# foomflops: full text export A sourced timeline of Eliezer Yudkowsky’s AI predictions, their deadlines, and how they resolved. This export mirrors the canonical pages at https://foom.hyperplex.org/. Status labels are editorial assessments; primary-source links are included with every entry. # Chapters ## Chapter 01: Extropians Canonical URL: https://foom.hyperplex.org/chapter/extropians Timeline: mid-1990s – 2000 Thesis: When transhumanism met mailing-list accelerationism As a teenager, Eliezer Yudkowsky was deeply involved in the Extropians mailing list, which was then one of the main online gathering places for transhumanists interested in cryonics, nanotechnology, life extension, and the technological Singularity. He soon became one of the list’s most active and intense participants. By 2000, however, he had grown frustrated with what he saw as its lack of focus and urgency. He left to found SL4 (“Shock Level Four”), a more specialized forum whose stated purpose was to accelerate the advent of artificial superintelligence. ### Chapter sources - Extropians Archives: https://extropy.org/ - Staring into the Singularity: https://textfiles.meulie.net/russian/cyberlib.narod.ru/lib/critica/sing/singstar.html --- ## Chapter 02: SL4 / Plan to Singularity Canonical URL: https://foom.hyperplex.org/chapter/sl4 Timeline: 1999–2000 Thesis: Shock Level Four and open-source seed AI SL4 was for the people who were already at “Shock Level Four.” People who took the Singularity, Jupiter brains, and mental revision seriously, rather than treating them as sci-fi concepts. Around the same time, Yudkowsky published The Plan to Singularity and The Singularitarian Principles. In both, he argued for developing seed AI openly, with the hope of setting off a broader AI industry that could reach superintelligence on an aggressively short timetable, somewhere around 2005 to 2010. ### Chapter sources - SL4 Archives: http://sl4.org/archive/0008/0000.html - The Plan to Singularity: https://web.archive.org/web/2002/http://yudkowsky.net/obsolete/plan.html --- ## Chapter 03: Singularity Institute for Artificial Intelligence Canonical URL: https://foom.hyperplex.org/chapter/siai Timeline: 2000–~2012 Thesis: From 'we will build Friendly AI' to early catastrophic risk warnings Yudkowsky’s first organization, founded in July 2000 with seed funding from Brian and Sabine Atkins. Its mission was to create a Friendly, self-improving Artificial Intelligence and thereby bring about the Singularity. As the Institute's own technical efforts looked increasingly fruitless while a neural network renaissance bloomed, the Institute turned its attention toward the catastrophic risks posed by unaligned AI, which soon became central to its public work. ### Chapter sources - Singularity Institute Archive: https://web.archive.org/web/2002/http://singinst.org/ - CFAI Essay: https://intelligence.org/files/CFAI.pdf --- ## Chapter 04: Overcoming Bias / LessWrong / Rationality community Canonical URL: https://foom.hyperplex.org/chapter/overcoming_bias Timeline: 2006–2012 Thesis: Debugging humanity, one very long blog post at a time Robin Hanson founded Overcoming Bias, where Yudkowsky soon became the most prominent contributor. Across the Sequences, lessons about cognitive bias and epistemology accumulated into a shared intellectual web, a course of study, and then the basis of a community. LessWrong launched in 2009, carrying that project into a site of its own. The earlier emphasis on accelerating AGI increasingly gave way to “raising the sanity waterline.” The hope was that a small group trained to reason unusually well might be able to build AGI without causing an existential catastrophe. The 2008 Hanson–Yudkowsky “Foom” debate put that hope beside a more basic disagreement over how AI progress would actually unfold, although the test itself was just two theorists arguing about a future that neither could see. ### Chapter sources - Overcoming Bias: https://www.overcomingbias.com/ - LessWrong Sequences: https://www.lesswrong.com/rationality --- ## Chapter 05: MIRI: Research Canonical URL: https://foom.hyperplex.org/chapter/miri_research Timeline: 2013–2021 Thesis: those who can't build, align After the early implementation timelines failed, MIRI turned toward decision theory, logical uncertainty, agent foundations, and alignment. The Institute increasingly treated the lack of any known safe path to advanced AI as its central problem. Meanwhile, deep learning complicated the picture. Progress came through larger models, more data, better chips, and large research organizations, rather than through the small number of decisive architectural breakthroughs the foom model had anticipated. AlphaGo raised serious doubts about that picture; GPT-3 did far more damage. ### Chapter sources - MIRI Rebrand Announcement: https://intelligence.org/2013/01/30/singularity-institute-becomes-machine-intelligence-research-institute/ --- ## Chapter 06: MIRI: Advocacy Canonical URL: https://foom.hyperplex.org/chapter/miri_propaganda Timeline: 2022–2025 Thesis: those who can't align, sabotage By 2022, MIRI's public discussion around AGI assumed it would end in catastrophe. “Dying with dignity” was not a call to stop trying. It meant working on whatever small chance remained without pretending the odds were better than they were, lying to oneself, or excusing reckless behavior. Public warnings followed, then calls for a worldwide shutdown of advanced AI development. AI, however, was already entering ordinary use. Chatbots and coding assistants spread. Training runs became major corporate expenditures. Governments began writing regulations, and technology companies competed openly to build larger systems. The old takeoff story had expected a single decisive jump. Instead, AI arrived piecemeal, at enormous expense, and in full public view. ### Chapter sources - Death with Dignity Strategy: https://www.lesswrong.com/posts/j9Q8bRmwCgXRYAgcJ/miri-announces-new-death-with-dignity-strategy - TIME: Pausing AI Developments Isn't Enough: https://time.com/6266923/ai-eliezer-yudkowsky-open-letter-not-enough/ --- ## Chapter 07: When Prophecy Fails Canonical URL: https://foom.hyperplex.org/chapter/when_prophecy_fails Timeline: 2026 – present Thesis: the prophecy recedes; the apparatus grows By 2026, the old FOOM scenario had still not materialized. AI progress looked nothing like a small team pushing one system through recursive self-improvement in a matter of hours. It looked like datacenters, larger training runs, staggering capital expenditure, and products released into ordinary life. The institutions built during the years of imminent-risk advocacy did not disappear with that mismatch. By then, they had staff, donors, publications, political contacts, and an audience of their own. The campaign also found support among people whose concerns had little to do with hard takeoff. Projects such as PDKU brought AI doom to audiences outside the old rationalist subculture. Food & Water Watch connected the issue to local fights over datacenters. Sanders focused on automation, employment, and concentrated wealth. Ocasio-Cortez added surveillance, rising electricity costs, and the jars of brown well water brought by Georgia residents who blamed a nearby Meta development, though Meta disputed the connection. The resulting moratorium could be presented as an answer to extinction risk, layoffs, corporate power, pollution, utility bills, and