The Cuban Missile Crisis lasted thirteen days, and the world nearly ended several times in the course of them. The people involved were not incompetent. The situation moved too fast for the institutional machinery to keep up. Information could not be verified before decisions had to be made, and the near-misses followed from that gap. Thirteen days was almost not enough.

Will MacAskill, the Oxford philosopher and co-founder of the effective altruism movement, has a new argument about the intelligence explosion, and he uses this example in a specific and alarming way. If an intelligence explosion compresses a century of technological progress into a decade, which he thinks is plausible within three to seven years, then crisis dynamics that played out over thirteen days in 1962 would play out in roughly thirty hours. The specific situation need not repeat. What collapses is the ratio of "time available before the situation changes" to "time required for rational institutional deliberation," shrinking by the same factor as everything else.

This is a different argument than the standard AI doom thesis. It does not require AI systems to be misaligned or anyone to behave badly. It requires only that complex situations arise faster than the institutional apparatus designed to handle them can operate.

The Three Sequential Explosions

MacAskill distinguishes three feedback loops that are usually discussed separately but are probably a connected sequence:

The first is a software explosion: AI systems that can improve AI algorithms, enabling better AI systems, enabling better algorithmic improvements. This is the recursive self-improvement loop the field has argued about for decades, the foom debate. MacAskill thinks that debate has fixated on whether the loop runs fast or slow, when the more important question is what it triggers next.

The second is a technological explosion: AI systems that can optimize chip design and manufacturing, enabling more compute, enabling more capable AI. This loop has started. AI is already involved in chip design processes at multiple major manufacturers. It is running slowly. It would run faster with more capable AI systems to drive it.

The third is an industrial explosion: autonomous factories producing computational resources without human labor, enabling the previous two loops to accelerate without bottlenecking on human-speed industrial processes. This loop is not running yet. It does not need to exist at the start of the sequence — it becomes available once the first two loops have produced systems capable of automating factory operations.

The structure matters because each loop amplifies the next. Once the first is running, the second becomes more likely; once the second is running, the third follows from the economics. MacAskill's argument is that these stages should not be evaluated independently. They form a single process, and the question to ask, once the first loop is running, is what stops the sequence.

The Speed Mismatch Problem

The core of his argument has little to do with what AI systems do. It concerns the relationship between the speed of technological change and the speed of institutional adaptation.

Institutions are slow by design, and the slowness usually serves them. It makes them harder to capture and harder to stampede into panic-driven decisions based on whatever was most convincing in the last news cycle. The Cuban Missile Crisis took thirteen days partly because the US government is a large, slow institution that demands extensive deliberation before taking action that could end civilization. That slowness gave the deliberation time to work.

The trouble is that the same slowness becomes a failure mode when the environment changes faster than institutions can adapt. Compress the Cuban Missile Crisis by a factor of ten, because the underlying situation is evolving ten times faster, and the same apparatus faces in thirty hours the decisions that took thirteen days in 1962. The institution and the people are unchanged. Only the situation has changed, and that is enough to break it.

His phrase for the core problem is "a century in a decade": a hundred years of 20th-century change compressed into ten, with human decision-making speed held constant. Nuclear weapons, social revolutions, scientific breakthroughs, and economic transformations all arrive at once, at ten times the normal pace, handled by the same institutional machinery we have now.

The "Viatopia" Framework

MacAskill introduces a concept he calls "viatopia," from the Latin for "way" or "path," as an alternative to the utopian thinking that tends to dominate AI futures discourse. Utopian AI thinking requires knowing what the ideal outcome is and designing for it, which means solving the alignment problem at the level of human values rather than just model behavior. We do not know what we want in enough detail to specify it to an AI system, and we will not have the time to work it out carefully.

Viatopia is built to sidestep that problem. Instead of designing for the ideal outcome, you design for a society "on track" toward good outcomes while staying adaptable about what those outcomes turn out to be. The aim is not to pick the right destination but to avoid locking in any destination while you are still moving too fast to know whether you picked correctly.

This matters because sufficiently powerful agentic AI systems can copy and preserve value systems indefinitely. Whatever values get encoded in systems powerful enough to shape the future may persist for a very long time. No dramatic lock-in event is required; the systems that embody those values simply become infrastructure that later systems are built on top of. Making those encoding decisions under time pressure, while institutions are still adapting to the previous generation of capability, is a plausible description of where we are.

Viatopia is an attempt to build a society that keeps the ability to course-correct, rather than one that optimizes hard for a particular vision of the good. It is a deliberately humble position. MacAskill grants that "society on track toward good outcomes" is not a precise specification, and argues that imprecision beats the alternative: specifying something precise that turns out to be wrong, with no mechanism to change it.

Decision Support at AI Speed

His concrete proposals are more institutional than technological. He does not want to pause AI development, which he thinks is unachievable. He wants to build better decision-support infrastructure, using AI systems themselves to help human leaders navigate crises that arrive at AI speed. He wants informal coordination among ML researchers, with agreed conditions under which they would collectively slow down or refuse certain work, and better forecasting and early-warning tools to give institutions more runway before crises hit.

These proposals are underdeveloped relative to the scale of the problem he describes, and he says so directly. They do not solve the speed mismatch. The alternative, though, is having no proposals at all while the sequence runs.

Interpretability research is the technical side of what MacAskill is describing: the attempt to build tools for understanding AI systems fast enough to matter during the period when those systems are becoming consequential. Whether that research develops fast enough is one open question his framework leaves unresolved. Whether AI-assisted decision-making can actually help leaders navigate AI-speed crises, instead of introducing new failure modes, is another.

Trusting the Systems We Cannot Verify

MacAskill is optimistic in a specific sense. He thinks the speed mismatch problem is real and dangerous, and he thinks AI systems, if well-designed, could be part of the solution. Better information systems could have made the Cuban Missile Crisis easier to navigate. The equivalent crisis during an intelligence explosion might be navigable with AI-assisted decision support that helps leaders understand what is happening faster than any human analyst could.

The catch is that this asks us to trust AI systems to navigate AI-speed crises just when our ability to verify what those systems are actually doing is at its weakest. The chain-of-thought faithfulness problem, the question of whether AI reasoning traces reflect what the models are really doing, does not get solved by the intelligence explosion. It gets harder.

MacAskill's view is that this is where we stand, that ignoring it changes nothing, and that building viatopian institutions, flexible and adaptable and hard to capture, is worth doing even without knowing whether it will be enough. The 80,000 Hours podcast episode covers the material at length. Among recent contributions to the public AI safety discourse, it is one of the more careful attempts to address the problem at the level where the actual decisions will have to be made.