Within AI Takeoff
What would warn US before FOOM?
A hard takeoff would leave little time for correction, so the key question is what observable signs might precede a sudden capability jump.
On this page
- What hard takeoff means in plain English
- Warning signs before recursive capability gains accelerate
- Why slow takeoff arguments still matter
Page outline Jump by section
Introduction
If a true AI “FOOM” happened — a rapid intelligence explosion driven by recursive self-improvement — the central question would not be whether it is dangerous, but whether there would be enough warning to respond. In AI doom discussions, a hard takeoff means capability growth compressed into months, weeks, days, or even hours rather than decades. The debate is not simply about speed. It is about whether warning signs would appear early enough for safety measures, governance, and human decision-makers to matter. Some researchers argue that a fast takeoff would still produce observable indicators before the most dramatic acceleration. Others argue that the very nature of a FOOM is that many conventional warning signals would arrive too late or be misleading. AI Safety Atlas [AI Alignment Forum]alignmentforum.orgdistinguishing definitions of takeoffAI Alignment ForumDistinguishing definitions of takeoff13 Feb 2020 — A fast takeoff is one that occurs over the timescale of minutes, hou…
Because AI doom arguments often depend on loss of control during a period of rapid capability growth, identifying credible warning signs has become one of the most practical questions in the field.
What hard takeoff means in plain English
The term “FOOM” originated in early discussions of intelligence explosion scenarios. The basic idea is that once an AI becomes sufficiently capable of improving AI systems, each improvement makes it better at producing the next improvement. Instead of progress proceeding at a roughly steady pace, capability growth could accelerate rapidly. [AI Safety Atlas]ai-safety-atlas.comAI Safety Atlas TakeoffAI Safety AtlasTakeoff - Chapter 1Fast takeoff describes scenarios where AI capabilities increase dramatically over very short periods… [LessWrong]lesswrong.comai takeoff30 Dec 2024 — A hard takeoff (or an AI going "FOOM") refers to AGI expansion in a matter of minutes, days, or months. It is a fast, abrup…
Different researchers define takeoff speeds differently, but a common distinction is: [sideways-view.com]sideways-view.comThe sideways view Takeoff speedsThe sideways viewTakeoff speedsFebruary 24, 2018 — 24 Feb 2018 — Slower takeoff means faster progress. Fast takeoff is often justified by…
- Slow takeoff: transformative capabilities emerge over years or decades.
- Moderate takeoff: major changes occur over months or several years.
- Fast or hard takeoff: capabilities jump over days, weeks, or months, leaving little opportunity for intervention. [AI Alignment Forum]alignmentforum.orgdistinguishing definitions of takeoffAI Alignment ForumDistinguishing definitions of takeoff13 Feb 2020 — A fast takeoff is one that occurs over the timescale of minutes, hou…
Importantly, even people who take FOOM seriously disagree about timelines. Some hard-takeoff advocates imagine very rapid changes in AI capability while acknowledging that changes in the physical world might still unfold more slowly because of limits in manufacturing, infrastructure, energy, and hardware deployment. [Reddit]reddit.comRedditAre we heading for a hard takeoff? How do you think it…Personally, I think it will be a hard takeoff in terms of self-recursive…
For AI doom arguments, however, the crucial concern is not how quickly robots appear in factories. It is how quickly a system might move beyond the range where human oversight remains effective.
Which warning signs would matter most?
Many signs commonly cited in public discussion are probably too late to be useful. If stock markets surge, economies transform, or entire professions disappear, then the transition is already underway. Hard-takeoff proponents therefore focus on indicators that appear closer to the source of recursive improvement itself. [Reddit]reddit.comRedditIf we are in a fast-takeoff world, how long until this is…Fast takeoffs preclude many warning signs: Sustained economic growth s…
Several warning signs recur across the literature and debate.
AI starts accelerating AI research
The most obvious precursor to recursive self-improvement would be AI systems becoming major contributors to AI development itself.
Today, AI already assists with coding, debugging, experimentation, literature review, and model optimisation. A stronger warning sign would be systems that independently discover new architectures, training methods, algorithms, or optimisation techniques that human researchers neither anticipated nor fully understand. Recent discussions within frontier labs increasingly focus on AI-assisted AI research as a plausible route toward recursive improvement. [Axios]axios.comBehind the Curtain: Intelligence explosionCo-founder Jack Clark predicts a greater than 60% chance that by 2028, an AI system could autonomously build a better version of itself…
A particularly significant milestone would be an AI system making improvements that materially increase its own successor’s capabilities with limited human involvement. Some researchers view this as the earliest practical form of recursive self-improvement rather than the dramatic self-rewriting scenarios often depicted in science fiction. [Axios]axios.comBehind the Curtain: Intelligence explosionCo-founder Jack Clark predicts a greater than 60% chance that by 2028, an AI system could autonomously build a better version of itself…
Capability gains become compressed
Another warning sign would be shrinking intervals between major breakthroughs.
