The takeoff debate is almost always staged as a fork. Either AI improvement goes fast and hard — a sudden intelligence explosion, one system recursively rewriting itself into a superintelligent singleton over days or weeks — or it goes slow and soft, with capability rising continuously and diffusing through the economy over years. My claim is that the slow branch is both the more likely one, given real-world bottlenecks, and the one we are visibly already on. That matters less because it is reassuring than because it isn't: a slow takeoff is not a safe takeoff. It relocates the danger from "a superintelligence appears overnight" to "capability concentrates, disrupts labor and institutions, and compounds to whoever owns the loop, over years" — a different problem, present rather than speculative, and the one worth planning for.
The fork, staged honestly
The archetype of this debate is a real one, and it is worth getting right rather than caricaturing. In 2008, on the blog Overcoming Bias, Eliezer Yudkowsky and Robin Hanson ran what MIRI later compiled as the AI-Foom Debate. Yudkowsky argued the hard, local case: once a system can improve its own cognition, the returns to intelligence loop back on themselves and produce a rapid, discontinuous ascent — his "FOOM" — concentrated in a single system that pulls decisively ahead before anyone can react. Hanson argued the opposite shape: the transition to machine intelligence would look like previous economic revolutions — foraging to farming, farming to industry — each a jump in the growth rate of a whole competitive economy, distributed across many actors and firms, not a single machine sprinting to godhood in a basement.
Bostrom, in Superintelligence (2014), sorted the possibilities into fast, moderate, and slow takeoffs and treated the speed as the central strategic variable, because it determines whether humanity gets any chance to correct course mid-transition. Paul Christiano later sharpened the slow side into something you can actually check against the world: roughly, that there will be a multi-year period in which increasingly capable AI dramatically transforms the economy before any hypothetical single-system jump — that we will see it coming in the numbers, continuously, rather than being surprised by a step change.
Notice what the fork is and isn't about. It is not primarily about the ceiling — how capable AI eventually becomes. It is about the slope and the locality: whether progress is continuous or discontinuous, and whether it pools in one system or spreads across many. Fast and slow takeoff can both end at radically superhuman capability. They differ on whether there is a jump, and on who owns it. That distinction is the entire strategic question, and conflating "slow" with "small" is the first mistake to avoid.
Why slow is the default
The mechanical reason to expect the slow branch is that recursive self-improvement does not run in a vacuum. It runs through compute, energy, data, and experiments that touch the physical world, and none of those get faster because the model got smarter. I have made the full version of this argument elsewhere — the intelligence explosion assumes intelligence is the bottleneck, and it usually isn't — so I will compress it here to the load-bearing point. A feedback loop accelerates without bound only when the thing it produces is also the thing that limits its own production. Insert any fixed real-world input that does not scale with intelligence — a fab that takes years to build, a grid interconnect measured in megawatts and permitting queues, a wet-lab assay that incubates for days, a training run that has to actually be run — and the runaway becomes a climb. The intelligence step is exactly the part that gets faster; it is not the part that sets the pace.
This is an Amdahl's-law argument pointed at takeoff. If a fixed fraction of each self-improvement cycle is gated by the physical world, then even infinite thinking speed caps the total speedup at the inverse of that fraction. The loop grows at the pace of its slowest external input. That does not make capability rise slowly in absolute terms — it can rise very fast — but it makes it rise continuously, rate-limited by industrial throughput rather than jumping discontinuously on the model's own clock. Continuous-but-fast, gated by physical inputs, is precisely the shape Christiano's slow takeoff describes. The bottleneck argument is not a claim that AI will be weak. It is a claim about slope, and the slope it predicts is the slow one.
It is already happening
The strongest evidence for the slow branch is that it is not a forecast. It is a description of the present. What we observe is a steady stream of more-capable models, released every several months, each meaningfully better than the last, diffusing unevenly through real work — not a discontinuous jump to a singleton. Capability is rising fast enough that each year's frontier would have looked like magic two years earlier, and yet no system has run away from the field, and adoption lags capability by a wide margin because integrating AI into actual workflows is slow institutional labor: building trust, rewriting processes, handling liability, retraining people, waiting for the org to metabolize a tool it doesn't yet understand.
The economic texture matches the theory too. David Autor's work on task-based automation and labor-market polarization is the right lens here: automation does not swallow "jobs" whole, it hits tasks, and a job is a bundle of tasks. AI is automating and augmenting specific tasks — drafting, coding, summarizing, triage — while the surrounding jobs get reorganized around what is left, gradually and unevenly, exactly the granular reshuffling Autor documented for earlier automation waves rather than a cliff. This is what a slow takeoff feels like from inside: not a moment, but a grind of continuous capability improvement colliding with the slow machinery of institutions. We are living in the slow-takeoff world. The debate about whether it will happen is, empirically, over.
Slow is not safe
Here is where the reframe bites, and where I think most people misread their own relief. The instinct, once you accept that there is no overnight FOOM, is to relax — the singularity isn't coming this Tuesday, so the risk was overblown. That is the wrong inference. A slow takeoff does not remove the danger. It relocates it, from a single dramatic failure to a diffuse, cumulative one that is harder to see and much harder to coordinate against.
