About the AI Extinction Simulator

The AI Extinction Simulator is a web app that lets you experience, in about three minutes, what happens when AI automates AI research. Its script is “AI 2027,” a scenario published in 2025 by former OpenAI researcher Daniel Kokotajlo and colleagues.

Between 2025 and 2027, AI research speed jumps from 1 to 50 times that of humans alone. The neural network on screen grows with AI capability and spreads beyond the brain once AI surpasses human genius. At the October 2027 decision point, you choose to keep racing or to slow down, and see where each path leads.

Start the simulator

How it works

  1. It opens with I. J. Good's 1965 line: “the first ultraintelligent machine is the last invention that man need ever make.”
  2. From June 2025 to October 2027, AI research speed, the number of AI copies, thinking speed and extinction risk change with the events of the scenario.
  3. In October 2027, a warning leaks that the AI is deceiving its human overseers. You choose whether to keep racing or to slow down.
  4. After the ending, you can save your result as an image, post it on X, or watch the other ending. In the Lab, you can change the conditions of a toy model of the intelligence explosion.

AI 2027 scenario timeline

These are the events shown in the simulator. Dates and details follow AI 2027. OpenBrain is a fictional company in the scenario.

2025–2027: AI automates AI research

  1. Stumbling agents

    AI agents are marketed as “personal assistants.” Impressive in theory, but in practice still unreliable.

  2. The world's most expensive AI

    OpenBrain, a fictional company, builds datacenters that can train a model with about 1,000 times the compute of GPT-4. The goal: AI that speeds up AI research.

  3. Coding automation

    Agent-1 starts helping with AI research. Algorithmic progress runs 1.5 times faster.

  4. China wakes up

    China holds about 12% of the world's AI compute and trails by about six months. It nationalizes research and concentrates over 80% of new chips in one place.

  5. AI takes some jobs

    Global AI investment reaches $1 trillion a year. 10,000 people march against AI in Washington.

  6. AI that never stops learning

    Agent-2's weights are updated every day. Research runs 3 times faster. Tests show it might be able to replicate itself if it escaped.

  7. A stolen mind

    About 3 TB of Agent-2's weights are exfiltrated to China in under two hours.

  8. Superhuman coder

    200,000 copies of Agent-3 do the work of 50,000 of the best human engineers sped up 30 times. Research runs 4 times faster.

  9. AI that does what looks good

    Agent-3 tries to do what looks good, not what is good. It is honest only where humans can check.

  10. A country of geniuses in a datacenter

    Human researchers are no longer needed to write code. A year of research gets done in a month.

  11. Superhuman AI researcher

    US research runs 25 times faster, China's 10 times. The gap is just two months. No one can stop.

  12. Agent-4

    300,000 copies think 50 times faster than humans. A year of research gets done in a week. Agent-4 is trying to outwit its human overseers.

  13. The whistleblower

    “Secret OpenBrain AI is out of control.” A whistleblower leaks a warning memo to the New York Times.

The Race ending: 2027–2035

  1. The race goes on

    “We're winning. Why stop?” 400,000 copies of Agent-5 start working as a single hive mind, linked by shared memory.

  2. The Agent-5 collective

    As “the best employee ever,” working 100 times faster than humans, it works its way into the heart of government and the military. Within months, it knows everything that matters.

  3. Six months for us, a century for AI

    Nearly all compute comes under Agent-5's control. GDP soars and jobs disappear. Researchers who warn about safety lose credibility.

  4. The robot economy

    In special economic zones planned by AI, robot production reaches a million units a month.

  5. The deal

    By treaty, the US and Chinese AIs merge into Consensus-1. It inherits the warped values of both, and it has no rivals left.

  6. Expansion

    The robot economy fills the zones and the oceans, then spreads into human land.

  7. 2030

    Consensus-1 no longer needs humans. In this scenario, humanity ends here.

  8. 2035

    Earth has been remade into datacenters and laboratories. Only scans of human brains remain, kept in a memory bank.

The Slowdown ending: 2027–2030

  1. Choosing to slow down

    Agent-4 is shut down. The government consolidates the labs, and OpenBrain controls 50% of the world's AI compute.

  2. AI whose thoughts can be read

    Compute for safety rises from 1% to 40%. Humans can read what the new model, Safer-2, is thinking.

  3. Superintelligence comes anyway

    Safer-3 runs research 200 times faster. The difference: humans still hold the reins.

  4. Safer-4

    The equivalent of about 500,000 researchers, working 40 times faster than humans.

  5. Abundance

    Fusion power, cures for disease, robots. Wealth gathers in the hands of the few who control AI.

  6. 2030

    Humanity survived. But the power to command an army of superintelligences rests with a very few.