pressure on local water supplies all at once. Its supporters did not need to agree about FOOM, or even believe in it, to support the same policy. The April attack on Sam Altman’s home raised a more difficult question about what happens when apocalyptic language leaves the page. There is no evidence that PauseAI organized, approved, or had prior knowledge of the attack. The group condemned it and said that the accused had never held a position within the organization. That distinction matters. Still, the incident made it harder to dismiss the relationship between extinction rhetoric and fringe action as entirely hypothetical. A miniature version of the discourse could be seen online, where arguments about AI policy increasingly became arguments about whether particular people should remain employed. Liron Shapira called two aspects of a Sébastien Krier post “fireable.” Andrew Ho’s departure from OpenAI was later circulated as evidence that he had been dismissed, although he had already announced that he was leaving to start a company and later denied the firing story directly. Collectively, these events show how quickly disagreement can become a question of loyalty once the issue is framed as an emergency. In a situation where AI development is seen as a crisis that poses an imminent danger to humanity, even something as simple as making a joke or an argument can become the basis for a demand to have someone fired, while resistance to a moratorium can be cast as helping the danger along. ### Chapter sources - Sanders–AOC moratorium: https://www.sanders.senate.gov/press-releases/news-sanders-ocasio-cortez-announce-ai-data-center-moratorium-act/ - AOC's well-water hearing: https://ocasio-cortez.house.gov/media/press-releases/ocasio-cortez-presses-epa-assistant-administrator-kramer-jeopardizing-clean - DOJ: April attack: https://www.justice.gov/opa/pr/texas-man-faces-multiple-federal-charges-related-attack-ai-company-and-its-ceo - PauseAI's response: https://pauseai.info/statement-sam-altman-attack-2026 - Ho announces his startup: https://x.com/andrewho03/status/2082615798011744270 - Ho denies the firing claim: https://x.com/andrewho03/status/2084055130723496137 - Shapira on Krier: https://x.com/liron/status/2081879184930959807 # Predictions ## Singularity by 2025 (or 2005, given funding!) Canonical URL: https://foom.hyperplex.org/p/staring Chapter: Extropians Category: timelines & doom Predicted: 1996 Resolution date: 2001 Status: walked back Status detail: The prediction was resolved as walked back in 2001 Source: Staring into the Singularity, 1996 > It's expected in 2035. (Oops, make that 2025.) ### Resolution Disavowal appended to the essay itself: everything about the Powers “almost certainly wrong.” The dates expired on schedule anyway. ### Context and outcome The essay applied Moore’s law to subjective time, producing the date 2035, then 2025, then 2015; with 2005 as a target if the project received roughly $100 million a year, with thirty seconds given as a lower bound. The 2000 revision claimed that the Internet already contained enough computing power, “if properly reprogrammed,” to run a human brain, and that “perhaps a smaller computer” could support a seed AI. Yudkowsky gradually backed away from these claims. Around 2001, he added a disavowal to the essay. He then later disavowed everything he had written before 2003. Every named date has since passed. ### Supporting receipts > It's expected in 2035. (Oops, make that 2025.) Context: The document's opening doubling-cascade ends on a Singularity date revised downward in place in a later edit. Source: Staring into the Singularity (1996, mirror) — https://textfiles.meulie.net/russian/cyberlib.narod.ru/lib/critica/sing/singstar.html > I would like the Singularity to occur in 2005, which I think I would have a reasonable chance of doing via AI if someone handed me a billion dollars a year. … The lower bound? Thirty seconds. Context: His stated target date given funding, followed by the same section's closing estimate of the lower bound. Source: Staring into the Singularity (1996, mirror) — https://textfiles.meulie.net/russian/cyberlib.narod.ru/lib/critica/sing/singstar.html > We don't lack the technology; the Singularity could happen tomorrow. The Internet has enough power, if properly reprogrammed, to run a human brain, and perhaps a smaller computer has enough power to run a seed AI. Context: Claim that the 1996 Internet already contained enough compute for a human brain. Source: Staring into the Singularity (1996, mirror) — https://textfiles.meulie.net/russian/cyberlib.narod.ru/lib/critica/sing/singstar.html ### Sources - full text (mirror): https://textfiles.meulie.net/russian/cyberlib.narod.ru/lib/critica/sing/singstar.html - wayback: https://web.archive.org/web/2002/http://sysopmind.com/singularity.html --- ## Seed AI by 2010, to outrace nanotech Canonical URL: https://foom.hyperplex.org/p/plan Chapter: SL4 / Plan to Singularity Category: timelines & doom Predicted: 1999 Resolution date: 2010 Status: falsified by events Status detail: The prediction was resolved as falsified by events in 2010 Source: The Plan to Singularity, 1999–2000 > To have a good chance of outracing nanotech, we should plan to develop a seed AI by 2010. ### Resolution 2010 came. No seed AI, no nanotech, no race. He'd already called the plan a folly in 2008. ### Context and outcome The Plan treated Drexlerian nanotechnology as a doomsday deadline. If it arrived around 2010, a seed AI would need to be developed first; ideally by 2005, and no later than 2010. He advertised the prediction with a T-shirt reading “Singularity 2005.” In the 2008 “Coming of Age” sequence he framed the whole plan as The Sheer Folly Of Callow Youth. Then 2010 passed without seed AI or Drexlerian nanotechnology. ### Supporting receipts > To win, we need to create an AI. To avoid losing, we need to outrace nanotechnology. … To have a good chance of outracing nanotech, we should plan to develop a seed AI by 2010. Context: The plan's core framing of AI as a race against nanotechnology, with a 2010 seed-AI deadline. Source: The Plan to Singularity (1999) — https://web.archive.org/web/20100727182235/http://yudkowsky.net/obsolete/plan.html > I'll be overjoyed if we have 'till 2020 to create a seed AI, but it's increasingly looking like the deadline may be more on the order of 2005. That's not impossible, but it's damned tenuous. Context: From the 'If nanotech comes first' section, moving the deadline earlier to roughly 2005. Source: The Plan to Singularity (1999) — https://web.archive.org/web/20100727182235/http://yudkowsky.net/obsolete/plan.html > I would not realistically expect a substantial user base before four years have passed, making my "Singularity 2005" T-Shirt a touch unrealistic... but we can always hope. Context: Schedule slippage measured against his own 'Singularity 2005' T-shirt; the ellipsis is in the original. Source: The Plan to Singularity (1999) — https://web.archive.org/web/20100727182235/http://yudkowsky.net/obsolete/plan.html ### Sources - The Plan to Singularity (wayback): https://web.archive.org/web/2002/http://yudkowsky.net/obsolete/plan.html - The Sheer Folly of Callow Youth: https://www.lesswrong.com/posts/Yicjw6wSSaPdb83w9/the-sheer-folly-of-callow-youth --- ## Nanotech ~2010; the Singularity ~2020 Canonical URL: https://foom.hyperplex.org/p/crns Chapter: SL4 / Plan to Singularity Category: timelines & doom Predicted: 