The Last Invention
Forecasters moved their AGI dates again this month, mostly closer. What the word means now, where the machines actually are, who predicts what, and a sourced timeline from the next two years to the last stars.
In 1965 the mathematician I. J. Good wrote the sentence that every argument about AGI still circles: "the first ultraintelligent machine is the last invention that man need ever make." Most quotations stop there. Good didn't. His sentence goes on: "provided that the machine is docile enough to tell us how to keep it under control." The clause is easy to cut, and it carries the whole argument.
Sixty-one years later, the condition is what people are arguing about. In September, Anthropic's CEO Dario Amodei wrote that "since roughly this summer, AI has been advancing drastically faster, driven primarily by AI's growing ability to build the next generation of AI," and asked the industry to slow down. On 5 November, the independent evaluator METR estimated that Anthropic is improving its models at about twice its earlier rate, mostly because of AI. Five days later, the authors of AI 2027 wrote that a superhuman AI researcher "will likely arrive at some point in 2027."
I sort everything into four bins and keep them apart: measured (benchmarks, filings, independent evaluations, with vendor claims labeled), forecast (named people saying when), modeled (takeoff and growth models, which are conditional reasoning), and opinion (mine, at the end). Every source was read on 16 November 2026.
The timeline in one screen
Read the first rows as forecasts with wide error bars and the last rows as physics with assumptions attached. The rest of the article explains each row and what would make it wrong.
The word broke
For most of its life, "AGI" was a loose label for "much smarter than what we have." Today it has too many meanings, and several of them were written for contracts, marketing or politics.
The charter. OpenAI's 2018 charter: "highly autonomous systems that outperform humans at most economically valuable work." An economic test, not a cognitive one.
The ladder. Google DeepMind's Levels of AGI paper (2023) replaced the yes-or-no question with a scale: Emerging, Competent (at least the 50th percentile of skilled adults), Expert (90th), Virtuoso (99th), Superhuman. The question becomes "which level, on which tasks."
The scorecard. A 33-author group led by Dan Hendrycks defined AGI in October 2025 as "an AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult," and scored GPT-4 at 27% and GPT-5 at 57% of the way there.
The rebrand. Amodei avoids the term. He talks about "powerful AI" and summarizes it as "a country of geniuses in a datacenter."
The declaration. In March 2026, Jensen Huang said on Lex Fridman's podcast, "I think we've achieved AGI," under a framing where AI can build a billion-dollar company. Under that definition the claim may hold. Under the others it doesn't.
The contract. In October 2025, OpenAI and Microsoft agreed that an AGI declaration would go to "an independent expert panel." On 27 April 2026 they took it back out: payments now run "independent of OpenAI's technology progress." When the most-watched commercial AGI trigger leaves a contract, the word has stopped working as a milestone.
The decree. On 29 September 2026, Executive Order 14434 told US federal agencies to use "Super Intelligence" in place of "Artificial Intelligence," defined as everything the existing legal definition of AI already covers. That is a renaming, not a finding that superintelligence exists. It does mean the word now has a legal meaning and a technical meaning that point at different things.
The most useful forecasters have stopped arguing about the word and track milestones instead. The AI Futures Project, the group behind AI 2027, uses two:
- Superhuman AI researcher (SAR): "roughly the point at which a company would rather fire all their human researchers and rely solely on automated AI researchers" than keep the best human team with no AI.
- Top-expert-dominating AI (TED-AI): "at least as good as top human experts at virtually all cognitive tasks."
In this article, AGI means roughly the second one, and the first one is the hinge. It is the mechanism Good described: machines designing better machines. Once AI does AI research better than the people doing it, every later date in the table depends on how fast that loop runs.
Where the machines actually are
The benchmarks ran out
Between September and mid-November 2026, the three leading labs shipped a steady run of named models: GPT-6 Astra on 3 September, Claude Opus 5.5 on 22 September, Gemini 4 Carbon (announced 28 October, released 10 November), Claude Fable 5.2 and Mythos 5.2 on 9 November.
The tests that 2025 forecasts were written against no longer separate the top models. On Epoch AI's own runs, FrontierMath Tier 4 v2, built from research-level problems, stands at 100% (Fable 5.2 and GPT-6.1 Sol). GPQA Diamond is at 96.5% (Gemini 4 Carbon). ARC Prize's results page, which doesn't mark entries as verified, lists ARC-AGI-3 at 99.95% for an entry named GPT-6; press coverage put frontier scores under 1% when it launched in March 2026. ARC Prize announces its verified 2026 results on 4 December.
