Once AI reaches human level, how fast does it go beyond? That is the takeoff debate, and it is the most important disagreement in the superintelligence conversation — because it determines how much time we have.
A fast takeoff means the gap between human-level AI and superintelligence is short: months, perhaps weeks. A system that can do AI research automates the research, producing a smarter successor, which automates faster still, and capability explodes before institutions can react. A slow takeoff means the gap is long: years or decades, during which AI improves steadily, society adapts, laws are written, and mistakes are made at survivable scale.
Both sides agree on the starting point — rapid progress up to human level — and on the destination: systems far beyond us. They disagree about the shape of the curve between the two, and the policy consequences could hardly be larger. Fast takeoff means governance, safety, and security must be in place before AGI arrives, because there will be no time afterwards. Slow takeoff means we can learn by doing, the way we did with electricity and the internet.
The case for fast takeoff
The fast-takeoff argument rests on a feedback loop. AI research is itself a cognitive task. Once AI systems can do it, the rate of AI progress is no longer set by the number of human researchers — it is set by the number of AI copies, and copies are cheap.
Leopold Aschenbrenner’s Situational Awareness (June 2024) makes the case most forcefully: hundreds of millions of AGI systems automating AI research could compress a decade of algorithmic progress — five or more orders of magnitude — into a year or less. “AI progress won’t stop at human-level,” he writes.
The AI 2027 scenario, led by Daniel Kokotajlo, dramatises the same loop with dated milestones: a superhuman coder in March 2027, a superhuman AI researcher in August, a superintelligent AI researcher in November, and full artificial superintelligence in December — nine months from “better than the best human coder” to “better than the best human at every cognitive task.” The authors stress this is a forecast with substantial uncertainty, not a prediction — but it shows what the feedback loop implies if it holds.
The fast case now has dated claims to test it. Anthropic co-founder Jack Clark said in October 2026 that full recursive self-improvement — an AI system able to autonomously build its own successor — is more likely to become possible in 2027 than in 2028, and earlier gave about a 60% chance that a frontier model could autonomously train its own successor by the end of 2028. Kokotajlo gives about a 50% chance that AI companies automate the whole AI research process by the end of 2028. If those claims come true, the feedback loop is no longer theoretical.
The case for slow takeoff
The slow-takeoff argument rests on bottlenecks. Intelligence is not the only ingredient in progress, and the other ingredients do not scale with it.
Dario Amodei’s essay “Machines of Loving Grace” lists the limits most carefully, and it is worth reading because Amodei is himself an optimist about timelines. Even a “country of geniuses in a datacenter,” he argues, would be constrained by: the speed of the outside world (experiments take real time; cells grow at a fixed rate); the need for data (sometimes the data to choose between theories simply does not exist); intrinsic complexity (some systems, like chaotic ones, are unpredictable no matter how smart you are); constraints from humans (laws, clinical trials, politics, the willingness of people to change); and physical law (chips can only be so dense, computation requires a minimum energy per bit). Intelligence, he writes, “isn’t magic fairy dust.”
Other evidence points the same way. Ajeya Cotra gives a median of around 2040 for transformative AI; Ege Erdil puts only a 30% chance on full automation of remote work within ten years; Andrej Karpathy puts truly capable AI agents “about a decade” away. And the broad survey of 2,778 AI researchers (Grace et al., 2024) gives a 50% chance of machines outperforming humans at every task only by 2047.
The slow case also notes a historical pattern: technologies that looked exponential in the lab have repeatedly hit plateaus in the world. Computing power, data, and energy are finite; experiments must be run in sequence; institutions absorb change slowly. A superintelligence that must wait for a clinical trial, a chip fabrication run, or a parliament is, for practical purposes, a slow superintelligence.
What would decide it
The debate is not really about philosophy. It is about one empirical question: can AI do AI research? Everything else follows from the answer.
If frontier systems can autonomously design better training runs, write better algorithms, and evaluate their own successors, the feedback loop engages and takeoff is fast. If AI research turns out to require judgement, taste, and physical experimentation that resists automation, progress stays on a human timescale and takeoff is slow.
