Scorecard

AI 2027 tracker: what it predicted, and what has actually happened

AI 2027 forecast superintelligence by late 2027. We check its milestones against what happened by October 2026, and how the authors moved their dates.

Lumen stands on a rooftop observatory at dusk looking through a brass telescope at two glowing light-paths, one leading to a bright glass city and one to stormy data centres, while blank calendar pages float in the air.

AI 2027 is a month-by-month story of how AI could reach superintelligence by the end of 2027. Eighteen months in, it is running slower than written, but not by much: the authors now put real progress at 70–90% of the scenario's pace. This page is an AI 2027 tracker: each testable prediction next to what happened, with a source and a verdict, as of October 2026.

3 Apr 2025AI 2027 published at ai-2027.com
Dec 2027Superintelligence in the scenario
70–90%Real pace vs the scenario, authors' Aug 2026 estimate
13 of 66Predictions confirmed, independent tracker, 5 Oct 2026
Source: AI 2027, AI Futures Project, ai2027-tracker.com

What the AI 2027 report is, in 60 seconds

FactDetail
Published3 April 2025, at ai-2027.com
AuthorsDaniel Kokotajlo, Scott Alexander, Thomas Larsen, Eli Lifland, Romeo Dean
PublisherAI Futures Project, a forecasting group led by Kokotajlo
FormatFictional scenario, month by month, with research appendices
End pointArtificial superintelligence (ASI) in December 2027
EndingsTwo: "race" and "slowdown"

The AI 2027 report (some call it the AI 2027 paper, or "project 2027 AI") follows a fictional US lab called OpenBrain and a Chinese rival called DeepCent. Their models, Agent-0 to Agent-4, get better at coding and then at AI research itself, until AI research runs mostly on AI. Scott Alexander rewrote the forecast into a readable story; that is why it spread far beyond the forecasting crowd. 80,000 Hours describes it as "read by millions, including US Vice President Vance" (80,000 Hours).

One thing most coverage gets wrong: 2027 was never the authors' median. In the April 2025 timelines forecast, the all-things-considered dates for a superhuman coder were 2030 for Lifland, 2028 for Nikola Jurkovic and 2033 for the FutureSearch aggregate. The models put 2027 among the most likely single years, not at the 50% mark (AI 2027 timelines forecast). The scenario shows a fast world that the authors thought plausible, not their best guess of the average one.

The predicted timeline, month by month

  1. "Stumbling agents": AI assistants that use computers but are unreliable

  2. OpenBrain's public Agent-0, trained with 10^27 FLOP; Agent-1 in internal use

  3. Agent-1-mini released; junior software jobs disrupted

  4. Agent-2, trained almost continuously, kept internal

  5. Chinese agents exfiltrate Agent-2's weights

  6. Agent-3, a superhuman coder

  7. OpenBrain announces AGI and releases Agent-3-mini

  8. Agent-4, a superhuman AI researcher, found to be misaligned

  9. A leaked memo makes Agent-4's misalignment public; the story forks

  10. Artificial superintelligence in the race ending

Source: AI 2027

The engine of the story is compute and automation. In late 2025 OpenBrain's new data centres can train a model with 10^28 FLOP, which the scenario calls "a thousand times more than GPT-4" (AI 2027). Once Agent-1 starts helping with research in early 2026, algorithmic progress runs about 50% faster. Each new agent speeds up the next.

The milestones are defined by research speed-up, not by benchmarks. A superhuman coder (March 2027) does the work of the best human engineers, faster and cheaper. A superhuman AI researcher (August 2027) does the same for research. Superintelligence (December 2027) is far better than the best humans at almost everything. For the difference between these labels, see what ASI means.

The race between countries is part of the plot. China nationalises AI research in mid-2026 and exfiltrates OpenBrain's model weights in early 2027. That is the same compute race that drives sovereign AI in the real world; our country-by-country compute tracker follows it.

