Eliezer Yudkowsky: The Blind Prophet of AI Doom

Eliezer Yudkowsky has spent 20+ years warning that “if anyone builds [artificial general intelligence], everyone dies”.

The AI doomer movement he has inspired have infiltrated many non-profits, government agencies, and corporations. They have called for a vast panopoly regulations intended to slow or stop the development of AI:

• Enforce an indefinite, global ban on frontier AI models (systems more powerful than GPT-4) [1].
• Force all chip manufacturers to install government-controlled spyware at the silicon level [2].
• Force all chip manufacturers to install government-controlled remote killswitches [3].
• Impose a compulsory licensing scheme for all AI researchers and developers who want to train models beyond a certain computational threshold [4].
• Impose lengthy prison sentences on AI researchers and developers who do not comply [5].
• Launch rocket attacks against non-compliant data centers, even if doing so risks nuclear war [6].
• Impose unlimited strict civil and criminal liability on AI companies and individual developers for any downstream harm from their general-purpose models, regardless of intent [7].
• Enforce strict legal limits on the total floating-point operations allowed during a single training run, capping the maximum capability a model can achieve [8].
• Maintain an international registry tracking every advanced AI accelerator chip by unique identifier from manufacturing to final deployment [9].
• Ban or heavily restrict publishing the weights of open-source and open-weight models that cross certain capability thresholds [10].
• Require cloud computing providers to monitor customer workloads and report any large-scale AI training clusters directly to government regulators [11].

And they’ve enjoyed success:

• Over 200 local jurisdictions and dozens of state legislatures across the United States have enacted or introduced moratoriums, zoning bans, and development freezes on new AI data center construction.
• National legislation introduced by Bernie Sanders to prohibits the development or deployment of AI systems that match or exceed human cognitive performance across a broad range of domains (with criminal penalties of up to 20 years imprisonment for individuals who violate the law).
• Passage of the EU Artificial Intelligence Act (Regulation EU 2024/1689),

Yes, the demands for new totalitarian government powers rest on the doomer predictions about the future of AI.

So what’s his track record?

Abstract

X.com user [12] scored twenty of Yudkowsky’s public AI predictions from 1996 to 2022 against their own deadlines and resolution criteria. Thirteen are falsified by their own deadlines, two were retracted by Yudkowsky himself, one capability bet hit at 16% odds, one was a hedge, and three remain open. Every dated timeline he ever published failed. The pattern is not subtle, and it should change how much weight you give to his AI-doom pronouncements.[13]

For easier readability, this brief extracts all twenty scored predictions in the order they were made, with each one’s hero quote, full-source link, and prediction page on the site.

Judge the pattern for yourself:

1996 – Singularity by 2025 (or 2005, given funding!)

• Status: Retracted (he disavowed the essay around 2001, then everything pre-2003) [14]
• Hero quote: “It’s expected in 2035. (Oops, make that 2025.)”
• What happened: Every named date has passed. The essay applied Moore’s law to subjective time (2035 -> 2025 -> 2015), offered 2005 as a target with ~$100M/year funding, gave “thirty seconds” as a lower bound, and claimed the 1996 Internet already had enough compute to run a human brain. Yudkowsky added a disavowal around 2001 and later disavowed all pre-2003 writing.
• Source: Eliezer Yudkowsky, “Staring into the Singularity” (1996) – https://textfiles.meulie.net/russian/cyberlib.narod.ru/lib/critica/sing/singstar.html
• Prediction page: https://foom.hyperplex.org/#prediction-staring

1999 – Seed AI by 2010, to outrace nanotech

• Status: Falsified [15]
• Hero quote: “To have a good chance of outracing nanotech, we should plan to develop a seed AI by 2010.”
• What happened: He advertised the prediction with a “Singularity 2005” T-shirt. In the 2008 “Coming of Age” sequence he reframed the whole plan as “The Sheer Folly Of Callow Youth.” 2010 passed with neither seed AI nor Drexlerian nanotechnology.
• Source: Eliezer Yudkowsky, “The Plan to Singularity” (1999-2000) – https://web.archive.org/web/20100727182235/http://yudkowsky.net/obsolete/plan.html
• Prediction page: https://foom.hyperplex.org/#plan

1999-2000 – Nanotech 50% by 2010; Singularity by 2020

• Status: Falsified ^[3]
• Hero quote: “The 50% point [for nanotech] may be 2010, or earlier. … AI at 2020 CRNS.”
• What happened: Both were labeled “best guesses” in the CRNS appendix. Neither deadline produced the forecast event.
• Source: The Plan to Singularity, CRNS Time appendix (1999-2000) – https://web.archive.org/web/20100727182235/http://yudkowsky.net/obsolete/plan.html#nav_crns
• Prediction page: https://foom.hyperplex.org/#prediction-crns