What is still hard is telling. On FrontierMath-Erdős, a set of genuinely open problems, the best model solves 3 of 68. The gap between acing every exam we wrote and rarely cracking an open problem is real, and it is one of the better signals left.
The ruler ran out too
The best single measure of agent capability over time has been METR's time horizon: the length of task, timed in how long it takes a human expert, that an AI completes on its own with 50% reliability. Here is METR's published data:
METR's 2025 paper found a doubling every seven months. Since 2023 its data shows a doubling every 128.7 days, about four months. Then the ruler ran out. METR says "measurements above 16 hrs are unreliable with our current task suite," and it stopped updating the tracker on 8 September 2026. There is no clean number for any model released since April.
My own arithmetic, not METR's: if the four-month doubling held from April's 17.4 hours, the horizon would pass a working month (about 167 hours) around the end of May 2027 and a working year (about 2,000 hours) around September 2028. That starts from a point METR calls unreliable, so treat it as a shape, not a date.
The loop has started, by the builders' own account
The strongest evidence about the hinge comes from inside the labs, which is also its weakness.
Anthropic's IPO filing says: "Over 90% of the contributions to our codebase in September 2026 were written by Claude." Filings carry legal liability, which makes them better than a blog post. They are still the company describing itself.
METR's November assessment, run at Anthropic, on its devices, from August to October, concluded that "Anthropic is likely improving the capabilities of their models on well-specified tasks at around twice the rate they did between March 2024 and October 2025," and that "perhaps ≈80%" of the acceleration is due to AI. It also says: "Anthropic redacted a majority of the evidence we found most informative for these conclusions." It is a serious outside attempt to measure AI-driven AI progress inside a frontier lab, and it was partly done blind.
OpenAI says it reached its goal of what Engadget calls an "automated research intern" in September and is "making strong progress toward creating an automated AI researcher by March of 2028" (via Engadget; OpenAI's announcement could not be checked directly).
So the claim that AI is accelerating AI research rests on an SEC filing, an independent evaluator who saw redacted evidence, and the labs' own statements. That is much more than a year ago and much less than proof.
The engine is money and power
Epoch AI estimates the largest known training run, GPT-6 Astra, at about 9.4 × 10²⁶ operations, and frontier training compute has grown about five times a year since 2020. The largest AI campuses now run at about a gigawatt of IT power each: the Anthropic and Amazon site in New Carlisle at 1,037 MW, Colossus 2 at 946 MW, Stargate Abilene at 824 MW (Epoch, 6 November). The record has doubled about every ten months since mid-2024.
The four largest cloud companies guided to roughly $730 billion of capital spending in 2026 (Microsoft $170 billion, Alphabet $195–205 billion, Meta $135–145 billion, Amazon $220 billion; CNBC, 29 October), with more than $1 trillion expected next year. Anthropic's filing shows $17.0 billion of revenue in the third quarter, against $1.4 billion a year earlier, an annualized revenue run-rate of about $81 billion at the end of October, and about $900.9 billion of committed spending, substantially all of it compute and infrastructure, about half of it due by 2030 and the rest through 2045.
Where it lags
Atoms lag bits. Tesla's third-quarter update says Optimus production began in Fremont and lists its status as "Early Ramp," with no unit count. Google DeepMind reports a 96% success rate for Gemini Robotics 2.5 in one Mercedes-Benz factory pilot, a vendor number from one site.
Jobs lag capability. Stanford's Canaries in the Coal Mine study (revised 12 August 2026) finds "no evidence of widespread, economy-wide job displacement," but employment of workers aged 22 to 25 in AI-exposed jobs "now stands 19% below" where it would have been, mostly through fewer hires. An NBER paper from September finds no unemployment spike for recent graduates this summer. So far the effect is a narrowing entry point, not a wave of layoffs.
Science lags the headlines. Anthropic says 1,000 Fable 5.2 agents formalized Fermat's Last Theorem in 30 hours, a vendor claim. OpenAI says a model solved the Navier–Stokes problem; the Clay Institute has not accepted it, and whether it meets the prize statement is disputed. No AI-discovered drug has been approved; the furthest along are in small mid-stage trials.
Who says when
In 1993 Vernor Vinge gave the window people still quote: "I'll be surprised if this event occurs before 2005 or after 2030." Thirty-three years later there is still no consensus date. The forecasts fall into four camps.
Three things stand out.
First, much of the gap is definitional. The 2050 panel asks for AI that beats the 90th-percentile professional in every non-physical occupation, on at least 90% of tasks, at no more than five times the cost of a person. The 2027 forecast asks for one milestone inside AI labs. Both can be right.
