There is a particular silence that follows a decision made at the speed of politics. In September 2025, when the former President dismissed the collective call from three of America's most powerful artificial intelligence laboratories to slow the pace of capability advancement, what reverberated was not merely a policy position. It was the sound of a civilization choosing its own speed—a choice made without ceremony, without the philosophical weight such a decision deserves, and with an almost casual certainty that history would vindicate the acceleration.
The laboratories in question—Anthropic, OpenAI, and xAI—had each, in their own way, articulated concerns about the trajectory of what they were building. The media would later compress these nuanced positions into a unified plea: slow down. But code is law, and narrative is truth, and the truth of what happened in those September chambers was far more complex than any headline could contain. The question I find myself returning to is not whether acceleration or deceleration is correct—that is a question for economists and ethicists—but rather: what does this moment tell us about the stories we tell ourselves regarding power, accountability, and the nature of technological progress?
I have spent the better part of a decade studying how narratives function within technological ecosystems. I have audited smart contracts that promised decentralization and delivered consolidation. I have watched yield-farming protocols that spoke of community ownership collapse under the weight of their own internal contradictions. And I have learned, through the slow accumulation of evidence, that the most dangerous moment in any technological transition is not when the technology fails, but when the narrative surrounding it becomes untethered from reality. What unfolded in September 2025 feels, to me, like precisely such a moment.
To understand what this policy pivot means, we must first understand the landscape it disrupted. For the preceding two years, the conversation around frontier AI development had been dominated by a peculiar form of institutional theater. The leading laboratories—those same three whose names now appear in regulatory filings and congressional testimony—had begun to articulate what they called "responsible scaling" frameworks. These were documents of genuine intellectual substance, proposing voluntary commitments to safety evaluations, capability thresholds, and pre-deployment assessments. They represented, in essence, an attempt at industry self-governance at a moment when formal regulatory frameworks remained years behind the technology itself.
The irony of this arrangement was not lost on those of us who had spent time in the DeFi trenches. We had seen this movie before. In decentralized finance, voluntary security practices had proven insufficient to prevent the systematic extraction of value from protocols that preached autonomy while depending on centralized backdoors. The same moral hazard that infected yield farming—where the promise of sustainable returns obscured the Ponzi dynamics underneath—now seemed to be replicating itself in the AI governance discourse. The laboratories that called for slower development were, by and large, the same laboratories that possessed the most to gain from maintaining competitive advantage. When Anthropic published its responsible scaling policy in 2024, the document was simultaneously a genuine contribution to safety discourse and a strategic positioning maneuver. This is not cynicism; it is pattern recognition.
What the laboratories were actually proposing, when one stripped away the press releases and carefully worded statements, was not a halt to progress but a managed deceleration—a slowing of the perceived acceleration, not necessarily the actual one. The distinction matters enormously, yet it was systematically collapsed in the coverage that followed. Liquidity flows, but trust evaporates, and the trust that was being asked for in those calls for "pacing" was trust that the laboratories themselves had spent years undermining through a pattern of broken commitments, moving goalposts, and what I can only describe as ethical drift.
The administration's response, however, represented its own form of narrative manipulation. By framing the safety calls as a "negative force" that "should never have been raised," the political calculus shifted from evaluating the substance of the concerns to questioning the legitimacy of the questioners. This is a familiar playbook, one that I have observed repeatedly in the blockchain space whenever developers or researchers raised concerns about protocol security or token economics. The ad hominem framing—what we might call the "whose interests are they serving" attack—does not address the underlying technical or ethical question. It changes the subject. And in the September moment, it succeeded brilliantly. The conversation moved from "should frontier AI development be paused" to "are these companies trying to slow American competitiveness," and the latter question was far easier to answer with patriotic certainty.
The technical implications of this policy pivot are substantial, even if they remain largely invisible to the public discourse. When we speak of "accelerating AI development," we are not speaking of a single variable but of an entire ecosystem of interdependent factors: training compute, data quality, algorithmic efficiency, deployment infrastructure, and—critically—the regulatory environment that governs each of these components. The decision to reject safety pauses is not merely a statement about intent; it is a signal that the United States will not impose meaningful constraints on any of these variables in the name of safety. This has direct consequences for the architecture of frontier models, the deployment timelines of new capabilities, and the relative priority of safety research within these organizations.