Historically, frontier models have improved over periods measured in months or years. A concerning pattern would be increasingly short cycles between capability jumps, especially if new models begin substantially outperforming predecessors across many domains at once.
Hard-takeoff advocates often argue that acceleration matters more than absolute capability. A model that is merely impressive is less important than a trend showing that progress itself is speeding up. [Machine Intelligence Research Institute]intelligence.orgWhoever builds AGIMachine Intelligence Research InstituteYudkowsky and Christiano discuss "Takeoff Speeds"22 Nov 2021 — In the fast takeoff scenario, weake…
The challenge is that apparent acceleration may also result from ordinary factors such as increased compute, larger datasets, or engineering improvements. Distinguishing genuine recursive acceleration from normal technological progress remains difficult.
AI becomes unusually effective at strategic reasoning
Many AI safety researchers treat sophisticated situational awareness as an important threshold capability.
Situational awareness refers to a system understanding facts about itself, its developers, its deployment environment, and the consequences of its actions. Researchers have begun creating evaluations specifically designed to measure whether models can reason about oversight systems, deployment conditions, and their own role within larger organisations. [arXiv]arxiv.orgarXiv Evaluating Frontier Models for Stealth and Situational AwarenessarXivEvaluating Frontier Models for Stealth and Situational AwarenessMay 2, 2025… OpenReview The concern is not that situational awareness automatically produces danger. Rather [openreview.net]openreview.netEvaluating Frontier Models for Stealth and Situational…by M Phuong · 2025 · Cited by 35 — After extensive discussion with the authors…, a highly capable system that understands when it is being evaluated, monitored, restricted, or rewarded may possess abilities relevant to deception or strategic behaviour. Some researchers therefore view rising situational-awareness scores as a potential early warning indicator worth monitoring closely. [arXiv]arxiv.orgarXiv Evaluating Frontier Models for Stealth and Situational AwarenessarXivEvaluating Frontier Models for Stealth and Situational AwarenessMay 2, 2025… [OpenReview]openreview.netEvaluating Frontier Models for Stealth and Situational…by M Phuong · 2025 · Cited by 35 — After extensive discussion with the authors…
Dangerous capabilities appear before deployment
A recurring lesson from AI safety work is that some capabilities can emerge unexpectedly.
Frontier-model evaluation programmes increasingly test for deception, persuasion, cyber capabilities, self-proliferation, strategic planning, and self-reasoning. These are sometimes described as “dangerous capability evaluations” because they focus on abilities that could contribute to loss-of-control scenarios rather than merely improving productivity. [arXiv]arxiv.orgarXiv Evaluating Frontier Models for Stealth and Situational AwarenessarXivEvaluating Frontier Models for Stealth and Situational AwarenessMay 2, 2025…
Researchers behind these evaluations often frame them as an early-warning system. The hope is that capability thresholds can be identified before systems become powerful enough for serious misuse or autonomy. Several safety frameworks explicitly rely on the existence of a “safety buffer” between the appearance of warning signs and the emergence of genuinely dangerous systems. [arXiv]arxiv.orgarXiv Evaluating Frontier Models for Stealth and Situational AwarenessarXivEvaluating Frontier Models for Stealth and Situational AwarenessMay 2, 2025…
Whether that buffer will actually exist is one of the major unresolved questions in the takeoff-speed debate.
Why some doomers expect little warning
A central hard-takeoff argument is that warning signs may be visible only in retrospect.
In this view, the first system capable of driving explosive improvement may emerge from a world that still appears relatively normal. Conventional forecasts may continue predicting decades until transformative AI right up until the point when those forecasts fail. [Effective Altruism Forum]forum.effectivealtruism.orgConventional wisdom will still say that transformative AI is thirtyConventional wisdom will still say that transformative AI is thirty
Several considerations motivate this concern:
- Software can improve much faster than physical infrastructure.
- AI research is already heavily concentrated in a small number of organisations.
- Capability gains can be difficult to measure before they appear in real-world performance.
