Takeoff speeds rule everything around me

Core claim

Short AGI timelines are now near-consensus, so they no longer explain disagreement about AI x-risk. The remaining disagreement is secretly about takeoff speed: many sceptics combine very short timelines to “AGI” with very long timelines to AI transforming the physical world. Operationalising takeoff in terms of the rate of real-world technological progress makes the disagreement legible, and that single parameter determines the answers to most strategic and policy questions.

The shift in the debate

• A decade ago, AI x-risk debates centred on AGI timelines: over a century away, or plausibly as soon as 20 years?
• Today even people relatively sceptical of x-risk often hold very short timelines — ten years, five years, or “negative two months”.
• So short timelines no longer account for the remaining disagreement about the level of imminent risk and the appropriate policy stance.
• Cotra’s diagnosis: it is still mostly about timelines, but of a different kind — sceptics pair short timelines to AGI with long timelines to AI transforming the real world, i.e. a very slow takeoff.

Caveat on "AGI"

The definition of AGI is often unproductively watered down. Those with the very shortest timelines (1–2 years) are disproportionately likely to be forecasting a milder version of “AGI,” which in turn exaggerates how long they believe the period from AGI to radical superintelligence will be.

Operationalising takeoff

• The classic definition (Bostrom, Superintelligence, 2014) is the time from AGI to superintelligence — but both terms are slippery, and “how super” an intelligence is can’t be specified without naming concrete real-world feats.
• Cotra’s proposed yardstick: the speed of progress in physical technology, benchmarked against the human-driven pace of the last ~20 years across hard-tech fields (hardware, batteries, materials, industrial chemistry, robotics, spacecraft).
• The framing question: suppose that in 2028 we develop AI that automates all the intellectual work of scientists across these fields. How many “years of progress” at the old human pace does each calendar year then deliver?
• An alternative yardstick is economic output / gross world product — used by Paul Christiano (2018), Epoch’s GATE model, and Tom Davidson’s 2021 takeoff speeds report — but it introduces complications inessential to the basic argument.

Three views, conditional on the same timeline

All three views below assume full cognitive automation of science in mid-2028; they differ only in what happens next.

Fast takeoff

• Once AI fully automates AI R&D, a strong super-exponential feedback loop produces unfathomably superhuman AI within months.
• During that software-based intelligence explosion, physical technology does not visibly improve much.
• Once god-like AI exists, sci-fi technologies (molecular nanotechnology, whole brain emulation, reversible computing, near-light-speed spacecraft) arrive in months, weeks or days — things that would have taken centuries at human rates.

Slow takeoff

• There is still a super-exponential AI-improving-AI feedback loop.
• The difference: getting the loop going requires physical automation of the entire AI stack, so it is more gradual — and physical signatures appear before the fully automated scientist arrives.
• Even so, “slow” takeoff means going from automated science to an unrecognisably sci-fi world within a matter of years.

No takeoff

• Cotra’s read of most x-risk sceptics: AI automating scientific research isn’t that big a deal.
• Perhaps it makes R&D modestly faster; perhaps automation is merely what keeps us at our previous pace when we would otherwise have stagnated.
• On this view AGI is the next innovation sustaining ~2% frontier growth a while longer, or adding maybe half a percentage point.

Why the physical-technology graph

A graph of AI cognitive capability would grow rapidly but smoothly across 2026–2029 on all views (cf. the AI Futures Project’s takeoff model). The physical-technology graph better illustrates how takeoff will feel to people outside AI companies.

Why this one parameter dominates

• Decisive strategic advantage: if a six-month lead in AI translates into a centuries-long lead in weapons technology, the first country to AGI plausibly gains the ability to impose its will on everyone else.
• Extinction risk from misalignment: plausible if misaligned systems can trivially develop a superplague and upload themselves into self-replicating nanocomputers to survive without humans.
• The case for slowing down: worth it if it buys a year or two to absorb the impacts of 2100-level technology before facing 2200-level technology.
• Because powerful AI (at least an 8 on Nate Silver’s Richter Scale) now feels imminent to doomers and sceptics alike, the deeper disagreement over how powerful and how quickly has been obscured.
• The gap in worldviews: some see AGI as an innovation sustaining frontier growth; others see humanity as “t minus a few years from first contact with an unfathomably advanced alien species, entities we would view as nothing less than gods.”

Conditioning caveat

The three curves hold the timeline to fully-automated science fixed for illustration. In reality, slower takeoff speeds probably also imply longer timelines to that initial milestone — and there would be room for more acceleration prior to full automation.