Watching the Frontier

I have been sitting with a proposal from one of the major AI laboratories about slowing the pace of development, and the more closely I examine it, the more its unresolved gaps become apparent. The instinct behind it is easy to respect. The difficulty is that a proposal of this kind has to hold in four dimensions at once, and each of them raises questions the laboratory that wrote it cannot answer alone.

The problem of collective participation

The first difficulty is participation. A single company placing external reviewers alongside its own researchers demonstrates willingness, and willingness is not nothing. It does not, however, establish industry-wide compliance. Persuading competitors to accept comparable oversight, in the absence of a regulatory framework that compels them, is a different order of problem entirely. It asks rivals to hand a competitor a view into the very thing they are racing toward.

The antitrust implications of competing laboratories coordinating directly with one another are precisely why any serious version of such an arrangement may ultimately require governments to occupy the center of it. Voluntary coordination at this scale is not a private understanding that a few executives can settle over a call. It is closer to treaty-making, and treaties are what states resort to when the cost of disorder finally exceeds the cost of restraint.

Who benefits from stricter rules

The second question is who benefits from stricter rules. A safety regime built around embedded evaluators, extensive interpretability research, rigorous testing, and slower releases is expensive, and expense is never distributed evenly. The largest laboratories are far better positioned to absorb those costs than smaller organizations or companies still attempting to catch up.

That does not automatically make the underlying concerns insincere, and I want to be careful here, because the lazy reading collapses the whole thing into a question of motive. Genuine concern and competitive advantage can coexist inside the same proposal. A rising cost floor preserves the position of whoever already stands above it. Both facts have to be held at once rather than choosing the one that makes the argument easier to dismiss.

Consider what compliance would actually cost in practice. An embedded evaluation team is not a checklist; it is salaries for a dozen or more specialists, dedicated compute for red-teaming and interpretability runs, legal review for every disclosure, and delays measured in quarters rather than weeks. A frontier laboratory absorbs that as overhead. A university group fine-tuning an open-weights model on a borrowed cluster, or a small team in Kathmandu or Nairobi trying to adapt a model to local languages, experiences the same requirement as an existential filter. I have seen how quickly well-intentioned standards become, in effect, a licensing regime for incumbents, not because anyone designed them cynically, but because cost is a constraint that binds unevenly.

The geopolitical constraint

The third issue is geopolitical, and it is one no company can simply set aside. Any attempt by democratic countries to coordinate the pace of AI development is constrained by the technological lead those countries currently hold over potential competitors. Preserving that lead depends partly on export controls, security measures, semiconductor restrictions, and a broader set of strategic policies continuing to work. None of those questions is permanently settled. They remain active disputes rather than stable assumptions, and no long-term technological strategy can safely rest on ground that is still being argued over.

The supply chain makes this concrete rather than abstract. Training a frontier model depends on lithography machines built in a single Dutch town, fabrication concentrated on one island, packaging and memory spread across a handful of firms, and power and networking at data-center scale. Each of those chokepoints has been the subject of export-control debates from 2022 onward, with rules tightened, reinterpreted, and worked around in turn. Verification is the harder half: with nuclear programs, inspectors can count centrifuges. With weights that fit on a hard drive and training runs that can be split across clouds, there is no equivalent counting regime yet. Any slowdown agreement without verification is a promise; any verification regime without state power behind it is a suggestion.

What laboratories promise and what their people believe

The fourth issue sits closer to the laboratories themselves. Around the same period that such proposals were being discussed publicly, individuals inside AI companies were raising concerns about the speed of development and about the possibility that increasingly capable, potentially self-improving systems could create serious safety risks.

That does not invalidate the proposals. It does, however, highlight a distinction worth keeping in view: what an organization promises externally and what the people working inside it believe internally are not always the same thing. External evaluators are only as effective as the access they actually receive, not the access described in a policy document. Between those two phrases sits the entire question of whether the arrangement is a commitment or a press release.

What I am actually trying to say

Beneath all of these questions is a warning that deserves more attention than it receives: we may be underestimating what is approaching. Systems capable of discovering vulnerabilities, transforming entire professions, accumulating extraordinary economic and strategic leverage, and operating at scales beyond individual human supervision are no longer purely theoretical ideas. They are being built, funded, and deployed. At the same time, nothing about the current trajectory guarantees that development will naturally slow down.

What troubles me most is the possibility that some of the people building these systems privately believe things could go seriously wrong within the next decade. Whether every such concern is justified is a separate question, and I do not pretend to adjudicate it. What matters is that the possibility is being taken seriously by people standing close enough to the technology to read its trajectory clearly.

Moving rapidly on alignment and safety only makes sense if we are highly confident that no safer path exists. I am not convinced humanity has earned that level of certainty yet. We are accelerating through a landscape we have not finished mapping, and the instruments we would use to map it are the same ones we are accelerating with.

There is still room for hope, and I mean that plainly rather than as consolation. Real incidents, genuine warning signals, and a widening awareness of systemic risk may make coordination between laboratories more conceivable than it once was. The broader trajectory, however, still appears to point toward competition, and preventing an uncontrolled race may require measures far more substantial than voluntary statements or informal agreements. Perhaps that eventually means slowing capability development itself.

Coordination has precedents, just not at this speed

It helps to remember that humanity has solved slower versions of this problem. Civil aviation did not become safe because pilots individually tried harder; it became safe through mandatory incident reporting, independent investigation boards, and a culture where a near-miss in one airline rewrites procedures in all of them. Nuclear energy runs on inspections and material accounting. Even finance, hardly a model of restraint, eventually accepted audited statements as the price of public trust. Each of these regimes took decades and at least one disaster to build, and each rests on something AI currently lacks: a slow-moving physical substrate that gives institutions time to catch up.

Software diffuses faster than any of those precedents allow for. A safety-relevant technique published on Monday is replicated by Friday; a set of weights released openly cannot be recalled. That asymmetry is why I keep returning to compute as the one observable input: it can be metered, sited, and, in principle, governed, the way aviation fuel and fissile material are. Whether that observation becomes policy depends less on laboratories than on whether governments decide that the capability-governance gap is a public problem rather than an industry courtesy. Until then, proposals to slow down will read as what they are: serious people describing a real gap, without the instruments to close it.

The larger question is not whether AI will become extraordinarily powerful. Most people seriously involved in building these systems already assume that it will.

The question is whether technological capability is advancing faster than our collective ability to understand it, govern it, evaluate it, and coordinate around it.

That gap is what matters. Not a particular prediction. Not a particular date. The gap itself.