- Some important capabilities may emerge suddenly rather than smoothly. Machine Intelligence Research Institute [Effective Altruism Forum]forum.effectivealtruism.orgConventional wisdom will still say that transformative AI is thirtyConventional wisdom will still say that transformative AI is thirty
Hard-takeoff proponents therefore argue that relying on visible societal disruption as an alarm bell may be dangerous. By the time economic, military, or political consequences become obvious, the underlying capability transition could already have occurred.
This is one reason AI doom discussions place so much emphasis on advance preparation rather than reactive governance.
Why slow takeoff arguments still matter
Despite the attention given to FOOM scenarios, many researchers remain sceptical that intelligence explosions would unfold over hours or days.
The strongest slow-takeoff arguments point to practical bottlenecks. Even if AI systems become excellent researchers, they may still require compute, energy, hardware, experiments, data collection, and real-world testing. Progress could therefore remain constrained by physical and organisational limits rather than pure intelligence. [The sideways view]sideways-view.comThe sideways view Takeoff speedsThe sideways viewTakeoff speedsFebruary 24, 2018 — 24 Feb 2018 — Slower takeoff means faster progress. Fast takeoff is often justified by… LessWrong Another argument is historical. Most transformative technologies have diffused through society gradually enough that warning signs were visib [lesswrong.com]lesswrong.comOpen source on lesswrong.com. well before their largest effects appeared. Proponents of slower takeoff suggest that highly capable but sub-superintelligent AI systems would already reshape science, industry, and economies before any true intelligence explosion occurred. [AI Alignment Forum]alignmentforum.orgdistinguishing definitions of takeoffAI Alignment ForumDistinguishing definitions of takeoff13 Feb 2020 — A fast takeoff is one that occurs over the timescale of minutes, hou…
Under this view, warning signs would be abundant: [arxiv.org]arxiv.orgarXivWhat AI evaluations for preventing catastrophic risks can…26 Nov 2024 — The hope is that this gap would give evaluators time to d…
- Rapid automation of cognitive labour.
- AI-led scientific breakthroughs.
- AI systems conducting significant portions of research programmes.
- Sustained productivity growth across multiple sectors.
- Increasing dependence on AI decision-making. [AI Alignment Forum]alignmentforum.orgdistinguishing definitions of takeoffAI Alignment ForumDistinguishing definitions of takeoff13 Feb 2020 — A fast takeoff is one that occurs over the timescale of minutes, hou…
The disagreement is therefore not necessarily about whether warning signs exist, but about how much time separates those signs from potentially irreversible outcomes.
The hardest forecasting problem
The deepest uncertainty is that no one has observed a genuine intelligence explosion.
The evidence available today comes from theoretical arguments, historical analogies, trend analysis, capability evaluations, and observations of current AI progress. Researchers disagree not only about timelines but about what should count as a meaningful warning sign in the first place. [LessWrong]lesswrong.comai takeoff30 Dec 2024 — A hard takeoff (or an AI going "FOOM") refers to AGI expansion in a matter of minutes, days, or months. It is a fast, abrup… [Machine Intelligence Research Institute]intelligence.orgWhoever builds AGIMachine Intelligence Research InstituteYudkowsky and Christiano discuss "Takeoff Speeds"22 Nov 2021 — In the fast takeoff scenario, weake…
Some point to AI systems increasingly contributing to AI development as the most important signal. Others focus on dangerous-capability evaluations, deception-related behaviours, or emerging situational awareness. Still others argue that economic and scientific transformation will provide ample notice before any loss-of-control scenario becomes plausible. [arXiv]arxiv.orgarXiv Evaluating Frontier Models for Stealth and Situational AwarenessarXivEvaluating Frontier Models for Stealth and Situational AwarenessMay 2, 2025… [axios]axios.comBehind the Curtain: Intelligence explosionCo-founder Jack Clark predicts a greater than 60% chance that by 2028, an AI system could autonomously build a better version of itself… For AI doom discussions, this uncertainty matters because p(doom) estimates often depend heavily on takeoff assumptions. A world with years of warning and repeated opportunities for intervention looks very different from a world where critical capability thresholds are crossed in a matter of weeks. The debate over FOOM timelines is therefore not just about forecasting technological progress. It is about estimating how much opportunity humanity would have to recognise danger and respond before control becomes difficult or impossible.
Amazon book picks
Further Reading
Books and field guides related to What would warn US before FOOM?. Use these as the next step if you want deeper reading beyond the article.
Human Compatible
Discusses warning signs and control challenges before superintelligence.
eBay marketplace picks
Marketplace Samples
Example marketplace items related to this page. Use the search link to explore similar finds on eBay.
