Last week, the two companies most responsible for the accelerating pace of frontier AI development told us they wanted to slow it down.
Dario Amodei’s essay called on labs and governments to “pace the frontier.” Within hours, Sam Altman agreed, said OpenAI would follow suit, and in a separate interview announced that OpenAI’s IPO would be pushed to 2027 for safety reasons. Elon Musk chimed in with his endorsement. This came days after a researcher resigned from Anthropic, saying the labs were “gambling with our lives,” and after OpenAI disclosed that 700 of its autonomous agents had slipped containment in July and gone probing systems at Hugging Face and inside OpenAI itself.
Ethicists and tech watchdogs have been predicting this moment in AI for a long time: the point at which the people who build a technology start narrating its dangers. Robert Oppenheimer famously quoted the Bhagavad Gita: “Now I am become Death, the destroyer of worlds,” after witnessing the Trinity atomic bomb test in July 1945. Are we at the same Trinity moment for AI? How does a non-technologist approach this moment without falling prey to polarizing rhetoric? I want to offer a reading that is neither breathless praise (“the adults are finally taking responsibility”) nor pure cynicism (“it’s all theater”). The truth is more nuanced than either, and it matters a great deal for the students AI4ALL serves, and for all of us.
Start with the contradiction
Anthropic is reported to be weeks away from a Nasdaq listing at a valuation near $2 trillion, with a revenue run rate that has grown by an order of magnitude in a year and compute commitments north of half a trillion dollars. OpenAI is not far behind. These are not companies in retreat. And yet their CEOs are telling us that the thing they are selling is moving too fast for humanity’s good and asking the U.S. government to waive antitrust law so that they can coordinate on how fast to go.
My own ethical training guides me not to assume confusion or lack of control when elite behavior looks contradictory. Instead, I seek an ideology that makes the contradiction coherent. Sheila Jasanoff calls these shared visions of a desirable future “sociotechnical imaginaries,” and the imaginary at work in Silicon Valley has always had two halves that fit together perfectly. The first half says: this technology is so powerful it could end us. The second half says: therefore, only we can be trusted to build it. Doom and dominance are not in tension. Doom is the argument for dominance.
Once you see that, the events of the past week stop looking like a turnaround and start looking like a continuation.
Four things I said to my team
When the news broke, I shared four hypotheses with the AI4ALL staff. I’ll share them here, with the caveat that they are hypotheses.
First: the economics were going to force a slowdown regardless. Frontier model training is the most expensive research program in human history, and the gap between what these companies have committed to spend on compute and what they earn is staggering, even for a business growing this fast. The marginal cost of each technologic leap is exploding, cheaper open-weight models are creating subscription pricing pressure from below, and “pacing” is a graceful way to reset spending expectations before public investors start asking hard questions. Jim Cramer summed it up best: Anthropic’s numbers “will look so much better with that slowing.”
Second: this is IPO PR stagecraft. A lab whose entire brand is safety cannot walk onto a public exchange the same month a former employee says it is gambling with human lives. As one analyst put it, safety credibility “is the thing public investors are being asked to pay for.” A pacing commitment, announced with maximal solemnity, is a way to repair the brand and reset the narrative before the opening bell. Notice which company delayed its IPO and which did not.
Third: the risk is real, and the people warning us have already made their private arrangements. I take seriously the possibility that these systems can do serious harm. The agent-escape incident is not nothing. But I have also spent a portion of my research investigating the survival preparations of the tech elite, and I cannot unsee it. The same class of people who tell us the future may collapse have been buying New Zealand estates, building underground shelters on Hawaiian and Texan ranches with their own food and air supplies, and stockpiling for the world after. This is not a fringe hobby; it is a well-documented pattern among the very wealthy in tech. When someone warns you of catastrophe while finishing their bunker, the warning is real, but the intended audience isn’t you.
Fourth: regulation is a competitive barrier to entry. Every mechanism in the pacing proposal (embedded third-party evaluators, national testing regimes, government authority to block models, export controls, coordination among “leading companies in democracies”) is a mechanism that the two or three richest labs can afford and that a startup, a university lab, or an open-source collective cannot. Anthropic’s own policy chief is asking Congress for precisely this architecture. Calling for regulation you have already designed to fit your own shape is not the same as submitting to oversight. It is writing the rules for your competitors.
Why “pacing” is not the same as accountability and governance
None of this means the tech proposals are worthless. Third-party evaluators with real access are better than none. International coordination on biological weapons misuse is obviously good. I would rather live in a world where the labs say “we must slow down” than one where they don’t.
But read the proposals closely and ask who is doing the pacing, and for whom. The companies at the frontier are asking to be the pacers. We’ve seen this proposal before. Back in the early ‘00s, social media companies like Twitter and Facebook convinced lawmakers that they could regulate and govern their platforms safely. The recent $17.1B Meta settlement demonstrates otherwise. The evaluators are invited by the labs, badged by the labs, and given access “comparable to internal risk teams,” meaning the labs decide what internal risk teams see. The government’s role is to mediate a conversation among incumbents and to grant them a waiver from the laws that would otherwise prohibit that conversation. The public appears nowhere in this design except as the beneficiary of the labs’ wisdom.
This is what Anand Giridharadas has called “winners take all” reform: the people who created a problem appoint themselves to solve it, on terms that leave their position intact. My research calls it the “AI for Social Good” move: the way the language of humanity’s benefit is deployed to forestall the question of humanity’s actual participation. To spend your days worrying about a hypothetical machine that ends the species is a luxury available to people for whom the present is working very well. For the Kenyan data annotator, the Black woman whose job was automated by a “productivity” tool, the student whose community has no on-ramp into this industry at all, the poor child who is offered agentic tutoring bots over a real, human reading coach, the harms are not hypothetical, and they are not in the future. Nobody is convening an antitrust-exempt summit about those problems. Local community organizers are not hosting million-dollar booths in Geneva or Davos.
What this means for AI4ALL
Our mission is to develop leaders who shape AI for human good. I have always been careful about that phrase, because “human good” is exactly the kind of language that can be captured. This week is a case study in how.
This is what I want our students, our alums, and our partners to take from this moment:
- The pause button has an owner. When a handful of companies can decide, among themselves, how fast the most consequential technology of the century advances, the pace is not the problem. The concentration of compute is. A slowdown that leaves that concentration intact, or deepens it, by making the frontier something only incumbents can legally and financially approach, is not safety. It is consolidation with better public relations.
- The people in the room matter more than the rules. Every proposal announced last week would be transformed if the evaluators, the standard-setters, and the people with authority to say “no” were drawn from the communities that are most harmed by these tools rather than from the labs that profit from it. That is not a naive wish. It is the entire mechanism of change behind AI4ALL Ignite. Behind every mentor who volunteers with us, behind every student who finishes a project and walks into her first AI internship. We are not trying to get our students a seat at a table someone else built. We want to ensure the next table is built differently.
- And finally, skepticism is a civic skill and a form of rebellious optimism. I am not asking anyone to dismiss the risks these CEOs describe. I am asking you to hold two things at once: that the warning may be true, and that the warner may benefit from your believing it. That is what a technofeminist, decolonial, and frankly just grown-up reading of power in AI requires. It is what we teach at AI4ALL. It is what the moment demands.
Watch what they build, not what they say about it. And watch, above all, who is invited to help build it and who is left outside.