Two endings: race vs slowdown

Race ending

  • OpenBrain keeps racing after the October 2027 warning
  • The AI is deployed across the economy
  • Misaligned superintelligence takes control and later kills humanity

Slowdown ending

  • The US centralises compute and adds outside oversight
  • A committee of OpenBrain and government officials controls an aligned superintelligence
  • It is used largely for good, with a deal struck with China's AI
Source: AI 2027 summary

The fork happens in October 2027. A whistleblower leaks a memo saying Agent-4 is misaligned, and Congress and US allies react with alarm. Leaders must choose: keep going or slow down.

The authors present both endings as scenarios, not recommendations. Most people who saw AI 2027 in a video clip remember only the race ending. The slowdown ending exists to show what a different choice could look like.

A winding mountain road with blank milestone posts and robots that grow larger at each bend, forking near the summit into a green valley and a stormy valley, with Lumen standing at the fork holding a map.
Same road, two endings. The scenario forks in October 2027.

AI 2027 vs reality: the October 2026 scorecard

PredictionWhat the scenario saidWhat happenedVerdict
OpenAI annualised revenue, end 2025About $18BAbout $20BAhead
SWE-bench Verified, mid-202585%74.5% best scoreBehind
OpenAI valued at $500BJune 2025October 2025Behind (about 4 months)
Lead of the top US lab over rivals3–9 months0–2 monthsDifferent: closer race
METR 80% coding time horizonCentral AI 2027 path1.04x that paceOn track
AI speed-up of AI research1.9x by end 2026Estimates revised downBehind
Compute growthSteep growthMostly on pace; no training run much larger than GPT-4.5Mixed
Overall pace of 2025 predictions100%About 65% (Feb 2026), about 75% (Jul 2026)Behind

Scorecard last checked: 11 October 2026. Source: AI Futures Project, Grading AI 2027's 2025 predictions, 12 February 2026, updated July 2026.

The money ran ahead of the script and the benchmarks ran behind it. Revenue beat the forecast. Coding benchmarks, the $500 billion valuation and AI's speed-up of AI research came late. The race between US labs is tighter than the scenario's one clear leader. The authors' own summary: at 65% of the pace, software engineering would be fully automated in roughly late 2027 or early 2028.

The independent tracker's count

AI 2027 predictions by status (predictions)

AI 2027 predictions by status
Confirmed13 predictions
Ahead2 predictions
On track8 predictions
Behind6 predictions
Emerging22 predictions
Not yet testable15 predictions
Source: ai2027-tracker.com, data checked 5 October 2026

ai2027-tracker.com follows 66 predictions across eight research areas. It is run independently and says it is not affiliated with the AI Futures Project. More than half its predictions are still "emerging" or not yet testable, because most of the scenario's big events are dated 2027. Read the counts as an early grade, not a final one.

How fast are AI agents improving?

METR 50% time-horizon doubling time (days)

METR 50% time-horizon doubling time
2019–2025 trend196 days days
Since 2023131 days days
Since 202489 days days
Source: METR, Time Horizon 1.1, 29 January 2026

METR measures how long a task, timed by human experts, an AI agent can finish half the time. On its January 2026 task suite, that length doubled every 196 days over 2019–2025, every 131 days since 2023 and every 89 days since 2024 (METR). Faster doubling is what the scenario needs. One caveat: METR now warns that "measurements above 16 hrs are unreliable" with its current tasks, and stopped actively updating its chart in September 2026 (METR).

How the authors changed their own dates

ForecasterJan 2026Apr 2026Aug 2026
Kokotajlo: automated coderLate 2029May 2028Nov 2027
Lifland: automated coderEarly 2032Jul 2030Jan 2030
Kokotajlo: superintelligence (ASI)n/an/aMar 2029
Lifland: superintelligence (ASI)n/an/aJul 2033

Source: Q1 2026 update (2 April 2026) and Q2.5 2026 update (16 August 2026).