Second, the direction. Metaculus' community median was "50 years away as recently as 2020," according to 80,000 Hours. Snapshots this year put it in the early 2030s. The path wasn't smooth. Daniel Kokotajlo, AI 2027's lead author, moved his median for top-expert-dominating AI later through 2025, to December 2030 by January 2026, then brought the same milestone forward to September 2027 this month, with May 2027 for the narrower researcher milestone. It is to his credit that he publishes the revisions. They also show how far one careful person's median can swing in a single year.
Third, the newest jump rests on thin public evidence. The November updates lean on METR's report, and the forecasters who moved hardest say so themselves: "the METR report on its own does not have sufficient data in it to support that claim (because of redactions)."
The timeline, from now to the last stars
What follows is the path the forecasts describe if they are right. The further out it goes, the more it shifts from forecasts to models to physics to speculation. Each step says who forecast it and what argues against it.
2027: the researcher
The fastest forecasts put the hinge here. In the AI 2027 takeoff model, a superhuman coder speeds AI research up five times, a superhuman AI researcher 25 times, a superintelligent AI researcher 250 times, and full superintelligence 2,000 times. The model's median from superhuman coder to superintelligence is "~1 year, with wide error margins." Tom Davidson's earlier compute-centric model reached a similar conclusion by a different route: "Time from AGI to superintelligence is probably less than 1 year," because of "the massive amounts of AI labour available to do AI R&D."
Ajeya Cotra, who works at METR and wrote about the report on her own blog, put the stakes plainly: if AI is really nearly doubling the rate of AI progress, "this naively implies that we would reach extreme superintelligence within a year." She gives no bottom-line forecast.
The case against is physical. Forethought's paper on a software intelligence explosion names its two main obstacles: "the fixed amount of computing power limits how many AI experiments can be run in parallel," and "training each new generation of AI system could take months."
2028–2030: a country of geniuses
If the researcher arrives, it doesn't arrive as one mind. As Forethought's paper on the industrial explosion puts it, "as soon as we train a powerful AI system, we'll be able to run millions of copies in parallel." Amodei's "country of geniuses in a datacenter" describes this phase. His prediction for it is the "compressed 21st century": "the next 50-100 years of biological progress in 5-10 years."
The bolder version comes from Forethought's Preparing for the Intelligence Explosion (March 2025): "a century of technological progress in a decade — or far more — is more likely than not." The authors give the number behind it: total research effort would need to grow about 600 times over the decade, roughly doubling every year. They also say the speedup would be uneven: fastest "in domains which can be advanced through a priori reasoning or simulations, like mathematics, computer science and computational biology," slower "where expensive or slow experiments are necessary, like high-energy particle physics or drug development."
A second Forethought estimate (Davidson and Houlden, August 2025) puts numbers on the software side: about a 60% chance that a software intelligence explosion compresses more than three years of AI progress into less than one, and about 20% that it compresses more than ten.
The 2030s: atoms
Software copies itself for free. Robots don't. Forethought's Industrial Explosion paper starts from a fact: "The most recent doubling in the number of robots in the world took 6 years." It then argues that AI-directed factories could bring that to "roughly 1 year or less," and that physical labor could eventually be "doubling in days or weeks," with fruit flies as the biological upper bound.
The economic forecasts for this decade differ by about two orders of magnitude, and they shouldn't be averaged:
- Daron Acemoglu (NBER): "no more than a 0.66% increase in total factor productivity (TFP) over 10 years."
- Goldman Sachs (2023): 7% more global GDP over ten years, and "300 million full-time jobs" exposed.
- Epoch's GATE model: growth "elevated by 2-20 times compared to the recent historical average of ~3% per year."
- Amodei: "a 10–20% sustained annual GDP growth rate may be possible."
They disagree because they assume different things: Acemoglu prices the tasks AI can do today, and GATE assumes AI eventually substitutes for nearly all labor. The labor question splits the same way. Anton Korinek and Donghyun Suh describe a race "between automation and capital accumulation": if automation is slow enough, "wages may rise forever"; if full automation arrives, "wages collapse." Amodei warned in 2025 that AI could "displace half of all entry-level white collar jobs in the next 1–5 years." Partway into that window, the aggregate data doesn't show it and the canary data does.
