For six years I kept a physical folder in a Melbourne filing cabinet labeled "Signed Off Reluctantly." It holds eleven audit reports where I attached my name to code I did not fully trust — not because it was exploitable, but because the people deploying it had stripped out the checks that would have caught what they could not foresee. When I read that the Trump administration had "dismissed AI safety concerns," my first instinct was not political. It was archival. I recognized the sentence. I had written a variant of it years ago, in the language of Solidity, when a founder looked at a reentrancy guard and asked me why we needed it if nothing had gone wrong yet.
The answer, as anyone who lived through June 2016 understands, is that nothing has gone wrong yet precisely because the guardrail exists. Remove it and you have not discovered a durable system. You have merely postponed the discovery of its fragility — and, critically, you have transferred the cost of that discovery onto the people who arrived after you.
That is the lens I want to bring to this policy moment. Not partisan, not technophobic. Audited.
To understand what actually changed, one has to go back to October 2023, when the Biden administration signed Executive Order 14110. That document was, by any measure, the most stringent frontier-model governance framework in the world. Its central mechanism was deceptively simple: any foundation model trained above a compute threshold of 10^26 floating-point operations was required to report to the federal government and submit to safety testing, including red-team exercises designed to surface dangerous capabilities before deployment. The threshold was not arbitrary. It was calibrated to capture the class of models — GPT-4 scale and above — whose emergent behaviors cannot be reliably predicted from smaller training runs.
The Trump administration replaced this with a different philosophy, articulated under the banner of "Removing Barriers to American Leadership in AI." The operational content was the rescission of mandatory reporting and enforced red-teaming. The rhetorical content was the elevation of "maintaining American leadership" as the supreme objective, above and beyond any competing consideration, including safety. This was not a technical judgment. It was a frame-level switch — from a precautionary principle to what policy scholars call accelerationism.
The shift arrived bundled with personnel signals. The appointment of David Sacks as the administration's AI and crypto czar marked the ascendancy of the technology-optimist wing; the AI safety cohort that had populated the NIST AI Safety Institute and related State Department positions saw its influence collapse. Both the Brookings Institution and the Center for Security and Emerging Technology observed this pattern in early 2025, and it was visible internationally too. At the Paris AI Action Summit in February 2025, the American delegation refused to sign a joint statement containing language about "inclusive and sustainable AI," and publicly criticized what it described as European over-regulation.
I want to be precise about what I am and am not claiming. The coverage that triggered this reflection compressed all of it into a single headline — "Trump dismisses AI safety concerns" — and that compression is itself worth examining. Politicians almost never say the literal words. They say things like "we must balance innovation and safety" or "we cannot let excessive regulation kill American leadership." "Dismisses" is an editorial annotation of priority, not a quotation. But the direction it captures is real. The reporting obligations are gone. The red-teaming mandate is gone. What remains is a bet.
The bet is not "safety doesn't matter." It is "safety follows from markets, not mandates." That is the actual proposition, and it deserves to be examined on its merits rather than on its caricature. The accelerationist argument runs roughly: the United States is in a genuine strategic competition with China; that competition will be decided by who reaches frontier capabilities fastest; and any friction that slows American labs — including safety testing — cedes ground to rivals who will not observe those constraints anyway. In this view, safety caveats are unilateral disarmament.
The argument is not stupid. It is simply incomplete, and the incompleteness is where the danger lives.
To see why, one has to disaggregate "AI competition" into its actual constituent warfronts, because the policy tools that help on one front can actively harm another. There are three: the compute supply chain, governed by chip export controls; the model capability frontier, governed by lab-level racing; and standards and ecosystem influence, governed by which governance framework the world adopts as default.
On the first front, the policy is continuity, not change. Advanced chip export controls toward China have tightened continuously since October 2022, with repeated updates to the Bureau of Industry and Security rules, and the Biden administration's late-January 2025 AI Diffusion Rule established a tiered export framework. The Trump administration has kept the tool but adjusted its application — loosening toward allies, tightening toward China. The May 2025 chip agreements with Saudi Arabia and the UAE are the clearest expression of that adjustment. The safety deregulation does not touch this channel at all, and the fact that the headline collapsed both into one story is the first sign that the reporting was describing a mood rather than a policy.