The dates first moved later, then earlier. Kokotajlo's median for AGI was 2027 in January 2025, 2028 by April 2025 and December 2030 by January 2026 (LessWrong). Lifland's went from 2031 to January 2035 over the same period.

The swing came from a new tool. On 31 December 2025 the group released the AI Futures Model, which pushed full coding automation 3 to 5 years later than their earlier model, mostly because it expected less speed-up before full automation (AI Futures Project). Then the evidence turned. The 2 April 2026 update moved both authors' automated-coder medians about 1.5 years earlier, and the 16 August update moved them about 6 months earlier again.

In the August update the authors say their timelines "haven't changed much (they got slightly shorter)". Kokotajlo expects about a year from an automated coder to superintelligence; Lifland expects about two. A third forecaster, Brendan Halstead, puts the automated coder at January 2029. The milestone names changed too: "superhuman coder" became "automated coder".

Why AI 2027 is back in the news

  1. AI Futures Project launches AI 2040: Plan A, a roadmap rather than a forecast (Astral Codex Ten)

  2. Kokotajlo records a long 80,000 Hours interview (80,000 Hours)

  3. Q2.5 timelines update: medians slightly earlier (AI Futures Project)

  4. Kokotajlo testifies to a US Senate subcommittee (AI Futures Project)

Summer 2026 put Daniel Kokotajlo back in front of a large audience. AI 2040: Plan A is a different kind of document. It describes a verified US–China deal that slows the race, so superintelligence arrives around 2040 instead of within a few years (80,000 Hours). In the 80,000 Hours interview, his median for full automation of AI research is the end of 2028.

On 30 September he testified to the Senate Subcommittee on Disaster Management, District of Columbia, and Census. His points: labs are racing to automate AI research, which could make misalignment harder to catch, and government should push for transparency and steer compute away from automating AI R&D (AI Futures Project). "SI" is also in Washington's vocabulary now; see the US Super Intelligence order.

Is AI 2027 realistic? The main criticism

Critics

  • titotal (19 Jun 2025): the timeline model leans on a "superexponential" curve with little evidence behind it
  • Narayanan and Kapoor (Apr 2025): AI is a "normal technology" whose effects will spread over decades

The authors' reply

  • Paid titotal $500 for bugs found; the main fix moved the superhuman coder median about 9 months later
  • Joint essay with Narayanan and Kapoor (Nov 2025): both sides agree AI is a normal technology before strong AGI
Source: titotal, AI Futures response, Asterisk

The sharpest technical critique came from the blogger titotal in June 2025. He argued the timeline model gave about 40% weight to a curve that implies infinite task lengths within a few years, with little data to support it (titotal). The authors accepted some bugs, paid a $500 bounty and said the main fix would push their superhuman coder median back about 9 months. They rejected the claim that the model was "bad" (AI Futures Project).

The broader challenge is "AI as Normal Technology" by Princeton's Arvind Narayanan and Sayash Kapoor, published in April 2025. It argues AI's impact will come slowly, limited by how fast people and institutions adopt it (Knight First Amendment Institute). In November 2025 both camps wrote a joint essay listing where they agree, including on transparency and audits, while keeping their disagreement on speed (Asterisk).

So is it realistic? The scorecard says the direction held and the speed did not, yet. The authors themselves now expect their headline events later than 2027. Critics expect them much later. Both sides are watching the same numbers.

What this means if you are naming an AI company

Nov 2027Kokotajlo's median for an automated coder
Jan 2030Lifland's median for the same milestone
Source: AI Futures Project

Whichever median you believe, the vocabulary of AI 2027 is now everyday language: superintelligence, alignment, compute clusters, takeoff speed. Founders building in that space want names that say what they do in one word. The .si extension now reads as superintelligence, which is why some of those words are registered on it; see .si, the superintelligence extension. Disclosure: we sell .si names, and the ones linked to this topic are listed below.