From a structural perspective, what we are witnessing is the institutionalization of what I would call "moral hazard accelerationism." The laboratories have been given explicit political cover to pursue capability expansion without meaningful safety constraints, with the implicit promise that any consequences of this expansion will be externalized onto the public. This is the same structure we observed in the 2020 DeFi summer, when yield farming protocols offered double-digit percentage returns that were structurally impossible without continuous new capital inflows. The protocol operators knew this. The early investors knew this. The political and regulatory environment at the time—characterized by aggressive innovation advocacy and minimal enforcement—created the conditions for the inevitable correction. The question was never whether the correction would come, but when, and who would bear its costs.
The analogy extends further than is comfortable. In DeFi, the moral hazard emerged from a specific institutional configuration: regulatory ambiguity created space for novel financial engineering, which attracted capital, which generated returns that attracted more capital until the structure collapsed under its own weight. In frontier AI, the moral hazard emerges from a different but structurally similar configuration: geopolitical competition has created political cover for capability expansion, which attracts capital, which generates capabilities that attract more political attention until the system reaches some as-yet-undefined boundary condition. The difference is that the boundary condition in AI may not be a market correction but something more fundamental—an irreversible capability, an uncontainable system, a moment when the technology exceeds our ability to govern it.
I want to be precise here about what I am not saying. I am not arguing that AI development should be paused indefinitely, nor am I arguing that the laboratories' calls for deceleration were motivated by pure altruism. What I am arguing is that the binary framing—acceleration versus deceleration—obscures a more important question: accountability. The question is not whether we build powerful AI systems, but who bears the cost when those systems fail, who is responsible when capabilities are deployed that cause harm, and what governance structures exist to ensure that the distribution of benefits and risks aligns with any reasonable conception of the public interest.
The contrarian angle I want to offer here is that the laboratories' calls for "slowing down" may have been, in a structural sense, the wrong ask. By framing the debate in terms of speed, they inadvertently ceded the more important terrain: the terrain of accountability and governance. They asked: "Can we please build more carefully?" What they should have asked, and what I suspect some of their researchers were asking internally, is: "Who is accountable when this goes wrong, and do we have the governance structures to ensure that accountability is meaningful?"
This reframing matters because speed is a variable that is extraordinarily difficult to govern. You can pass laws requiring safety evaluations before deployment, but the definition of "safety evaluation" is contested, the timelines are negotiable, and the enforcement mechanisms are weak. By contrast, accountability structures—liability frameworks, mandatory incident reporting, transparency requirements, and genuine oversight mechanisms—are potentially more robust, because they address the question of consequence rather than the question of process. The laboratories could be allowed to move fast, as long as they moved with genuine skin in the game, real liability for harms, and meaningful transparency about what they were building and why.
The current policy environment, however, seems designed to prevent exactly this conversation. By framing safety advocacy as un-American, by suggesting that any constraint on capability expansion is a concession to foreign adversaries, the political discourse has foreclosed the possibility of a mature governance debate. This is not unique to AI, of course. I have watched the same dynamic play out in blockchain regulation, where legitimate concerns about investor protection and financial stability were repeatedly reframed as attacks on innovation and American competitiveness. The result was a regulatory patchwork that ultimately protected neither innovation nor investors, and that left the most egregious actors—those who had never intended to build anything beyond a tokenized ponzi scheme—largely immune from consequence.
What the September 2025 moment revealed, then, was not merely a policy disagreement but a fundamental disagreement about the nature of technological progress itself. Is progress a value that trumps all others? Is competitive advantage in AI so important that it justifies the externalization of risk onto the public? Or do we believe, as I do, that progress without accountability is not progress but acceleration toward an undefined destination?