On the second front, the picture is more genuinely alarming for the optimistic thesis. In January 2025, DeepSeek released R1, and with a single release it destabilized the assumption that compute is a moat. A Chinese team, operating under constraints the American labs do not face, produced reasoning capabilities approaching the frontier. That event did not merely embarrass some valuation models. It weakened the underlying policy logic that says "if we control the chips, we control the gap." When your opponent closes the gap under restriction, restriction stops being a strategy and starts being a tax on your own supply chain.
I have watched this dynamic before, in a much smaller arena. In 2017, I audited a series of ICO-era projects so consumed with speed that they treated security as a delay function. The argument was always the same: the market will punish bad code, so we do not need to delay our launch for audits. The market did eventually punish bad code. What it did not do was punish it before the founders had moved to the next project and the tokenholders had absorbed the loss. Speed, in the absence of accountability, does not distribute upside and downside symmetrically. It concentrates upside on the builders and downside on everyone who trusted them.
That is the asymmetry I keep returning to when I read the deregulation arguments. Markets are excellent at pricing risk after it materializes and terrible at pricing it before. The safety-reporting regime was never a guarantee against catastrophe. It was an attempt to shrink the interval between capability emergence and public knowledge of it — the same function, incidentally, that a decent audit performs. Remove that interval and you have not eliminated the risk. You have made it invisible for longer.
The third front — standards and ecosystem influence — is where deregulation is most counterintuitive, because it is the front on which the United States may be quietly losing while congratulating itself on the first two. China's open-weight model ecosystem — Qwen, DeepSeek, Kimi, GLM — has accumulated a substantial and growing share of downloads and derivative models on platforms like Hugging Face. This is not a marginal development. Open weights travel. They become the substrate that developers in the Global South, in Europe, and in American universities build upon. If the models that shape the world's tooling are governed by Beijing's preferences — even loosely — then the standard-setting war is being lost by the side that talks most about winning it.
And here is the piece the accelerationists rarely confront: safety, reliability, and controllability are not costs that sacrifice competitiveness. They are increasingly the product. Enterprises buying AI for regulated functions — clinical decision support, financial compliance, legal discovery — do not purchase raw capability. They purchase the ability to defend a decision to a regulator. A model that cannot explain itself, that hallucinates under pressure, or that ships without documented evaluation, is not a bargain at any price for a hospital or a bank. The premiums in this market are paid for trust, and trust is manufactured by exactly the kind of rigorous, boring, unglamorous work that a deregulated environment de-prioritizes.
I saw this from an unusual angle in 2024, when I advised a large Australian pension fund on its first crypto allocation. The trustees were not asking me whether the assets would go up. They were asking who would be accountable when something went wrong. I negotiated a clause directing five percent of the allocation toward open-source infrastructure, which was criticized at the time as unorthodox. But the clause existed because the trustees understood something the crypto-native crowd did not: institutional capital does not buy returns, it buys defensibility. The same logic is arriving at AI procurement. Safety is the price of admission to the parts of the economy that matter.
Watch how capital is already interpreting this. The equity transmission chain from AI deregulation runs, in order of conviction, through compute hardware, data-center infrastructure, and electrical generation. Names tied to accelerators, cooling, networking, and utility-scale power have all traded on the narrative of unrestrained build-out. The logic is not irrational — deregulation plus arms-race dynamics do raise the probability of a sustained capital expenditure cycle. But notice what is missing from that chain: any pricing of the tail risk that accompanies a witness-free training environment. The market has assigned enormous value to the upside of acceleration and approximately nothing to the probability that a serious incident triggers a regulatory pendulum swing back. That asymmetry — cheap insurance, expensive conviction — is where I would look if I were auditing the sector rather than celebrating it.
There is a second, more mechanical story that the headline obscures, and it concerns where the actual constraints live. If you want to understand American AI policy's effect on AI development, you should ignore the safety rhetoric and follow the electricity.
The three binding constraints on frontier AI in 2025 are not regulation, tariffs, or even capital. They are, in order: training compute cost, which rises exponentially per generation; inference cost, which scales with every user; and electrical power plus data-center construction lead time, which now runs two to four years. The last of these has quietly become the hardest. Several regions in the United States have interconnection queues exceeding three years, and hyperscalers are signing long-term nuclear and geothermal power agreements precisely because grid capacity cannot be summoned on demand.