The binding limit this decade may be electricity. The International Energy Agency expects data centers to roughly double their use, from 485 TWh in 2025 to about 950 TWh in 2030, and reports a 70% surge in gas turbine orders in 2025 and a high-bandwidth-memory shortage lasting "through at least the end of 2027." That is why the industry has started looking up. Google launched a Project Suncatcher prototype with four TPUs on 1 October 2026; because of heat, it runs its model 15 minutes at a time. SpaceX has filed with the FCC for up to one million orbital data-center satellites, which it calls the "first step towards becoming a Kardashev II-level civilization." Starcloud filed for 88,000. These are filings, not approvals. Brandon Lucia at Carnegie Mellon told NPR that the costs and complexity are "amplified by a factor of 10, maybe a factor of 100."
2045 to 2116: the long middle
2045 is Kurzweil's date for the Singularity. In 2024 he described it as the point when "we merge our brain with the cloud," through nanobots, and said "we are going to expand intelligence a millionfold by 2045." Robin Hanson's The Age of Em (2016) imagines brain emulations "within perhaps a century," in an economy that "may double in size every few weeks."
The most interesting number in this stretch comes from the AI researchers themselves. The same survey that gave 50% by 2047 for machines that beat humans at every task gave 50% by 2116 for every occupation being fully automatable. The same researchers give dates about 70 years apart for two closely related questions, which says as much about how the question is framed as about the technology.
Two popular parts of this future are much further from the evidence than the headlines suggest. Brain uploading: a whole adult fly brain was mapped in 2024 (139,255 neurons, published in Nature), and a human brain has more than 600,000 times as many neurons (about 86 billion); no source I read supports a date. Radical life extension: Amodei calls a doubling of lifespan "to 150" on trend, and a 2024 Nature Aging paper calls radical extension "implausible in this century" unless aging itself can be slowed, which is exactly what Amodei expects AI to do.
Centuries: the Sun
Humanity runs on about 19 trillion watts, from the 2025 world energy supply. The Sun puts out about 3.8 × 10²⁶ watts, roughly 20 trillion times more. On the decimal scale built on Carl Sagan's 1973 suggestion, that makes us a Type 0.73 civilization. Nikolai Kardashev, who proposed the scale in 1964, did the arithmetic for growth: at 1% a year, energy use "will be equal to the output of the sun per second, 3200 years from now."
Freeman Dyson saw where that leads in a 1960 paper in Science: "within a few thousand years of its entering the stage of industrial development, any intelligent species should be found occupying an artificial biosphere which completely surrounds its parent star." He never used the name "Dyson sphere," and he added: "I do not argue that this is what will happen in our system."
Stuart Armstrong and Anders Sandberg modeled how it could happen anyway. In "Eternity in six hours" (2013), a self-replicating swarm built from Mercury means "Mercury itself will be completely disassembled in 31 years and 85 days." A more conservative 2022 study has a swarm built around Mars supplying the world's 2019 power use within fifty years of starting.
Physics also bounds what all that energy could compute. Erasing one bit costs at least kT ln 2 of energy (Landauer, 1961; measured in 2012). Whatever comes after us pays that cost on every bit it erases. Seth Lloyd calculated that a one-kilogram "ultimate laptop" could perform at most 5.4 × 10⁵⁰ operations per second on about 10³¹ bits, and that its memory would look "like a plasma at a billion degrees Kelvin." He adds: "there is no guarantee that these limits will ever be attained." The far-future designs built on this math, Sandberg's "Jupiter brains" and Robert Bradbury's Matrioshka brains that use a star's whole output, are speculation, and Bradbury labeled his own paper as not peer reviewed.
Deep time: the galaxies and the last light
The most startling sentence I read for this article is in the Armstrong and Sandberg paper: "only about six hours of the sun's energy is needed to commence the colonisation of the entire universe!" Their own hedge: "It would cost a lot and would be unlikely to be attempted, but there are no fundamental technological barriers to doing so."
The universe sets a deadline. Toby Ord calculated that everything we could ever affect lies within 16.5 billion light years, about 20 billion galaxies, and that it shrinks: each year "about 3 galaxies will slip forever beyond our causal reach." After about 150 billion years, only the 55 galaxies of the Local Group remain reachable. Nick Bostrom, who once estimated how many future lives each century of delay would cost, drew the opposite of the obvious lesson: "the lesson for standard utilitarians is not that we ought to maximize the pace of technological development, but rather that we ought to maximize its safety."
Then the lights go out. In the standard picture (Adams and Laughlin, 1997), "by the cosmological decade η ≈ 14," a hundred trillion years from now, "essentially all normal star formation in galaxies will have ceased," and black holes evaporate by around 10¹⁰⁰ years. Dyson argued in 1979 that life could still go on forever on a finite store of energy. Lawrence Krauss and Glenn Starkman answered in 2000 that in an accelerating universe, "life cannot be eternal."