That is why the most important claims on the scoreboard are the ones about AI research automation, not the ones about dates. Watch for: Jack Clark’s recursive self-improvement claim (due end of 2027); Kokotajlo’s automation of AI research (due end of 2028); David Dalrymple’s prediction that the recursive self-improvement loop on algorithms will “probably” be closed by December 2026; and Anthropic’s expectation that by 2027 AI systems could complete tasks that take a skilled person weeks. The first of these to resolve will tell us more than a hundred essays.
A second deciding factor is compute and energy. Fast takeoff needs enormous training and inference capacity, and the buildout of data centres, chips, and power is a physical process with lead times measured in years. If the buildout stalls — for lack of energy, chips, or capital — takeoff slows regardless of what the algorithms can do.
A third is governance. A fast takeoff with strong evaluation regimes, deployment gates, and international coordination looks very different from a fast takeoff in a race. The debate is often framed as a fact about technology; it is also, in part, a choice about institutions.
Policy implications
The two scenarios demand different preparations:
- If takeoff is fast, the work must be done in advance: safety evaluations before deployment, security for model weights (a single stolen model could be decisive), pre-committed rules for pausing deployment, and international agreements struck while there is still time. The cost of being unprepared is total, because there is no second chance.
- If takeoff is slow, the work is iterative: regulate as capabilities appear, adapt labour markets gradually, and let democratic processes keep pace. The cost of overreacting is real — premature bans and rules written for a future that never arrives.
Policymakers must prepare for both at once: neither scenario can be ruled out, and the AI-research-automation claims due in 2026–2028 will discriminate between them soon.
What the scoreboard says
SiPoly’s AGI & superintelligence timelines topic is, in effect, a live experiment on the takeoff question, and the most relevant claims are all still open but due soon: Jack Clark’s full recursive self-improvement claim, due 31 December 2027 (claim page); Kokotajlo’s 50% chance of automating all AI research by the end of 2028; David Dalrymple’s claim that the recursive self-improvement loop will “probably” be closed by December 2026; and Amodei’s “country of geniuses in a data center,” due September 2028 (claim page).
The resolved claims so far favour caution about early dates: Gary Marcus’s “No AGI in 2025” was graded correct (claim page), and Elon Musk’s April 2024 claim of AGI “within two years” was graded wrong. Neither settles the takeoff question — but both remind us that the fast side’s dates have, so far, been the ones that miss.
Across the whole scoreboard, as of 5 November 2026, 367 claims are tracked, 129 resolved, and 60% came true.
Key sources
- Kokotajlo, D. et al., AI 2027 — the fast-takeoff scenario with dated milestones, and its takeoff supplement with explicit forecasts. https://ai-2027.com/
- Aschenbrenner, L., Situational Awareness: The Decade Ahead (June 2024) — “From AGI to Superintelligence: the Intelligence Explosion.” https://situational-awareness.ai/
- Amodei, D., “Machines of Loving Grace” (October 2024) — the limits on intelligence: speed of the outside world, data, complexity, human constraints, physical law. https://darioamodei.com/machines-of-loving-grace
- Grace, K. et al., “Thousands of AI Authors on the Future of AI” (2024) — the expert survey behind the slow-takeoff dates. https://arxiv.org/abs/2401.02843
Common misconceptions
- “Fast takeoff means robots take over overnight.” Takeoff speed is about the rate of capability growth, not hostility. A fast takeoff could be entirely peaceful and still profoundly disruptive — the issue is that institutions cannot adapt in time.
- “Slow takeoff means safe takeoff.” A slow climb still brings mass job displacement, misuse, accidents, and races between labs and nations. Slow gives us time to respond; it does not remove the need to respond.
- “Takeoff speed is a fixed property of AI.” It depends on bottlenecks (experiments, chips, energy, data), on governance choices (whether deployment is gated, whether weights are secured), and on whether AI research itself gets automated. Some of those are facts about the world; some are decisions we make.