The three laboratories whose names were invoked in that September moment are not monolithic entities. Anthropic, OpenAI, and xAI represent fundamentally different approaches to the question of AI development, different relationships to openness and transparency, and different internal governance structures. What they share is a position at the frontier of capability, a position that comes with both extraordinary power and extraordinary responsibility. The media narrative that collapsed these three distinct organizations into a unified "safety lobby" did a disservice to the complexity of the actual positions being articulated.
Anthropic, for instance, has been the most explicit about the potential for catastrophic risk from frontier AI systems. Their interpretability research and Constitutional AI framework represent genuine attempts to build accountability into the architecture of their systems, not merely to add safety evaluations after the fact. OpenAI, despite its recent corporate restructuring and commercial pressures, retains a research culture that has produced genuinely important work on alignment and evaluation. And xAI, whatever one thinks of its founder's public pronouncements, has made computational commitments that suggest a genuine belief in the importance of scale. These are not positions that can be meaningfully unified into a single "call to slow down."
The geopolitical framing of AI as a zero-sum competition—"whoever wins AI wins everything"—is itself a narrative choice, not a technical necessity. It is a frame that serves certain interests (those who benefit from reduced safety constraints and accelerated deployment) at the expense of others (those who will bear the costs of inadequate governance). I have seen this narrative frame deployed repeatedly in the blockchain space, where "we must move fast to compete with foreign jurisdictions" was used to justify regulatory arbitrage that ultimately harmed retail investors and legitimate businesses alike. The frame is seductive because it presents a false urgency, a manufactured crisis that demands immediate action without the inconvenient delays of democratic deliberation or safety evaluation.
What I find most troubling about the September moment is not the policy itself but the silence that followed. The laboratories, having been publicly rebuffed, largely retreated to carefully worded statements that did not challenge the political framing. The research community, already facing precarious employment conditions, had limited capacity to dissent publicly. And the policy community, having found a politically convenient position, had little incentive to revisit the question. This is how moral hazards become institutionalized—not through dramatic decisions but through the quiet accumulation of small choices that make accountability increasingly difficult to demand.
Don't trade the chart; trade the story. In the AI context, this means understanding that the relevant variable is not the current state of capabilities but the narrative structures that are being constructed around those capabilities. The story being told in September 2025 is a story about American competitiveness, about the necessity of speed, about the irrelevance of caution. That story serves certain interests, and we should be clear about whose interests those are.
Forward-looking assessment requires us to ask what happens next. If the accelerationist position holds, we should expect continued expansion of frontier model capabilities, increasing deployment pressure on sensitive applications, and diminishing space for safety research within the leading laboratories. We should expect the geopolitical framing to intensify, with any criticism of capability expansion being labeled as un-American or as aiding adversaries. And we should expect, eventually, a reckoning—either through a catastrophic incident that forces a policy reversal, or through a gradual erosion of public trust that makes the political position untenable.
The laboratories, for their part, face a choice that the September moment may have clarified for them. They can continue to operate within a political environment that demands capability expansion without meaningful accountability. Or they can find ways to build genuine accountability structures into their organizations—structures that would survive the current political moment and provide a foundation for sustainable governance. This is not a choice between speed and safety; it is a choice between external accountability and self-governance, between the narrative of progress and the reality of responsibility.
What I have learned, in my years of studying technological transitions, is that the narratives we tell ourselves matter enormously. The story that acceleration is synonymous with progress, that caution is weakness, that the only thing that matters is staying ahead in the competition—these are not truths but choices, narrative constructions that serve certain interests and harm others. The September moment was a choice, made by identifiable actors, with identifiable consequences. Whether those consequences are acceptable depends on the answer to a question that has not yet been asked: acceptable to whom, and accountable by whom?
The ghost in the blockchain is us, watching our own reflections in the code we write. And in the AI systems being trained on our collective knowledge, we might do well to ask what reflections we are encoding, what values we are embedding, and what accountability we are building into the foundations of the systems we are creating. The speed of progress is a variable we can influence. The question is whether we will influence it in the direction of human flourishing or in the direction of accelerated risk. That choice, unlike the choice of speed, cannot be delegated to the laboratories, the markets, or the politicians. It is a choice that belongs to all of us, whether we have the courage to claim it or not.