2000-2008 – Neural nets “certainly Not Sufficient”

• Status: Falsified [16]
• Hero quote: “Regardless of whether neural networks are Necessary, they are certainly Not Sufficient.”
• What happened: From GISAI through the 2008 Sequences he argued intelligence must be built from first principles and called connectionism magical thinking. Deep learning disproved the necessity of hand-built architecture; by 2023 he acknowledged GPT-4 had exceeded what he thought “stacking more layers” could do.
• Source: General Intelligence and Seed AI (2000-01) and the Sequences – https://foom.hyperplex.org/#prediction-connectionism
• Prediction page: https://foom.hyperplex.org/#connectionism

January 2001 – Hard takeoff in a basement within ten years

• Status: Falsified (by 2011, and again by 2020) [17]
• Hero quote: “…a hard takeoff occurs in SIAI’s or Webmind’s basement sometime in the next ten years or so.”
• What happened: Webmind collapsed in 2001; SIAI never built its seed AI; no takeoff by 2011 or 2020.
• Source: SL4 mailing list, Jan 2001 – https://foom.hyperplex.org/#prediction-basement
• Prediction page: https://foom.hyperplex.org/#basement

2001 – “Three hours to transhumanity”

• Status: Retracted [18]
• Hero quote: “…the ‘hard takeoff’ ‘seed AI’ three-hours-to-transhumanity scenario.”
• What happened: Creating Friendly AI treated the ultrarapid takeoff as most likely, ending in a “Sysop” with near-absolute power. He abandoned the picture in his blanket disavowal of pre-2003 work and the “Coming of Age” sequence. “No three-hour anything has occurred.”
• Source: Creating Friendly AI 1.0 (2001) – https://foom.hyperplex.org/#prediction-cfai
• Prediction page: https://foom.hyperplex.org/#cfai

June 2001 – Time to Singularity: “2005-2020, peak 2008-2010”

• Status: Falsified (every year in the range has passed) [19]
• Hero quote: “2005-2020 / peak at 2008-2010” (his entire SL4 poll reply)
• What happened: A month later he stressed the dates were engineered targets, not optimism. Every year in the range has now passed.
• Source: SL4 mailing list, June 2001 (http://sl4.org/archive/0106/1491.html)
• Prediction page: https://foom.hyperplex.org/#prediction-ets

2001-2002 – SIAI ships transhuman AI ~2008-2010

• Status: Falsified [20]
• Hero quote: “…reaching transhumanity sometime between 2005 and 2020, probably around 2008 or 2010.”
• What happened: Those years came and went. Flare (the custom language for the seed AI) was abandoned; the organization turned to alignment theory and became MIRI.
• Source: singinst.org mission pages (2001-02 archive) – https://web.archive.org/web/20020124071714/http://singinst.org/intro.html
• Prediction page: https://foom.hyperplex.org/#prediction-siai

June 2002 – “Dead certain” hard takeoff; days to nanotech

• Status: Falsified [21]
• Hero quote: “I would call it dead certain in favor of a hard takeoff.”
• What happened: Progress has been gradual, distributed across deployed systems and large organizations, constrained by compute, capital, and infrastructure. Recursive improvement has looked like an industrial process, not a single system rewriting itself.
• Source: SL4 mailing list, June 2002 – http://sl4.org/archive/0206/4129.html
• Prediction page: https://foom.hyperplex.org/#prediction-takeoff

June 2006 – “A planet-sized neural network would fail the same way”

• Status: Falsified [22]
• Hero quote: “But a neural network the size of a planet, given the same examples, would have failed in the same way.” (contra Bill Hibbard)
• What happened: The systems that later surprised him are supervised/self-supervised learning with local search at unprecedented scale – the exact configuration the claim ruled out.
• Source: SL4 mailing list, June 2006 (contra Bill Hibbard) – http://sl4.org/archive/0606/15161.html
• Prediction page: https://foom.hyperplex.org/#prediction-planet

2008 – AGI = ten insights, a small team, one desktop

• Status: Falsified [23]
• Hero quote: “Intelligence is about architecture… architecture is mostly about deep insights.” (Hanson-Yudkowsky foom debate)
• What happened: GPT-3 (175B parameters, industrial compute, a three-year-old architecture) was hard to reconcile with the “ten insights” picture. In 2021 he acknowledged being “very unpleasantly surprised” by how fast “More Compute” worked.
• Source: Hanson-Yudkowsky foom debate, 2008 – https://foom.hyperplex.org/#prediction-foom
• Prediction page: https://foom.hyperplex.org/#foom