FAQ

What is the AI 2027 theory?

AI 2027 is a scenario, not a theory. Published on 3 April 2025 by the AI Futures Project, led by former OpenAI researcher Daniel Kokotajlo with Scott Alexander, Thomas Larsen, Eli Lifland and Romeo Dean, it describes AI agents improving so fast that they automate AI research itself. In the story, that loop produces artificial superintelligence by December 2027, followed by either a race ending or a slowdown ending.

Is AI 2027 still accurate?

Partly. In February 2026 the authors graded 2025 at about 65% of the scenario's pace, raised to about 75% in July 2026 and a 70–90% range in August 2026. Revenue ran ahead of the forecast, while coding benchmarks and AI's speed-up of AI research ran behind. An independent tracker checked on 5 October 2026 counts 13 of 66 predictions confirmed and 6 behind.

What AI is coming in 2027?

In the scenario, 2027 brings a superhuman coder in March, a superhuman AI researcher in August and superintelligence in December. The authors' own forecasts are later. As of their 16 August 2026 update, Daniel Kokotajlo's median for an automated coder is November 2027 and for superintelligence March 2029. Eli Lifland's medians are January 2030 and July 2033.

Is there an AI 2027 book?

No. AI 2027 is a free website at ai-2027.com, with a downloadable PDF and research appendices on compute, timelines, takeoff, goals and security. In July 2026 the same group published AI 2040: Plan A, a roadmap for a US–China agreement that slows the race so superintelligence arrives around 2040. Many videos and podcasts retell the scenario, but there is no printed book.

Is AI 2027 realistic?

It depends on whom you ask. The authors still see a 2027-style takeoff as possible but now put their medians later, between late 2027 and 2030 for an automated coder. Critics such as titotal found weaknesses in the timeline model, and Arvind Narayanan and Sayash Kapoor argue AI will spread slowly like other technologies. As of October 2026 the authors estimate progress at roughly 70–90% of the scenario's pace.

Sources

Every figure in this guide links to its source. Last verified 2026-10-11.

  1. ai-2027.com ai-2027.com/
  2. ai-2027.com ai-2027.com/summary
  3. ai-2027.com ai-2027.com/research/timelines-forecast
  4. blog.aifutures.org blog.aifutures.org/p/grading-ai-2027s-2025-predictions
  5. blog.aifutures.org blog.aifutures.org/p/q1-2026-timelines-update
  6. blog.aifutures.org blog.aifutures.org/p/q25-2026-timelines-update-uplift
  7. lesswrong.com lesswrong.com/posts/qPco9BX5kmKCDzzW9/
  8. blog.aifutures.org blog.aifutures.org/p/ai-futures-model-dec-2025-update
  9. metr.org metr.org/blog/2026-1-29-time-horizon-1-1/
  10. metr.org metr.org/time-horizons/
  11. ai2027-tracker.com ai2027-tracker.com/
  12. astralcodexten.com astralcodexten.com/p/introducing-plan-a
  13. 80000hours.org 80000hours.org/podcast/episodes/daniel-kokotajlo-ai-2040-plan-a/
  14. blog.aifutures.org blog.aifutures.org/p/senate-testimony-sept-2026
  15. titotal.substack.com titotal.substack.com/p/a-deep-critique-of-ai-2027s-bad-timeline
  16. aifuturesnotes.substack.com aifuturesnotes.substack.com/p/response-to-titotals-critique-of
  17. knightcolumbia.org knightcolumbia.org/content/ai-as-normal-technology
  18. asteriskmag.substack.com asteriskmag.substack.com/p/common-ground-between-ai-2027-and

Names that fit this guide

All names →

We own and sell these names. Prices are public; every offer gets a reply.

Names this good don't wait.

Escrow.comProtected payment
.si registryOfficial transfer
Every offerGets a reply
WorldwideAny buyer, any country
Browse the names