The deregulation of safety reporting touches none of these constraints directly. Its effect is indirect and second-order: by removing the 10^26 FLOPs reporting threshold, it lowers the compliance friction on the very largest training runs. That may accelerate the launch rhythm of next-generation models. But it does not build a single megawatt of generation, does not shorten a single permitting queue, and does not manufacture a single HBM stack. The compute arms race is racing against physics, and physics does not read executive orders.
What this means is that the deregulation's most significant effect is on the timing of risk, not the magnitude of capability. Removing the reporting threshold may compress the interval between a frontier model's training and its deployment. It does not change how much compute exists or how fast models improve. It changes how long the world has to notice what has been built.
For anyone who has audited systems, this is the familiar and most dangerous variable. The failure mode almost never lies in the code that runs correctly. It lies in the code that runs correctly and silently until the day it does not, and no one has been watching. A reporting requirement is, at bottom, a witness. It ensures someone outside the builder's confidence is present. Remove the witness and the building does not fall. It just falls without a warning.
I have written about this before, in a private manifesto that was never meant for publication. In 2022, after FTX collapsed and my own six years of optimism curdled into something closer to grief, I withdrew to the Victorian bushlands and wrote an essay called "The Myopia of Decentralization." Its thesis was uncomfortable for the people who had once called me a believer: that the decentralization movement's deepest failure was not technical but epistemic. We had confused the absence of a single point of failure with the absence of failure itself. We had built systems that removed the trusted intermediary and forgot to ask who would answer when the untrusted many coordinated to do harm.
That manifesto leaked, and it cost me some friendships in the community. But I have never written anything I believe more completely. And I see its exact shape in the AI accelerationist posture. The belief that removing centralized oversight produces a freer, more robust system is, in both cases, a category error. What removing oversight actually produces is a system in which the entities with the most resources — the best lawyers, the largest compute budget, the deepest bench of engineers — absorb risk best, and everyone else absorbs it worst. Deregulation is not the diffusion of power. It is the concentration of a different kind of power: the power to be reckless without consequence.
The accelerationists have a response, and it is not without force: that in a genuine strategic competition, the alternative to American recklessness is not global caution but someone else's recklessness, combined with their dominance. This is a real argument. I do not dismiss it. But it conflates two different questions: what is the best way to win, and what is the best way to win without becoming the thing you are trying to outperform.
Here is the contrarian turn, and it may be the most useful thing I can offer. The AI safety camp has been arguing its case badly, and the deregulation window is the consequence. For years, the dominant safety rhetoric has been apocalyptic — AI as existential risk, alignment as humanity's last problem. That rhetoric works on people who already agree and repels everyone else. To a legislator or an executive in a competitive system, existential risk is an abstraction they cannot act on, and a decade of appeals to it have produced fatigue, not caution.
The genuinely persuasive case for safety infrastructure is not catastrophe. It is liability, and it is defensive capability. Who is responsible when a model is misused at scale? Who pays when a financial system's automated decisions cascade? What evidence does a company need to survive its first serious incident? Those questions are answerable, actionable, and concrete. They are also questions that institutions — insurers, boards, pension funds — are already asking without waiting for Washington. The market is developing its own safety infrastructure precisely because the state has withdrawn, and it is doing so for reasons more durable than regulation: self-interest.
That is the pragmatism test. If your only argument for guardrails is "otherwise everyone dies," then when the guardrails come down, nothing changes about behavior — the doomer is dismissed as a doomer, and the market proceeds. If your argument is "otherwise you cannot insure it, audit it, or sell it to anyone with a fiduciary duty," then the guardrails survive deregulation because they migrate from statute to contract. My wager — and I have made a career of this wager — is that the second argument is the one that actually protects people, because it survives the politics.
Fifteen months separated the signing of EO 14110 and its rescission. That interval is the clearest evidence of what the AI safety community has been slow to accept: that a safety framework resting on executive discretion is not a framework at all. It is a mood. The institutions that will actually hold the line are the ones with reasons of their own — the insurers who cannot underwrite the unaccountable, the enterprises that cannot deploy the unexplainable, the pension funds that cannot defend the indefensible.
The guardrail is coming down. The question that should occupy us now is not whether it returns at the federal level. It should be this: when the next frontier model is trained, will anyone outside the builder's confidence be watching? Or will we learn what was built the way we have so often learned it before — from the wreckage, and from the people who absorbed a cost they never agreed to bear?