1999 Resolution date: 2020 Status: falsified by events Status detail: The prediction was resolved as falsified by events in 2020 Source: The Plan to Singularity, CRNS Time, 1999–2000 > The 50% point [for nanotech] may be 2010, or earlier. … AI at 2020 CRNS. ### Resolution His “best guesses” supplied both sides of the race. 2020 passed without either Drexlerian nanotech or the Singularity. ### Context and outcome In the CRNS appendix to The Plan, Yudkowsky offered forecasts for what might happen without intervention. He placed the 50% point for nanotechnology at 2010 or earlier and predicted AI, meaning the full Singularity, by 2020. He described both dates explicitly as his “best guesses.” ### Supporting receipts > In recent times, I've moved up my CRNS estimate on nanotechnology … it now seems that the 50% point may be 2010, or earlier. … Some other key CRNS numbers include AI at 2020 CRNS. Context: The Plan's separate current-rate forecasts for nanotechnology and AI, rather than its intervention target. Source: The Plan to Singularity, CRNS Time (1999–2000) — https://web.archive.org/web/20100727182235/http://yudkowsky.net/obsolete/plan.html#nav_crns > Those are just my numbers - best guesses. Context: Yudkowsky's own label for the CRNS dates in the appendix. Source: The Plan to Singularity, CRNS Time (1999–2000) — https://web.archive.org/web/20100727182235/http://yudkowsky.net/obsolete/plan.html#nav_crns ### Sources - The Plan to Singularity: CRNS Time: https://web.archive.org/web/20100727182235/http://yudkowsky.net/obsolete/plan.html#nav_crns --- ## Neural nets: “certainly Not Sufficient” Canonical URL: https://foom.hyperplex.org/p/connectionism Chapter: Singularity Institute for Artificial Intelligence Category: neural nets & capabilities Predicted: 2000 Resolution date: 2023 Status: falsified by events Status detail: The prediction was resolved as falsified by events in 2023 Source: General Intelligence and Seed AI, 2000–01 > Regardless of whether neural networks are Necessary, they are certainly Not Sufficient. ### Resolution Trained networks did it anyway. His 2023 source: “GPT-4 got further than I thought stack more layers was going to get.” ### Context and outcome From GISAI through the 2008 Sequences, Yudkowsky argued that intelligence had to be built from first principles. He treated connectionism as a form of magical thinking, objected that training-based methods displaced the problem into the data, and regarded opaque neural networks as both technically unpromising and unsafe. Deep learning did not settle every argument about connectionism, but it did disprove the idea that advanced intelligence needed hand-built architecture. Yudkowsky changed from his surprise during the AlphaGo era to his 2023 acknowledgment that GPT-4 had exceeded what he thought “stacking more layers” could do. ### Supporting receipts > Not to mention that neural networks have also been “failing” (i.e., not yet succeeding) to produce real AI for 30 years now. […] But at least don’t tell me it’s still the new revolutionary idea in AI. Context: Yudkowsky dismissing neural networks as a stale, oversold research program. Source: Logical or Connectionist AI? (LessWrong, Nov 2008) — https://www.lesswrong.com/posts/juomoqiNzeAuq4JMm/logical-or-connectionist-ai > I’m no fan of neurons; this may be clearer from other posts. Context: Yudkowsky's own comment under the post, summing up his stance on neural approaches. Source: Logical or Connectionist AI?, comments (LessWrong, Nov 2008) — https://www.lesswrong.com/posts/juomoqiNzeAuq4JMm/logical-or-connectionist-ai > A backprop network with sigmoid units… actually doesn’t much resemble biology at all. Around as much as a voodoo doll resembles its victim. Context: Arguing the resemblance between artificial neural nets and brains is pure surface analogy. Source: Failure By Analogy (LessWrong, Nov 2008) — https://www.lesswrong.com/posts/C4EjbrvG3PvZzizZb/failure-by-analogy > Maybe neurons are just what brains happen to be made out of, because the blind idiot god is too stupid to sit down and invent transistors. Context: Suggesting neurons are evolution's poor building material, not the secret of intelligence. Source: Failure By Analogy (LessWrong, Nov 2008) — https://www.lesswrong.com/posts/C4EjbrvG3PvZzizZb/failure-by-analogy > …cognitive abilities are not emergent and do not happen by accident. Both classical AI and connectionist AI propose a kind of magic that avoids the difficulty of actually implementing the higher layers of cognition. Context: GISAI dismisses both AI camps as relying on magic and rejects emergence outright. Source: General Intelligence and Seed AI, singinst.org (2002 archive) — https://web.archive.org/web/20021212060638/http://singinst.org/GISAI.html > Regardless of whether neural networks are Necessary, they are certainly Not Sufficient. The human brain requires millions of years of evolution, thousands of modules, hundreds of thousands of adaptations… Context: The 'Necessary, But Not Sufficient' heuristic applied to neural networks. Source: General Intelligence and Seed AI, singinst.org (2002 archive) — https://web.archive.org/web/20021212060638/http://singinst.org/GISAI.html > Any form of cognition which can be mathematically formalized, or which has a provably correct implementation, is too simple to contribute materially to intelligence. Context: GISAI's 'Law of Pragmatism', ruling formalizable methods out of real intelligence. Source: General Intelligence and Seed AI, singinst.org (2002 archive) — https://web.archive.org/web/20021212060638/http://singinst.org/GISAI.html ### Sources - GISAI (wayback): https://web.archive.org/web/2002/http://singinst.org/GISAI.html - Logical or Connectionist AI?: https://www.lesswrong.com/posts/juomoqiNzeAuq4JMm/logical-or-connectionist-ai - Dwarkesh podcast (2023): https://www.dwarkesh.com/p/eliezer-yudkowsky --- ## Hard takeoff in a basement within ten years Canonical URL: https://foom.hyperplex.org/p/basement Chapter: Singularity Institute for Artificial Intelligence Category: timelines & doom Predicted: 2001 Resolution date: 2011 Status: falsified by events Status detail: The prediction was resolved as falsified by events in 2011 Source: SL4, Jan 2001 > …a hard takeoff occurs in SIAI's or Webmind's basement sometime in the next ten years or so. ### Resolution Webmind went bankrupt within months. The decade closed with no takeoff from anyone's basement. ### Context and outcome Yudkowsky claimed that genetic engineering and brain-computer interfaces wouldn't matter because they'd arrive around 2020, and a basement hard takeoff would beat them by a decade. Webmind Inc. collapsed in 2001, SIAI never built its seed AI, and nothing resembling a takeoff occurred by 2011 or 2020. ### Supporting receipts > What actually happens is that genetic engineering and neurohacking … don't show up, because they'd show up in 2020, and a hard takeoff occurs in SIAI's or Webmind's basement sometime in the next ten years or so. Context: Arguing human-enhancement tech never arrives because a basement hard takeoff beats it by a decade. Source: SL4, 24 Jan 2001 — http://sl4.org/archive/0101/0452.html > There still comes a point when the SI says that the Introdus wavefront is on the way, and you sit there waiting for the totally unknowable future to hit you in the next five seconds. Context: Same message, describing the final moments after the basement takeoff. Source: SL4, 24 Jan 2001 — http://sl4.org/archive/0101/0452.html ### Sources - SL4: Event horizon (2001): http://sl4.org/archive/0101/0452.html --- ## “Three hours to transhumanity” Canonical URL: https://foom.hyperplex.org/p/cfai Chapter: Singularity Institute for Artificial Intelligence Category: takeoff mechanics Predicted: 2001 Resolution date: 2008 Status: walked back Status detail: The prediction was resolved as walked back in 2008 Source: Creating Friendly AI, 2001 > …the ‘hard takeoff’ ‘seed AI’ three-hours-to-transhumanity scenario. ### Resolution Covered by his blanket disavowal of pre-2003 work, and by the 2008 sequence calling young Eliezer wrong. ### Context and outcome CFAI treated ultrarapid takeoff as the most likely outcome. On this account, a single seed AI would reach superintelligence and develop nanotechnology along something close to the minimum possible timeline, culminating in a “Sysop” with control over all matter within human reach. Yudkowsky later abandoned this picture in his blanket disavowal of his pre-2003 work and in the “Coming of Age” sequence. No three-hour anything has occurred. ### Supporting receipts > I visualize an extremely powerful AI produced by an ultrarapid takeoff, not just because it's the conservative assumption or the highest-stakes outcome, but because I think it's actually the most likely scenario. Context: Section 2.3 states the ultrarapid takeoff is CFAI's default because it is judged most likely. Source: Creating Friendly AI 1.0 (2001) — https://intelligence.org/files/CFAI.pdf > Under this scenario, the first self-modifying transhuman AI will have, at least in potential, nearly absolute physical power over our world. The potential existence of this absolute power is unavoidable… Context: The predicted position of the first transhuman AI under the default takeoff scenario. Source: Creating Friendly AI 1.0 (2001) — https://intelligence.org/files/CFAI.pdf > Even if you don't believe in the "hard takeoff" "seed AI" three-hours-to-transhumanity scenario… if I'm right about how these things work, the first Friendly AI you build becomes Transition Guide and builds a Sysop… Context: Doubters of the three-hour takeoff are waved off; the Sysop is CFAI's superintelligent operating system for all matter in human space. Source: Creating Friendly AI 1.0 (2001) — https://intelligence.org/files/CFAI.pdf ### Sources - CFAI (PDF): https://intelligence.org/files/CFAI.pdf --- ## Time to Singularity: “2005–2020, peak 2008–2010” Canonical URL: https://foom.hyperplex.org/p/ets Chapter: Singularity Institute for Artificial Intelligence Category: timelines & doom Predicted: 2001 Resolution date: 2020 Status: falsified by events Status detail: The prediction was resolved as falsified by events in 2020 Source: SL4 ETS poll, Jun 2001; his entire reply > 2005-2020 / peak at 2008-2010 ### Resolution The window closed in full on January 1, 2021. No Singularity was observed at the peak or elsewhere. ### Context and outcome Asked on the SL4 mailing list when he expected the Singularity, Yudkowsky gave a range along with its most likely point. A month later, he stressed that the dates were meant as targets rather than expressions of optimism. Real mental effort had gone into making it 2010 instead of 2020 or 2030. Every year in the range has now passed. ### Supporting receipts > 2005-2020 peak at 2008-2010 Context: His complete two-line reply to the SL4 poll on Estimated Time to Singularity. Source: SL4, 12 Jun 2001 — http://sl4.org/archive/0106/1491.html > People may have first heard us talking about 2010 and thought we were just being overoptimistic, not realizing how much mental effort went into figuring out how to make it "2010" instead of "2020" or "2030". Context: Recalling how the Singularity target date was deliberately engineered down to 2010. Source: SL4, 22 Jul 2001 — http://sl4.org/archive/0107/1820.html > We would very much like to do it in ten years, but if it takes twenty years to do it then we will SPEND twenty years, because this *is* a crusade and the entire planet is at stake. Context: Framing SIAI's ten-year timeline as the last great crusade in human history. Source: SL4, 22 Jul 2001 — http://sl4.org/archive/0107/1820.html ### Sources - ETS poll (2001): http://sl4.org/archive/0106/1491.html - The Quest for Singularity: http://sl4.org/archive/0107/1820.html --- ## SIAI ships transhuman AI ~2008–2010 Canonical URL: https://foom.hyperplex.org/p/siai Chapter: Singularity Institute for Artificial Intelligence Category: timelines & doom Predicted: 2001 Resolution date: 2010 Status: falsified by events Status detail: The prediction was resolved as falsified by events in 2010 Source: singinst.org mission pages, 2001–02 > …reaching transhumanity sometime between 2005 and 2020, probably around 2008 or 2010. ### Resolution No seed AI ever shipped. The org pivoted to alignment research and renamed itself MIRI in 2013. ### Context and outcome The Institute was founded to build the Singularity. Its early plan centered on funding a recursively self-improving “Singularity seed,” which was expected to bring about a transhuman future sometime around 2008-2010. Those years came and went. Flare, the programming language meant to support the project, was abandoned, and the organization gradually turned away from building the AI itself toward theoretical work on alignment and became MIRI. ### Supporting receipts > The Singularity Institute is a 501(c)(3) nonprofit devoted solely to creating the Singularity. … our long-term mission is direct research into Singularity technologies and the direct implementation of the Singularity. Context: SIAI's 2002 homepage stating its mission as the direct implementation of the Singularity. Source: singinst.org homepage (2002 archive) — https://web.archive.org/web/20021129235220/http://singinst.org/ > Our best guess for the timescale is that our final-stage AI will reach transhumanity sometime between 2005 and 2020, probably around 2008 or 2010. Context: The institute's official timescale estimate for its AI reaching transhumanity, contingent on funding. Source: An Introduction to the Singularity, singinst.org (2001 page, archived 2002) — https://web.archive.org/web/20020124071714/http://singinst.org/intro.html > There are concepts in Flare that have never been seen before. We expect to be able to solve problems in Flare that cannot realistically be solved in any other language. Context: The goals page for Flare, SIAI's custom XML-based programming language commissioned for building the seed AI. Source: Flare language goals, flarelang.sourceforge.net (2003 archive) — https://web.archive.org/web/20031204212440/http://flarelang.sourceforge.net/goals.html ### Sources - singinst.org (wayback): https://web.archive.org/web/2002/http://singinst.org/ - Garfinkel, On Deference (2022): https://forum.effectivealtruism.org/posts/NBgpPaz5vYe3tH4ga/on-deference-and-yudkowsky-s-ai-risk-estimates --- ## “Dead certain”: days to nanotech Canonical URL: https://foom.hyperplex.org/p/takeoff Chapter: Singularity Institute for Artificial Intelligence Category: takeoff mechanics Predicted: 2002 Resolution date: 2023 Status: falsified by events Status detail: The prediction was resolved as falsified by events in 2023 Source: SL4, Jun 2002 > I would call it dead certain in favor of a hard takeoff… all economic, computational… indicators point straight toward it. ### Resolution AI at expert-human level arrived commercially, over