacturer that cannot afford a large software team, a researcher working through a difficult body of evidence, or an operations team spending its day reconciling records. More capable and affordable tools could expand what each can accomplish.

These benefits are uncertain, and they will not arrive merely because a benchmark improves. They require implementation, investment, and capable people. But uncertainty does not justify assigning them a value of zero. A responsible decision must compare the expected costs of proceeding with the expected costs of waiting.

NVIDIA’s 2025 criticism of the AI Diffusion Rule advances a related economic argument: American strength grows through innovation, competition, and the widespread adoption of its technology. That statement concerned export policy, rather than Amodei’s essay. The underlying lesson is relevant here: leadership becomes more durable when an ecosystem of businesses, researchers, and developers can build on it. [NVIDIA’s policy statement](https://blogs.nvidia.com/blog/ai-policy/).

We should be trying to make useful intelligence more affordable and more widely available. A frontier that advances while its benefits spread can create opportunities for people who were excluded by the cost of expertise. Delaying those opportunities needs an explicit justification.

There is a fourth concern: the structure of the market that pacing could leave behind.

Established laboratories already possess capital, infrastructure, customers, and influence. If restrictions on further development are designed around their budgets and operating practices, challengers could face barriers that incumbents can comfortably absorb. Existing providers could keep collecting revenue while new competitors struggle to qualify for permission to compete.

That outcome does not require bad faith. It requires only rules whose costs fall unevenly. Any pacing proposal should therefore explain its effects on entry, independent research, customer choice, and the ability to switch providers. Safety requirements should be proportionate to demonstrated risks and applied consistently, regardless of a company’s size or political access.

The strongest objection to our position is more serious than any commercial argument: what if advanced AI creates a danger that cannot be contained, including during training or internal experimentation?

That possibility cannot be answered by saying that competitors are moving quickly. It requires evidence, independent investigation, and enforceable constraints on dangerous work. A failed containment test can justify stopping a particular experiment. An unresolved, demonstrated hazard can justify withholding a particular system. Continued progress should never be an excuse to ignore either.

But those decisions do not automatically establish that the whole frontier should advance more slowly. Other research paths, defensive tools, and useful applications may still deserve to move ahead. Restrictions need a defined target, an explanation of how they reduce risk, and conditions under which they can be lifted.

Amodei’s proposed capability checkpoints and outside scrutiny could contribute to that work. Our disagreement concerns the broader objective. We want safety obligations to produce demonstrably safer systems, while preserving every avenue of beneficial progress that can meet those obligations. Slower aggregate advancement is an inadequate measure of success.

For Finn, this leads to a practical agenda. Fund independent evaluations and alignment research. Invest in secure computing infrastructure and defensive tools. Require meaningful incident reporting. Make developers responsible for containment and operators responsible for the authority they grant. Establish clear consequences for exposing people to unjustified risks.

Then make it easier for businesses, researchers, and public institutions to put capable systems to useful work. Procurement, training, affordable computing, and reliable access deserve the same urgency as rulemaking.

Our interest is straightforward: we build software intended to help people get work done. We want better intelligence to become more useful, more dependable, and more accessible. That requires continued invention and the discipline to prove that systems are ready for the responsibilities they receive.

The hospital in our opening example still needs protection. Its security team needs tools it can understand, deploy, and trust. A successful AI policy should leave that team better equipped against the threats it actually faces.

**We absolutely must NOT pace the frontier. We must advance it with the institutions, defenses, and accountability needed to put its power to work.**

## Further reading

- [Dario Amodei: We Must Pace the Frontier](https://darioamodei.com/post/we-must-pace-the-frontier)
- [State Council: China issues guideline to accelerate AI Plus integration](https://english.www.gov.cn/policies/latestreleases/202508/27/content_WS68ae7976c6d0868f4e8f51a0.html)
- [Alex Karp: Our Oppenheimer Moment — The Creation of A.I. Weapons](https://www.palantir.com/assets/xrfr7uokpv1b/1wtb4LWF7XIuJisnMwH0XW/dc37fdda646a5df6c5b86f695ce990c0/NYT_-_Our_Oppenheimer_Moment-_The_Creation_of_A.I._Weapons.pdf)
- [NVIDIA: Open Secure AI Alliance for AI Safety and Security](https://blogs.nvidia.com/blog/open-secure-ai-alliance/)
- [NVIDIA: Statement on the AI Diffusion Rule](https://blogs.nvidia.com/blog/ai-policy/)

[Explore Finn AI OS](https://joinfinn.com/ai-os)

[Discuss your workflow](https://joinfinn.com/contact) · [How it works](https://joinfinn.com/demo/app/#/w/borough/overview) · [Get started](https://joinfinn.com/app/#/signup)