November 2021 – No GDP boom before doom

• Status: Falsified (per the site: no need to wait for any deadline) [24]
• Hero quote: “My mainline view is that growth stays at 5%/year and then everybody falls over dead in 3 seconds and the world gets transformed into paperclips.”
• What happened: AI showed up in corporate investment, valuations, and growth through a highly visible industrial buildout – the opposite mechanism. “All hell did not break loose in 2025.”
• Source: Yudkowsky and Christiano discuss “Takeoff Speeds,” LessWrong, Nov 2021 – https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds
• Prediction page: https://foom.hyperplex.org/#prediction-gdp

November 2021 – Self-driving cars approved only after the world ends

• Status: Falsified ^[13]
• Hero quote: “Once you can buy a self-driving car, the thing that Paul predicts with surety and that I shrug about has already happened.” (his own resolution criterion)
• What happened: By 2025 Waymo ran paid fully-driverless robotaxi service across Phoenix, San Francisco, Los Angeles, Austin, and Atlanta – approved through ordinary regulation, built by a non-AGI product team, hundreds of thousands of rides a week.
• Source: Yudkowsky-Christiano takeoff-speeds debate, LessWrong, November-December 2021 – https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds
• Prediction page: https://foom.hyperplex.org/#prediction-selfdriving

November 2021 – Trillion-scale AI boom “pretty darn prohibited”

• Status: Falsified ^[13]
• Hero quote: “this all sure does sound ‘pretty darn prohibited’ on my model”
• What happened: Christiano’s picture became ordinary news: hyperscaler capex ~$410B in 2025, projected past $700B in 2026; Nvidia above $4T (July 2025); frontier-lab rounds in the tens of billions.
• Source: Yudkowsky-Christiano takeoff-speeds debate, LessWrong, November-December 2021 – https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds
• Prediction page: https://foom.hyperplex.org/#prediction-trillionboom

November 2021 – Deepnets as ubiquitous as computers by 2035

• Status: Falsified (arrived roughly a decade early) ^[13]
• Hero quote: “if the world lasts until 2035, I could perhaps see deepnets becoming as ubiquitous as computers were in… 1995? 2005?”
• What happened: ChatGPT reached hundreds of millions of weekly users within two years; AI assistants are built into phones, browsers, office software, and dev tools.
• Source: Yudkowsky-Christiano takeoff-speeds debate, LessWrong, November-December 2021 – https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds
• Prediction page: https://foom.hyperplex.org/#prediction-deepnets

November 2021 – Codex: five years of stagnation or a massive leap in two

• Status: Hedged (his framing was compatible with either outcome – Christiano’s objection to the model) ^[13]
• Hero quote: “I feel like I would indeed not have the right to feel very surprised if Codex technology stagnated for the next 5 years, nor if it took a massive leap.”
• What happened: The leap path – GPT-4 (early 2023), then widespread coding assistants by 2024, growing faster than Christiano’s doubling curve. The site scores it yellow, not green: he did correctly name unregulated niches (Codex, “anime waifus”) as breakout candidates, but the claim was broad enough to fit either result.
• Source: takeoff debate, Nov 2021 – https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds
• Prediction page: https://foom.hyperplex.org/#prediction-codexbet

November 2021 – $1B/yr waifutech before self-driving “freedom of the road”

• Status: Open (no deadline; terms never settled) ^[13]
• Hero quote: “I think my sheer impression would be that we’ll get $1B/yr waifutech before we get AI freedom-of-the-road”
• What happened (mid-2026): Waymo annualized revenue above $350M; companion/character-AI revenue scattered and hard to compare. Autonomous driving is ahead in commercial deployment, but the comparison is too loosely defined for a clean verdict.
• Source: Yudkowsky-Christiano takeoff-speeds debate, LessWrong, November-December 2021 – https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds
• Prediction page: https://foom.hyperplex.org/#prediction-waifubet

January 2017 – Doom by Jan 1, 2030, at 2:1 odds (the Caplan bet)