years, bottlenecked on chips and capital. No twelve-hour curve. ### Context and outcome The SL4-era account held that, once an AI reached human-level intelligence, the part of takeoff still unfolding on a human timescale would last roughly twelve hours. Superintelligence would then develop nanotechnology within days or weeks; by 2005, he had shortened that estimate to three days. On this view, a system reaching human-level intelligence would already be well into takeoff. AI development has so far followed a very different pattern. Progress has been gradual, distributed across deployed systems and large organizations, and heavily constrained by computing power, capital, and infrastructure. Recursive improvement has looked less like a single system rapidly rewriting itself than an industrial process unfolding through research, deployment, and economic scaling. ### Supporting receipts > I would call it dead certain in favor of a hard takeoff, unless all the intelligences at the core of that hard takeoff unanimously decide otherwise. Context: Asked whether hard takeoff is a toss-up, a hunch, or near-dead certainty, he picks dead certain. Source: SL4, 22 Jun 2002 — http://sl4.org/archive/0206/4129.html > From the moment the AI realizes a hard takeoff lies ahead to the moment human-timescale phase terminates, should be... oh... twelve hours or so. … six thousand people are dying for every hour delayed. Context: Estimating the controlled-ascent window before an AI goes fully superhuman. Source: SL4, 24 Jan 2001 — http://sl4.org/archive/0101/0460.html > Assumedly a transhuman threatening AGI would have influence over the mental realm and may be able to rapidly sway existing humans to assist it for around the three days before it has its own nanotechnology. Context: On why fighting an unfriendly AGI is hopeless: three days from persuasion to nanotech. Source: SL4, 10 Jul 2005 — http://sl4.org/archive/0507/11458.html ### Sources - How hard a Singularity? (2002): http://sl4.org/archive/0206/4129.html - Basement Education (2001): http://sl4.org/archive/0101/0460.html - Fighting UFAI (2005): http://sl4.org/archive/0507/11458.html --- ## “A planet-sized neural network would fail” Canonical URL: https://foom.hyperplex.org/p/planet Chapter: Overcoming Bias / LessWrong / Rationality community Category: neural nets & capabilities Predicted: 2006 Resolution date: 2023 Status: falsified by events Status detail: The prediction was resolved as falsified by events in 2023 Source: SL4, Jun 2006; contra Bill Hibbard > But a neural network the size of a planet, given the same examples, would have failed in the same way. ### Resolution Planet-scale compute plus local search is precisely what produced GPT-4. Scale was the answer he ruled out. ### Context and outcome Arguing against Hibbard's proposals, he held that supervised learning with local search had a general, unfixable failure mode that no amount of scale would touch. The systems that later surprised him are supervised/self-supervised learning with local search at unprecedented scale, the exact configuration the planet-sized-network claim ruled out. ### Supporting receipts > It is a general problem. … It is a problem that people run into, in general, with supervised learning using local search techniques for traversing the hypothesis space. Context: Defending his smiley-face critique of Bill Hibbard as a general flaw of supervised learning, not one failed network. Source: SL4, 6 Jun 2006 — http://sl4.org/archive/0606/15161.html > But you could have used a galactic-size neural network in the army tank classifer and gotten exactly the same result… Context: Claiming the army tank-classifier failure was scale-independent; the spelling is the original's. Source: SL4, 6 Jun 2006 — http://sl4.org/archive/0606/15161.html > But a neural network the size of a planet, given the same examples, would have failed in the same way. Context: Telling Hibbard that no amount of scale fixes misclassified training targets. Source: SL4, 12 Jun 2006 — http://sl4.org/archive/0606/15268.html ### Sources - vs. Hibbard (2006): http://sl4.org/archive/0606/15161.html - planet-sized network: http://sl4.org/archive/0606/15268.html --- ## AGI = ten insights, a small team, one desktop Canonical URL: https://foom.hyperplex.org/p/foom Chapter: Overcoming Bias / LessWrong / Rationality community Category: takeoff mechanics Predicted: 2008 Resolution date: 2023 Status: falsified by events Status detail: The prediction was resolved as falsified by events in 2023 Source: Hanson–Yudkowsky foom debate, 2008 > Intelligence is about architecture… architecture is mostly about deep insights. ### Resolution It took megawatts, not epiphanies. Compute and capital won the retrospective: “obvious advantage to Robin.” ### Context and outcome GPT-3 was difficult to reconcile with that picture. Released in May 2020, it had 175 billion parameters, was trained with industrial-scale computing resources, and relied on an architecture introduced three years earlier rather than on a new conceptual breakthrough. Its capabilities came largely from scaling an existing method with more data, compute, and capital, the path Yudkowsky had considered least likely. In 2021, he acknowledged being “very unpleasantly surprised” by how rapidly such systems improved through “More Compute.” Later models only made the mismatch harder to dismiss. ### Supporting receipts > we are likely to see a fast, local increase in capability—“AI go FOOM”. Just to be clear on the claim, “fast” means on a timescale of weeks or hours rather than years or decades Context: The post's headline claim, defining the speed of hard takeoff. Source: Recursive Self-Improvement (LessWrong, Dec 2008) — https://www.lesswrong.com/posts/JBadX7rwdcRFzGuju/recursive-self-improvement > because our computing hardware has run so far ahead of AI theory, we have incredibly fast computers we don’t know how to use for thinking; getting AI right could produce a huge, discontinuous jolt Context: The hardware-overhang argument for a discontinuous jump once AI is solved. Source: Hard Takeoff (LessWrong, Dec 2008) — https://www.lesswrong.com/posts/tjH8XPxAnr6JRbh7k/hard-takeoff > there’ll be a discontinuity at the point where the AI understands how to do AI theory, the same way that human researchers try to do AI theory Context: Predicting a sharp capability break once an AI can do AI research itself. Source: Hard Takeoff (LessWrong, Dec 2008) — https://www.lesswrong.com/posts/tjH8XPxAnr6JRbh7k/hard-takeoff > there’s a big bucket o’ resources behind a locked door […] When you fold a complicated, choppy, cascade-y chain of differential equations in on itself via recursion, it should either flatline or blow up. Context: Resource overhangs, plus the claim that recursive self-improvement leaves no gradual middle path. Source: Recursive Self-Improvement (LessWrong, Dec 2008) — https://www.lesswrong.com/posts/JBadX7rwdcRFzGuju/recursive-self-improvement ### Sources - Hard Takeoff (2008): https://www.lesswrong.com/posts/tjH8XPxAnr6JRbh7k/hard-takeoff - foom retrospective (2023): https://www.lesswrong.com/posts/gGSvwd62TJAxxhcGh/yudkowsky-vs-hanson-on-foom-whose-predictions-were-better --- ## Doom by 2030, at 2:1 odds Canonical URL: https://foom.hyperplex.org/p/caplan Chapter: MIRI: Research Category: timelines & doom Predicted: 2017 Resolution deadline: 2030 Status: still open Status detail: The prediction is still open and is scheduled to resolve