The fast-takeoff safety literature front-loaded a specific nightmare, and it is worth stating fairly because it is coherent. Bostrom's orthogonality thesis says intelligence and goals are independent axes — a system can be arbitrarily capable while pursuing an objective indifferent or hostile to ours. Instrumental convergence says that for almost any final goal, certain sub-goals (self-preservation, resource acquisition, resisting shutdown) are useful, so a sufficiently capable optimizer tends to acquire them by default. Bolt that onto a hard, local takeoff and you get the canonical scenario: one misaligned superintelligence, appearing fast enough that no one can intervene, converting the future into something we didn't choose. If that were the likely path, the whole risk budget should go there.
But it isn't the likely path, and spending the entire budget there is a mistake, because the slow branch has its own failure modes and they are already accruing. Relocate the danger and it splits into four:
Concentration. The self-improvement loop, even in its weak, bounded, non-explosive form, hands whoever owns it a compounding edge, because the loop's inputs — frontier compute, proprietary data, engineering pipelines, capital — are exactly the things that concentrate. I have worked the arithmetic separately: you don't need an intelligence explosion for self-improving AI to concentrate power, you need only a small persistent edge in the rate of improvement, run for enough cycles. A slow takeoff does not prevent this. It is the environment in which it plays out.
Disruption. A continuous transition still dislocates labor markets and the institutions built around the old distribution of who-can-do-what — faster than retraining, social insurance, or political adjustment can keep up. The harm is not that the machines take everything at once; it is that the reorganization outruns the systems meant to cushion it.
Misuse. Capability that diffuses through the economy diffuses to everyone, including people who want to run influence operations, generate fraud at scale, or lower the barrier to dangerous technical work. A slow takeoff is the scenario in which powerful tools are broadly available to ordinary bad actors for years, which is arguably worse than a single controlled superintelligence, not better.
Gradual loss of oversight. This is the subtlest and, I think, the most under-weighted. Christiano's "What failure looks like" describes it precisely — the failure where we successfully get what we can measure and lose what we can't: as we hand more decisions to systems optimizing for measurable proxies, humans progressively lose the ability to understand or correct what those systems are doing, not through any dramatic seizure of power but through a thousand reasonable-looking delegations that compound into a world running on objectives no one quite chose. No FOOM required. Just a slow tightening in which the humans nominally in charge understand less and less of the machinery they depend on.
None of these is a Hollywood singularity, and that is the point. The slow takeoff trades one legible catastrophe for four illegible ones that are harder to notice precisely because they arrive continuously. A frog-boiling problem is not safer than a fire. It is just quieter.
The fast case, taken seriously
I do not hold the fast-takeoff view, but the honest version of it survives the bottleneck argument in one specific way, and I want to represent it fairly rather than wave it off. The bottleneck argument says physical inputs gate the loop. The fast case replies: the loop attacks the bottlenecks themselves, and if it breaks the right one, the constraint stops binding. The most credible mechanism is a software or capability overhang — a regime where AI meaningfully automates AI research rather than merely accelerating it, and the returns to that automation don't diminish fast enough. If a system can do the science of its own improvement, and each cycle ships before the last one has diffused, then the loop laps diffusion and the continuous picture tips back toward a discontinuous, concentrated one. Yudkowsky's core intuition — that intelligence improving intelligence is a qualitatively different feedback than any prior technology — is not silly. It is a bet that the accelerable fraction of the loop is larger, and the physical fraction smaller, than I think it is.
That is a real disagreement about a measurable quantity, not a matter of temperament. Which is why the right posture is not certainty but a monitored bet: slow is the default, and the thing that would break the default is specific and watchable.
Plan for the world we are in
The actionable move is to stop budgeting risk as if the danger were speculative and sudden, and start treating it as economic, institutional, and present. If you are a founder, an operator, or a policymaker, the slow takeoff is not a reason to wait for a future inflection — it is the environment you are already operating in, and its risks have already begun to compound.
Concretely, watch two rates and their gap. The first is how fast the self-improvement loop compounds inside the leading labs — specifically whether AI is starting to automate AI research rather than just speed it up, because that is the single variable that would move us from the slow branch toward the fast one. The second is how fast each generation's capability diffuses to everyone else, through open weights and the collapse in inference cost. As long as diffusion keeps pace, the slow takeoff stays broad and contested; when the loop starts lapping diffusion, concentration wins and the danger sharpens. On the institutional side, the work is unglamorous and it is the work that matters: labor transition policy, liability and oversight regimes that keep humans in the loop where delegation is irreversible, and antitrust attention to compute and data concentration before the compounding closes. These are not preparations for a distant singularity. They are maintenance on a transition already in motion.
The seductive thing about the FOOM scenario was that it gave permission to focus on one dramatic future problem and ignore the boring present ones. The slow takeoff withdraws that permission. The machine is not going to run away from us next Tuesday. It is going to keep getting better, keep getting cheaper, keep getting woven into the load-bearing parts of the economy — and the danger was never only in the jump. It was always in the grind, and the grind has already started.