• Status: Open – resolves January 1, 2030 [25]
• Hero quote: “Bryan Caplan pays Eliezer $100 now, in exchange for $200 CPI-adjusted from Eliezer if the world has not been ended by nonaligned AI before 12:00am GMT on January 1st, 2030.”
• Resolution criterion (his own): “If there are still biological humans running around on the surface of the Earth, it will not have been said to be ended.”
• Note: Caplan’s framing: at 5.5% real rates the $100 grows to $200 without any bet, so the wager mostly paid him to wait.
• Source: Bryan Caplan, EconLog, January 2017 – https://www.econlib.org/archives/2017/01/my_end-of-the-w.html
• Prediction page: https://foom.hyperplex.org/#prediction-caplan

February 2022 – AI takes IMO gold by end of 2025 (>= 16%)

• Status: HIT (resolved July 2025) [26]
• Hero quote: “I’ll stand by a >16% probability of the technical capability existing by end of 2025.”
• What happened: In July 2025, Google’s Gemini Deep Think received an officially certified gold-level score (35/42) and OpenAI reported the same for its model. Christiano had put the odds below 8% and pre-registered that a Yudkowsky win would move him from 30% to 50% on hard takeoff. [27]
• Source: Paul Christiano, “IMO challenge bet with Eliezer,” LessWrong, Feb 2022 – https://www.lesswrong.com/posts/sWLLdG6DWJEy3CH7n/imo-challenge-bet-with-eliezer ; resolution: “Google and OpenAI get 2025 IMO Gold,” LessWrong, Jul 2025 – https://www.lesswrong.com/posts/ZkgaPopsBkQgeA2k8/google-and-openai-get-2025-imo-gold
• Prediction page: https://foom.hyperplex.org/#prediction-imo

April 2022 – “Death with Dignity”: certain doom, no date

• Status: Open (unfalsifiable as stated – no date, no resolution criterion) [28]
• Hero quote: “…doubling our chances of survival will only take them from 0% to 0%.”
• What it claims: Humanity will not solve alignment or coordinate; the current prediction is certain extinction. His “dignity points” strategy: keep improving humanity’s log-odds without lying, abandoning ethics, or giving up.
• Source: Eliezer Yudkowsky, “MIRI Announces New Death with Dignity Strategy,” LessWrong, April 2022 – https://www.lesswrong.com/posts/j9Q8bRmwCgXRYAgcJ/miri-announces-new-death-with-dignity-strategy
• Prediction page: https://foom.hyperplex.org/#prediction-doom

Scoreboard

The pattern: every dated timeline prediction (1996-2008) failed by its own deadline; the 2021 takeoff-model claims were falsified economically rather than by deadline; the one hit is a capability claim, not a doom claim. The only date-certain prediction still standing is the Caplan bet.

Thirty years, twenty scored predictions, thirteen falsified, two retracted by the author himself. If a stock picker had this record, you would not buy his newsletter. If a weatherman missed this many hurricanes, the station would reassign him. Yet the AI-doom movement asks you to reorganize your politics, your portfolio, and your civilization’s technology policy around the forecasts of a man whose every published deadline has quietly expired.

The honest response to a failed forecaster is not hatred, and it is not censorship. It is discounting. Weight predictions by track record. Yudkowsky’s own retraction – he disavowed everything he wrote before 2003 – tells you even he does not stand behind the early material. The 2021 takeoff-model claims were falsified not by a deadline but by the real economy: trillions in capex, robotaxis on public roads, hundreds of millions of people using AI tools, all deployed through ordinary markets and ordinary regulation. Doom that never arrives on schedule is not a reason to accelerate the deadline. It is evidence the model is wrong.

Progress is real, it is measurable, and it is happening in the open – built by thousands of competing teams, not sealed in a basement. The people telling you to panic have been wrong on every date they dared to commit to paper. The people building the tools that will cure your diseases, extend your life, and raise global living standards are delivering, year after year, exactly the gradual, industrial, plural progress the doomers said was impossible.

Call to Action

• Before supporting any AI licensing bill or treaty, ask its sponsors to name the specific predictions they are using to justify the bill. A twenty-year record of missed deadlines is not a basis for handing a general-purpose technology to the state.
• Read the primary sources yourself. Every quote in this post links to Yudkowsky’s original essays and mailing-list posts, and every scored prediction has a page on the foomflops timeline (https://foom.hyperplex.org/). Do not take our word for the scoreboard; check the receipts.
• Push back on doom-flavored regulation in your own circles: city councils weighing data-center moratoria, legislatures drafting AI safety boards with subpoena power, employers treating model release dates as catastrophes. The credible counterargument is the record above.
• Support the builders: open-weight model releases, competitive AI labs, and the researchers actually shipping capability and safety work. Acceleration with eyes open beats blindness with a cane.

Footnotes

Notes