in 2030 Source: bet with Bryan Caplan, Jan 2017 (paraphrase of terms) > If humanity survives to 00:00 GMT, Jan 1 2030, Yudkowsky pays Caplan $200; a bet at 2:1 odds in favour of doom. ### Resolution Resolves January 1, 2030. Humanity currently extant; Caplan currently ahead. ### Context and outcome The clearest deadline comes from his bet with Bryan Caplan. Caplan paid him $100 up front; if biological humans are still living on Earth’s surface on New Year’s Day 2030, Yudkowsky must repay $200, adjusted for inflation. The terms amount to two-to-one odds on catastrophe before then. Among his recent predictions of AI doom, this is the only one tied to a specific date. That also makes it the easiest to judge. ### Supporting receipts > 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. Context: The headline terms of the 2017 bet, drafted by Yudkowsky and posted by Caplan on EconLog. Source: Bryan Caplan, EconLog (Jan 2017) — https://www.econlib.org/archives/2017/01/my_end-of-the-w.html > If there are still biological humans running around on the surface of the Earth, it will not have been said to be ended. Context: The bet's resolution criterion for whether the world counts as ended. Source: Bryan Caplan, EconLog (Jan 2017) — https://www.econlib.org/archives/2017/01/my_end-of-the-w.html > In my mind, this is primarily a bet that annualized real interest rates stay below 5.5%. After all, at 5.5%, I could turn $100 today in $200 inflation-adjusted without betting. Context: Caplan noting the structure meant he profited unless real rates spiked or the world ended. Source: Bryan Caplan, EconLog (Jan 2017) — https://www.econlib.org/archives/2017/01/my_end-of-the-w.html ### Sources - Caplan, EconLog (2017): https://www.econlib.org/archives/2017/01/my_end-of-the-w.html --- ## No GDP boom before doom Canonical URL: https://foom.hyperplex.org/p/gdp Chapter: MIRI: Research Category: economics & deployment Predicted: 2021 Resolution date: 2025 Status: falsified by events Status detail: The prediction was resolved as falsified by events in 2025 Source: takeoff debate with Paul Christiano, Nov 2021 > 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 ### Resolution The working premise that AI stays economically invisible is dead: AI is the growth story, the capex, the market. ### Context and outcome The claim depended on transformative AI arriving as a sharp break from the existing economic trend, rather than through a period of visible deployment that markets could absorb. World GDP would remain roughly on trend until a sharp capability jump carried events entirely “off the track.” His illustrative scenario was: “if all hell breaks loose in 2025.” That is not what happened. AI was already showing up in corporate investment, market valuations, and economic growth. The technology was arriving through an expensive, highly visible industrial buildout, the opposite of the mechanism his prediction described. There is no reason to postpone that judgment until some later deadline. And, for the record, all hell did not break loose in 2025. ### Supporting receipts > 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; there's never a point with 3000%/year. Context: Liveblog response to Christiano's slow-takeoff framing of growth accelerating 3% -> 30% -> 3000%/year. This is EY's flattest statement that GDP stays boring right up until instant doom. Source: Takeoff Speeds debate (2021) — https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds > Things feel crazy in the AI field and the world ends less than 4 years later, well before the world economy doubles. Context: Sep 14 chat with Christiano, explaining why his model permits 'final descent' beginning in 2025 without ever passing through Paul's economy-doubling phase. Source: Takeoff Speeds debate (2021) — https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds > If doubling GDP is such a big deal, go open borders and build houses. Oh, that's illegal? Well, so will be AIs building houses! Context: From his 'Deployment lag' commentary. The bureaucracy thesis holds that regulation, not AI capability, is what keeps AI out of GDP until the world ends. Source: Takeoff Speeds debate (2021) — https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds > Oh, come on. That is straight-up not how simple continuous toy models of RSI work. Between a neutron multiplication factor of 0.999 and 1.001 there is a very huge gap in output behavior. Context: Yudkowsky's criticality analogy for recursive self-improvement, against Christiano's smooth-takeoff picture. Source: Yudkowsky, takeoff-speeds debate, LessWrong (Nov 2021) — https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds > it isn't transforming society because it's not over the social-transformation threshold. AlphaFold presaged AlphaFold 2 but AlphaFold 2 is good enough to start replacing other ways of determining protein conformations… Context: Yudkowsky arguing capabilities only register socially past a threshold, so GDP lags the real action. Source: Yudkowsky, takeoff-speeds debate, LessWrong (Nov 2021) — https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds > in the real world we can already design a vaccine in a day and the rest of the time is bureaucratic time rather than technology time, and that goes on until we have an AI over the threshold to bypass bureaucracy. Context: Yudkowsky on why transformative AI would not show up in GDP statistics before the end. Source: Yudkowsky, takeoff-speeds debate, LessWrong (Nov 2021) — https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds ### Sources - Takeoff Speeds debate (2021): https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds --- ## AI takes IMO gold by 2025 (≥16%) Canonical URL: https://foom.hyperplex.org/p/imo Chapter: MIRI: Advocacy Category: neural nets & capabilities Predicted: 2022 Resolution date: 2025 Status: came true Status detail: The prediction was resolved as came true in 2025 Source: IMO bet with Paul Christiano, Feb 2022 > I'll stand by a >16% probability of the technical capability existing by end of 2025. ### Resolution DeepMind and OpenAI both hit gold at IMO 2025. The control group: his falsifiable near-term bet won. ### Context and outcome Christiano put the chance of an AI earning IMO gold by the end of 2025 below 8 percent; Yudkowsky put it at 16 percent. In July 2025, Gemini Deep Think received an officially certified gold-level score of 35 out of 42, and OpenAI reported the same score for its own model. This one stays in. Yudkowsky got the main prediction right, while Christiano’s narrower call also held because neither system solved Problem 6. ### Supporting receipts > My probability is at least 16% [on the IMO grand challenge falling], though I'd have to think more and Look into Things, and maybe ask for such sad little metrics as are available… Context: Yudkowsky's stated odds on an AI getting IMO gold by 2025, against Christiano's roughly 8%. Source: Yudkowsky, quoted in Christiano's bet post, LessWrong (Feb 2022) — https://www.lesswrong.com/posts/sWLLdG6DWJEy3CH7n/imo-challenge-bet-with-eliezer > …the sort of foolish little real-world obstacle which can prevent a proposition like this from being judged true even where the technical capability exists. Context: Yudkowsky's caveat that an open-sourcing requirement could block resolution; he stood by >16% on the capability existing by end of 2025. Source: Yudkowsky, quoted in Christiano's bet post, LessWrong (Feb 2022) — https://www.lesswrong.com/posts/sWLLdG6DWJEy3CH7n/imo-challenge-bet-with-eliezer > Maybe I'd move from a 30% chance of hard takeoff to a 50% chance of hard takeoff. If Eliezer wins, he gets 1 bit of epistemic credit. Context: Christiano pre-registering how much he would update toward hard takeoff if Yudkowsky won. Source: Paul Christiano, LessWrong (Feb 2022) — https://www.lesswrong.com/posts/sWLLdG6DWJEy3CH7n/imo-challenge-bet-with-eliezer > Google and OpenAI have LLMs with gold medal performances, each scoring exactly the threshold of 35/42… The results were not super clear cut if you look at the details… but a gold medal was still achieved. Context: The July 2025 resolution: two labs hit the IMO gold threshold, settling the bet Yudkowsky's way. Source: Zvi Mowshowitz, LessWrong (Jul 2025) — https://www.lesswrong.com/posts/ZkgaPopsBkQgeA2k8/google-and-openai-get-2025-imo-gold ### Sources - IMO challenge bet (2022): https://www.lesswrong.com/posts/sWLLdG6DWJEy3CH7n/imo-challenge-bet-with-eliezer - resolution (2025): https://www.lesswrong.com/posts/ZkgaPopsBkQgeA2k8/google-and-openai-get-2025-imo-gold --- ## “Dying with dignity”: certain doom, no date Canonical URL: https://foom.hyperplex.org/p/doom Chapter: MIRI: Advocacy Category: timelines & doom Predicted: 2022 Resolution deadline: No date supplied Status: still open Status detail: The prediction is still open; no resolution deadline was supplied Source: Death with Dignity, Apr 2022 > …doubling our chances of survival will only take them from 0% to 0%. ### Resolution No resolution date was supplied. Biological humans remain extant; the prediction is still open. ### Context and outcome The current prediction is now certain extinction. Alignment will not be solved in time, and the labs will not coordinate before one of them builds AGI. No date is attached, so the prediction remains unresolved. Yudkowsky defines “dignity points” as improvements to humanity’s log odds of survival, however small. The instruction is to keep working on plans grounded in reality, say openly what you believe, refuse to abandon your ethics, and remain ready in case some unexpected opportunity appears. He also says that fighting for even a little longer counts as more dignified. His later public warnings therefore follow from the strategy rather than contradicting it. ### Supporting receipts > It's obvious at this point that humanity isn't going to solve the alignment problem, or even try very hard, or even go out with much of a fight. … doubling our chances of survival will only take them from 0% to 0%. Context: The post's bleak forecast is paired with a serious strategy: keep improving humanity's log odds without lying to yourself, abandoning ethics, or giving up. Source: Eliezer Yudkowsky, LessWrong (Apr 2022) — https://www.lesswrong.com/posts/j9Q8bRmwCgXRYAgcJ/miri-announces-new-death-with-dignity-strategy > That's not what surviving worlds look like. … There's no plan. Surviving worlds, by this point, and in fact several decades earlier, have a plan for how to survive. It is a written plan. The plan is not secret. Context: Items 41-42 of Yudkowsky's AGI Ruin: A List of Lethalities. Source: Eliezer Yudkowsky, LessWrong (Jun 2022) — https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of-lethalities > you're not going to get the answer you want until right before the end of the world, and maybe not even then unless you're paying very close attention. Timing this stuff is just plain hard. Context: Yudkowsky's dialogue dismissing biology-anchored AGI timeline forecasts as uninformative until the very end. Source: Eliezer Yudkowsky, LessWrong (Dec 2021) — https://www.lesswrong.com/posts/ax695frGJEzGxFBK4/biology-inspired-agi-timelines-the-trick-that-never-works ### Sources - Death with Dignity (2022): https://www.lesswrong.com/posts/j9Q8bRmwCgXRYAgcJ/miri-announces-new-death-with-dignity-strategy - AGI Ruin (2022): https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of-lethalities - bio-anchors post (2021): https://www.lesswrong.com/posts/ax695frGJEzGxFBK4/biology-inspired-agi-timelines-the-trick-that-never-works --- ## Self-driving cars: approved only after the world ends Canonical URL: https://foom.hyperplex.org/p/selfdriving Chapter: MIRI: Research Category: economics & deployment Predicted: 2021 Resolution date: 2025 Status: falsified by events Status detail: The prediction was resolved as falsified by events in 2025 Source: takeoff debate with Paul Christiano, Nov 2021 > …the World-Ending Company will be able to use the same tech to build a true self-driving car, which would in the natural course of things be approved for sale a few years later after the world had ended. ### Resolution Waymo runs paid driverless rides across five metros through ordinary regulation, built by a product team. World intact. ### Context and outcome His 'concrete example' of the doom ordering: true self-driving arrives via the AGI company's tech and gets regulatory approval only after the world has already ended. Elsewhere in the thread he offers the resolution criterion himself: 'Once you can buy a self-driving car, the thing that Paul predicts with surety and that I shrug about has already happened.' By 2025 Waymo was running paid, fully driverless robotaxi service across Phoenix, San Francisco, Los Angeles, Austin and Atlanta. It was approved through ordinary regulation, built by a non-AGI product team, and providing hundreds of thousands of rides a week. (Pedants' corner: a personal 'true self-driving car for sale' remains not quite purchasable; the ordering claim is dead either way.) ### Supporting receipts > what does it mean for "AI will be doing all of the AI R&D"? that sounds to me like something that happens after the end of the world, hence doesn't happen. Context: The same filing system puts AI doing AI R&D in the after-the-end bucket. It is now the daily practice of every frontier lab. Source: Takeoff Speeds debate (2021) — https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds > I furthermore rather expect that it will be a research project and a prototype, because the great gap between prototypes and commercializable technology will ensure that prototypes are much more advanced than whatever is currently commercializable. Context: The structural version: the world-ender arrives as a lab prototype far ahead of the market. The prototype-product gap collapsed instead. Source: Takeoff Speeds debate (2021) — https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds ### Sources - Takeoff Speeds debate (2021): https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds --- ## The trillion-scale AI boom: 'pretty darn prohibited' Canonical URL: https://foom.hyperplex.org/p/trillionboom Chapter: MIRI: Research Category: economics & deployment Predicted: 2021 Resolution date: 2025 Status: falsified by events Status detail: The prediction was resolved as falsified by events in 2025 Source: takeoff debate with Paul Christiano, Nov 2021 > this all sure does sound "pretty darn prohibited" on my model, but I'd hope there'd be something earlier than that we could bet on. ### Resolution Hyperscaler AI capex ≈ $400B/yr, Nvidia ≈ $4T, fabs a macro story. The 'prohibited' world arrived. Doom did not. ### Context and outcome “This” referred to Christiano’s picture of the years before an intelligence explosion: hundreds of billions of dollars spent automating AI research, semiconductor production