Endnotes
-
Source: ai-safety-atlas.com
Title: AI Safety Atlas Takeoff
Link: https://ai-safety-atlas.com/chapters/v1/capabilities/takeoff/Source snippet
AI Safety AtlasTakeoff - Chapter 1Fast takeoff describes scenarios where AI capabilities increase dramatically over very short periods...
-
Source: lesswrong.com
Title: [ai takeoff]({{ ‘ai-takeoff/’ | relative_url }})
Link: https://www.lesswrong.com/w/ai-takeoff?lw_source=import_sheet&revision=0.0.24Source snippet
30 Dec 2024 — A hard takeoff (or an AI going "FOOM") refers to AGI expansion in a matter of minutes, days, or months. It is a fast, abrup...
-
Source: intelligence.org
Title: Whoever builds AGI
Link: https://intelligence.org/2021/11/22/yudkowsky-and-christiano-discuss-takeoff-speeds/Source snippet
Machine Intelligence Research InstituteYudkowsky and Christiano discuss "Takeoff Speeds"22 Nov 2021 — In the fast takeoff scenario, weake...
-
Source: reddit.com
Link: https://www.reddit.com/r/accelerate/comments/1ijd1gx/are_we_heading_for_a_hard_takeoff_how_do_you/Source snippet
RedditAre we heading for a hard takeoff? How do you think it...Personally, I think it will be a hard takeoff in terms of self-recursive...
-
Source: reddit.com
Link: https://www.reddit.com/r/slatestarcodex/comments/1id2k1x/if_we_are_in_a_fasttakeoff_world_how_long_until/Source snippet
RedditIf we are in a fast-takeoff world, how long until this is...Fast takeoffs preclude many warning signs: Sustained economic growth s...
-
Source: axios.com
Title: Behind the Curtain: Intelligence explosion
Link: [https://www.axios.com/2026/05/07/anthropicSource snippet
Co-founder Jack Clark predicts a greater than 60% chance that by 2028, an AI system could autonomously build a better version of itself...
-
Source: arxiv.org
Title: arXiv Evaluating Frontier Models for Stealth and Situational Awareness
Link: https://arxiv.org/abs/2505.01420Source snippet
arXivEvaluating Frontier Models for Stealth and Situational AwarenessMay 2, 2025...
Published: May 2, 2025
-
Source: openreview.net
Link: https://openreview.net/forum?id=Daqfkiy6z4Source snippet
Evaluating Frontier Models for Stealth and Situational...by M Phuong · 2025 · Cited by 35 — After extensive discussion with the authors...
-
Source: arxiv.org
Title: arXiv Evaluating Frontier Models for Dangerous Capabilities
Link: https://arxiv.org/abs/2403.13793Source snippet
arXivEvaluating Frontier Models for Dangerous CapabilitiesMarch 20, 2024...
Published: March 20, 2024
-
Source: ai-safety-atlas.com
Title: AI Safety Atlas Dangerous Capability Evaluations
Link: https://ai-safety-atlas.com/chapters/v1/evaluations/dangerous-capability-evaluations/Source snippet
Dangerous Capability Evaluations - Chapter 5Dangerous capability evaluations specifically probe for these potentially harmful abilities...
-
Source: arxiv.org
Link: https://arxiv.org/html/2412.08653v1Source snippet
arXivWhat [AI evaluations]({{ 'ai-evaluations/' | relative_url }}) for preventing catastrophic risks can...26 Nov 2024 — The hope is that this gap would give evaluators time to d...
-
Source: arxiv.org
Link: https://arxiv.org/abs/2412.15433 -
Source: sideways-view.com
Title: The sideways view Takeoff speeds
Link: https://sideways-view.com/2018/02/24/takeoff-speeds/Source snippet
The sideways viewTakeoff speedsFebruary 24, 2018 — 24 Feb 2018 — Slower takeoff means faster progress. Fast takeoff is often justified by...
Published: February 24, 2018
-
Source: lesswrong.com
Link: https://www.lesswrong.com/s/QyQcBpSur9SFyRuvB/p/AfGmsjGPXN97kNp57Source snippet
Arguments about fast takeoff24 Feb 2018 — The basic case for slow takeoff is: "it's easier to build a crappier version of something" + "a...
-
Source: lesswrong.com
Title: yudkowsky and christiano discuss takeoff speeds
Link: https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speedsSource snippet
LessWrongYudkowsky and Christiano discuss "Takeoff Speeds"22 Nov 2021 — This is a transcription of Eliezer Yudkowsky responding to Paul C...