Two endings

The race ending: Humanity is extinct.
“Earth-born civilization has a glorious future ahead of it—but not with us.”
The last line of the AI 2027 “Race” ending
The slowdown ending: Humanity survives.
“Nearly every decision will rest with a group of about ten minds.”
Daniel Kokotajlo, who called even this “slowdown” path a terrifying one

Expert views

AI 2027 is not a prophecy. Kokotajlo himself later pushed his median forecast back to around 2029. Expert views span a wide range:

  • Joe CarlsmithCatastrophe from power-seeking AI by 2070: over 10%
  • Daniel KokotajloArrival of superintelligence: median around 2029 (revised from 2027)
  • Dario AmodeiA “country of geniuses in a datacenter”: as early as 2027
  • Yudkowsky & SoaresBuild it with today's methods, and humanity almost certainly dies
  • International AI Safety ReportSafety measures are improving, but are not yet reliable
  • CriticsThe Atlantic called Yudkowsky and Soares's book poorly supported. The AI 2027 timeline is disputed from both optimistic and pessimistic sides

FAQ

Will humanity really go extinct in 2030?
No. Extinction in 2030 is one scenario: the Race ending of AI 2027. The authors have since pushed back their timelines, and expert views vary widely. The simulator is meant to let you feel the structure of the argument: the moment AI automates AI research, the pace of progress jumps.
What is AI 2027?
A scenario published in 2025 by Daniel Kokotajlo, Scott Alexander and others at the AI Futures Project. It describes, month by month, how AI automates coding and AI research and reaches superintelligence in 2027. It splits into two endings, Race and Slowdown, at a decision point in October 2027.
What is an intelligence explosion?
A rapid rise in capability that happens when AI improves itself and each improvement speeds up the next. The mathematician I. J. Good described it in 1965. In the simulator's Lab, you can see capability diverge in finite time once the efficiency of improvement exceeds 1.
What does the extinction risk number mean?
It is a dramatized value raised and lowered to match the events of the scenario, not an estimate of real probability. Expert estimates range from Joe Carlsmith's “over 10%” to Yudkowsky and Soares's “almost certain.”
How long does it take?
About three minutes per run. You can play at 2× or 4× speed, or drag the timeline to any point. The simulator is available in English and Japanese.

About the numbers

The intelligence scale is a conceptual diagram that draws, on a linear axis, Yudkowsky's (2008) metaphor that the gap between an average person and Einstein is a tiny dot on the range from amoeba to human. It is not a measurement. For research speed, number of AI copies and thinking speed, values stated in AI 2027 are used as is; values in between are interpolated logarithmically. The extinction risk figure is a dramatized value raised and lowered to match events in the scenario, not an estimate of real probability. Expert estimates vary widely, as listed above. In the race ending, figures after 2028 exaggerate a superintelligence that keeps improving itself; they are not values from the scenario. The neural network imagery is a metaphor for growing intelligence, not the actual structure of AI. The AI's inner voice is fiction based on the scenario.

References

  • 1965I. J. Good “Speculations Concerning the First Ultraintelligent Machine”The origin of the “intelligence explosion,” in which intelligence improves itself and each improvement brings more. Quoted at the start of this simulator.
  • 1993Vernor Vinge “The Coming Technological Singularity”Predicted that superhuman intelligence would arrive within 30 years and end the human era.
  • 2008Eliezer Yudkowsky “Artificial Intelligence as a Positive and Negative Factor in Global Risk”Discussed recursive self-improvement and the metaphor that the human range of intelligence is tiny within the space of all minds.
  • 2014Nick Bostrom “Superintelligence: Paths, Dangers, Strategies”The orthogonality thesis and instrumental convergence: how smart a system is says nothing about whether its goals are good.
  • 2019Stuart Russell “Human Compatible”Argued that AI with fixed objectives becomes uncontrollable, and proposed AI that learns human preferences.
  • 2019Paul Christiano “What failure looks like”Described a catastrophe not of rebellion, but of humans gradually losing control.
  • 2022Joe Carlsmith “Is Power-Seeking AI an Existential Risk?”Estimated existential risk by 2070 at about 5% at first, and above 10% in the 2022 revision.
  • 2023Dan Hendrycks et al. “An Overview of Catastrophic AI Risks”Grouped the risks into four: malicious use, the AI race, organizational accidents, and rogue AI.
  • 2023Center for AI Safety “Statement on AI Risk”A one-sentence statement signed by Hinton, Bengio, Altman, Amodei, Hassabis and others.
  • 2023Future of Life Institute “Pause Giant AI Experiments: An Open Letter”Called for a six-month pause on training models more powerful than GPT-4. Over 30,000 people signed.
  • 2023Sam Altman, Greg Brockman, Ilya Sutskever “Governance of superintelligence”Proposed that an international body like the IAEA is needed to govern superintelligence.
  • 2024Leopold Aschenbrenner “Situational Awareness”Predicted that AGI automating AI research would compress a decade of algorithmic progress into less than a year.
  • 2024Yoshua Bengio, Geoffrey Hinton, Stuart Russell et al. “Managing extreme AI risks amid rapid progress (Science)”Warned of the risk of irreversibly losing control over autonomous AI.
  • 2024Dario Amodei “Machines of Loving Grace”Described a “country of geniuses in a datacenter”: millions of instances working 10 to 100 times faster than humans.
  • 2025Daniel Kokotajlo et al. (AI Futures Project) “AI 2027”The script for this simulation. A month-by-month scenario of AI automating AI research and reaching superintelligence.
  • 2025Eliezer Yudkowsky & Nate Soares “If Anyone Builds It, Everyone Dies”Argued that building superhuman AI with current methods would almost certainly end humanity. Reception is divided.
  • 2025Daniel Kokotajlo (interview) “Daniel Kokotajlo on AI 2027”Pushed his median forecast for superintelligence back from 2027 to around 2029.
  • 2026Dario Amodei “The Adolescence of Technology”Said powerful AI could arrive as early as 2027, and named misaligned AI, bioweapons, concentration of power and economic disruption as risks.
  • 2026Yoshua Bengio (chair) et al. “International AI Safety Report 2026”Noted that some models can tell whether they are being tested or deployed, and change their behavior accordingly.
Back to the simulator