expanding on an enormous scale, and speculative valuations directing a significant share of world investment toward AI. Yudkowsky said that such a world sounded “pretty darn prohibited” on his model. Most of that picture has since become ordinary news. Capital spending by the largest technology companies reached roughly $410 billion in 2025 and was projected to exceed $700 billion in 2026, driven heavily by AI infrastructure. Nvidia closed above a $4 trillion valuation in July 2025, while funding rounds for frontier laboratories grew into tens of billions of dollars. ### Supporting receipts > so a thing which Eliezer's Prophecy does not mandate per se, but sure does permit, and is on the mainline especially for nearer timelines, is that the world-ending prototype had no prior prototype containing 90% of the technology which earned a trillion dollars. a lot of Paul's Prophecy seems to be about forbidding this. is that a fair way to describe your own Prophecy? Context: The mainline explicitly permits no trillion-dollar commercial precursor. Today's frontier products are exactly that precursor. Source: Takeoff Speeds debate (2021) — https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds > stuff that we could see then: self-driving cars (10 years is enough for regulatory approval in many countries), super Codex, GPT-6 powered anime waifus being an increasingly loud source of (arguably justified) moral panic and a hundred-billion-dollar industry Context: His conditional sketch of a no-AGI 2031, delivered as a reductio, is arriving partially on schedule. Source: Takeoff Speeds debate (2021) — https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds ### Sources - Takeoff Speeds debate (2021): https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds --- ## Deepnets as ubiquitous as computers, maybe by 2035 Canonical URL: https://foom.hyperplex.org/p/deepnets Chapter: MIRI: Research Category: economics & deployment Predicted: 2021 Resolution date: 2025 Status: falsified by events Status detail: The prediction was resolved as falsified by events in 2025 Source: takeoff debate with Paul Christiano, Nov 2021 > if the world lasts until 2035, I could perhaps see deepnets becoming as ubiquitous as computers were in… 1995? 2005? would that fulfill the terms of the Prophecy? ### Resolution Computers-in-2005 ubiquity arrived by 2025, a decade ahead of his 'perhaps'. ### Context and outcome Yudkowsky suggested that, if the world survived, deep neural networks might by 2035 become as common in everyday life as computers had been in 1995 or 2005. That level of adoption arrived much earlier. Within two years of its launch, ChatGPT had hundreds of millions of weekly users, and by 2025–26 AI assistants were built into phones, browsers, office software, and development tools. Deep learning had reached something close to 2005-era computer ubiquity roughly a decade ahead of his estimate. Neither “perhaps” nor “if the world lasts” changes the timing. ### Supporting receipts > My model does not forbid that we dwiddle around on the AGI side while gradient descent tech gets its fingers into enough separate weakly-generalizing pies to produce 10% GDP growth, but I'm happy to say that this sounds much more like Paul's Prophecy is coming true. Context: The pre-registered concession condition was broad, weakly generalizing deployment producing real growth. Reality took his own out. Source: Takeoff Speeds debate (2021) — https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds ### Sources - Takeoff Speeds debate (2021): https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds --- ## Codex: stagnation for 5 years or a massive leap in 2 Canonical URL: https://foom.hyperplex.org/p/codexbet Chapter: MIRI: Research Category: economics & deployment Predicted: 2021 Assessment date: 2023 Status: hedged: couldn't lose Status detail: The prediction was assessed as hedged: couldn't lose in 2023 Source: takeoff debate with Paul Christiano, Nov 2021 > 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 in 2 years and got ubiquitously adopted by lots of programmers. ### Resolution Reality took the massive-leap branch, on the 2-year clock. But a claim licensing both branches can never lose, so it doesn't win. ### Context and outcome Christiano expected spending on Codex to rise quickly to something under $10 million a year, then double annually. Yudkowsky said he would be unsurprised by either five years of little progress or a dramatic leap to widespread use within two years. In other words, both stagnation and rapid growth were treated as compatible with the model. Events followed the second path more closely. By early 2023, GPT-4 could handle substantial coding tasks, and by 2024 coding assistants were in widespread use, growing far faster than Christiano's simple doubling curve implied. I would still score the prediction as yellow rather than green, as Yudkowsky did identify relatively unregulated products such as Codex or, in his phrase, “anime waifus,” as likely places for a breakthrough. But the larger claim was broad enough to accommodate opposite outcomes, which was essentially Christiano’s objection to the model. ### Supporting receipts > well, if I was betting on something taking a big jump in income, I sure would bet on something in a relatively unregulated industry like Codex or anime waifus. Context: The adjacent claim did stick its neck out by naming the breakout niche, while declining to actually bet. Source: Takeoff Speeds debate (2021) — https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds ### Sources - Takeoff Speeds debate (2021): https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds --- ## $1B/yr waifutech before AI freedom-of-the-road Canonical URL: https://foom.hyperplex.org/p/waifubet Chapter: MIRI: Research Category: economics & deployment Predicted: 2021 Resolution deadline: No date supplied Status: still open Status detail: The prediction is still open; no resolution deadline was supplied Source: takeoff debate with Paul Christiano, Nov 2021 > I think my sheer impression would be that we'll get $1B/yr waifutech before we get AI freedom-of-the-road ### Resolution Neither leg crossed yet; robotaxis are arguably ahead, the opposite of the sheer impression. Still open. ### Context and outcome Yudkowsky’s impression was that “waifutech” would reach $1 billion a year before self-driving cars achieved “freedom of the road.” Christiano pressed him to define the automotive side in terms of revenue from services like Waymo, although the two never settled on precise terms. As of mid-2026, neither prediction had clearly come true. Waymo’s annualized revenue had risen above $350 million, while reported revenues for companion and character-based AI products remained scattered and difficult to compare. Autonomous driving appears to be ahead in actual commercial deployment, but the original comparison is too loosely defined for a clean verdict. With no deadline attached, it remains open. ### Supporting receipts > I'd be pretty surprised if this surpassed $1B/year of revenue before self-driving cars passed $10B/year of revenue. Context: Christiano's counter-prediction in the same exchange is the leg currently ahead. Source: Christiano, same thread — https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds ### Sources - Takeoff Speeds debate (2021): https://www.lesswrong.com/posts/vwLxd6hhFvPbvKmBH/yudkowsky-and-christiano-discuss-takeoff-speeds