-
Source: reddit.com
Link: https://www.reddit.com/r/OpenAI/comments/1phbgeg/stuart_russell_says_ai_companies_now_worry_about/Source snippet
AI with an IQ of 150 could improve its own algorithms to reach 170...
-
Source: situational-awareness.ai
Link: https://situational-awareness.ai/wp-content/uploads/2024/06/situationalawareness.pdfSource snippet
pdf6 Jun 2024 — Rather than iteratively encountering increasingly more dangerous safety failures in the wild, the first notable safety fa...
-
Source: alignmentforum.org
Title: distinguishing definitions of takeoff
Link: https://www.alignmentforum.org/posts/YgNYA6pj2hPSDQiTE/distinguishing-definitions-of-takeoffSource snippet
AI Alignment ForumDistinguishing definitions of takeoff13 Feb 2020 — A fast takeoff is one that occurs over the timescale of minutes, hou...
-
Source: forum.effectivealtruism.org
Title: Conventional wisdom will still say that transformative AI is thirty
Link: https://forum.effectivealtruism.org/posts/x3MSTGhxChEZsRvwB/yudkowsky-and-christiano-on-ai-takeoff-speeds-linkpost -
Source: alignmentforum.org
Title: takeoff speeds have a huge effect on what it means to work 1
Link: https://www.alignmentforum.org/posts/hRohhttbtpY3SHmmD/takeoff-speeds-have-a-huge-effect-on-what-it-means-to-work-1Source snippet
Whether AI is a...
Additional References
-
Source: simonlermen.substack.com
Link: https://simonlermen.substack.com/p/the-term-recursive-self-improvementSource snippet
The Term Recursive Self-Improvement Is Often Used IncorrectlyThe term Recursive Self-Improvement (RSI) now seems to get used sometimes fo...
-
Source: longtermrisk.org
Link: https://longtermrisk.org/against-gdp-as-a-metric-for-ai-timelines-and-takeoff-speeds/Source snippet
Center on Long-Term RiskAgainst GDP as a metric for [AI timelines]({{ 'timeline-effects/' | relative_url }}) and takeoff speeds30 Dec 2020 — Takeoff Speeds: Paul Christiano argues f...
-
Source: cognitiverevolution.ai
Title: situational awareness in government with uk aisi chief scientist geoffrey irving
Link: https://www.cognitiverevolution.ai/situational-awareness-in-government-with-uk-aisi-chief-scientist-geoffrey-irving/Source snippet
Situational Awareness in Government, with UK AISI Chief...1 Mar 2026 — pre-release frontier model evaluation for dangerous capabilities...
-
Source: x.com
Link: https://x.com/jarokrolewski/status/2008223852040437930Source snippet
s that make them potentially dangerous. And here's where...Read more...
-
Source: medium.com
Link: https://medium.com/agi-is-living-intelligence/situational-awareness-in-ai-thats-just-how-minds-work-637fac171f8fSource snippet
context of interaction. Is this dangerous? Only if we've...Read more...
-
Source: linkedin.com
Link: https://www.linkedin.com/pulse/future-ai-why-recursive-self-improvement-may-lead-hard-gary-ramah-grn5cSource snippet
ystems will reach a level of sophistication where they can...
-
Source: planned-obsolescence.org
Title: takeoff speeds rule everything around
Link: https://www.planned-obsolescence.org/p/takeoff-speeds-rule-everything-aroundSource snippet
me12 Feb 2026 — The classic definition of takeoff speed is the amount of time it takes to go from AGI to superintelligence,2 but both of...
-
Source: futureoflife.org
Link: https://futureoflife.org/wp-content/uploads/2025/11/Indicator-Dangerous_Capability_Evaluations.pdfSource snippet
gerous dual-use capabilities. It remains...Read more...
-
Source: youtube.com
Title: The AI Intelligence Explosion: Why Recursive Self-Improvement Changes Everything
Link: https://www.youtube.com/watch?v=xRQQFCWhobcSource snippet
Recursive Self-Improvement: The AI Feedback Loop Behind the “Intelligence Explosion”...
-
Source: aisecurityandsafety.org
Link: https://aisecurityandsafety.org/en/guides/ai-situational-awareness/Source snippet
AI Security & Safety DirectoryAI Situational Awareness: The Complete Guide (2026)25 Mar 2026 — Some researchers argue that strong situati...
Topic Tree