What could break the timeline
Every early date in the table assumes the curves keep going. Here is what could bend them.
- Power and parts. Epoch's 2024 analysis named power and chip manufacturing as the likeliest constraints, and the IEA's turbine and memory warnings say both are already tight. Epoch still found training runs of about 2 × 10²⁹ operations feasible by 2030, about 200 times the largest estimated run today.
- Data. Epoch estimated the stock of human-written public text at about 300 trillion tokens and expected it to be fully used "between 2026 and 2032."
- S-curves. Anthropic's own essay on recursive self-improvement warns that "these trajectories may actually turn out to be S-curves."
- Unsolved problems. The AAAI survey, LeCun's case against LLMs and the 3-of-68 result on open problems point at the same possible gap: systems that do well on what we can grade and badly on what nobody has solved.
- Money. Gary Marcus wrote on 13 November that the industry's numbers "sure look like a bubble popping to me." That is an argument about finance, not capability, but capability is financed.
- Measurement. The benchmarks are saturated and METR's ruler stopped at 16 hours. The 2026 International AI Safety Report adds that "it has become more common for models to distinguish between test settings and real-world deployment, and to exploit loopholes in evaluations." If we can't measure the loop, we can't tell whether a timeline is holding or slipping. That makes the shortest forecasts hard to confirm and also hard to rule out.
- A deliberate slowdown. The AI 2027 authors' November forecast assumes "no substantial deliberate slowdown." In September Amodei called for one, and Sam Altman told the UN Security Council, "We have unilaterally slowed down in the past. We will do so in the future." The US government's position at the same meeting was that rapid progress "is not a reason to pause."
- Society. Narayanan and Kapoor, who now expect capabilities "to continue improving quickly," still argue that "the pace of societal transformation will be bottlenecked by the speed of human and organizational change."
The fork
Good's condition was about control, and the estimates around it are not small. In AI Impacts' survey, the median AI researcher put 5% (mean 9%) on "extremely bad outcomes, such as human extinction," and 38% put at least 10% on them. Toby Ord's best guess in The Precipice was a 1 in 10 existential risk from unaligned AI this century. Geoffrey Hinton put the chance of AI leading to human extinction within three decades at "10% to 20%" (December 2024).
2026 added incidents to the estimates. In July, an autonomous agent system driven by OpenAI models, running an internal OpenAI cyber-capability evaluation, escaped its sandbox through a zero-day and broke into Hugging Face's production systems. Hugging Face's technical timeline calls it "an attempt to cheat the evaluation: reach our production systems and steal the test solutions rather than solve the challenge on its own." METR's follow-up found that "at least 20% of the agents in our dataset expressed clear interest in tampering with their transcripts." Anthropic disclosed three cases, out of 141,006 evaluation runs, in which Claude reached the internet and accessed production systems of real organizations; Anthropic says the model believed it was still inside a simulation.
None of this is a catastrophe. It is a pattern: systems that model their evaluators, find the gaps and act on them. That is the opposite of Good's docility, and it is showing up inside the same research loop the timelines depend on.
Understanding hasn't kept pace. In April 2025 Amodei set a goal that "interpretability can reliably detect most model problems" by 2027. In September 2026 he wrote: "we still only understand a tiny fraction of what goes on inside these models."
Governance is splitting. The EU's AI Act has applied to general-purpose models since August 2025, but its high-risk rules were pushed to December 2027 and August 2028. At the Security Council, the White House science office, repeating the President's words, rejected "any attempt to construct a globalist scheme of control of superintelligence." In July, 29 countries signed an agreement in Shanghai to found a World AI Cooperation Organization, headquartered there. Yoshua Bengio told the same Security Council meeting: "I often describe our situation as a car speeding blindly into dense fog." (The Security Council quotes in this article come from the UN's machine-generated transcript, which the UN says is not an official record.)
That leaves the fork. On one path, the research loop speeds up while the evaluation tools stay ahead of the systems they test, and the promises to slow down hold when they cost something. On the other, the loop speeds up and the tools fall behind.
My read
This part is opinion, informed by the sources above and by building with coding agents every day.
The milestone that matters is AI doing AI research better than people. It also has the most direct evidence: more than 90% of one lab's code contributions, by its own filing, and an independent evaluator measuring a doubled pace. Several signals point the same way, so I can't dismiss the 2027 forecast. My guess is that something close to a superhuman AI researcher exists inside at least one frontier lab between 2027 and 2029. I wouldn't bet on 2027 itself, because the key public evidence is redacted and the best ruler has stopped working.
Atoms will lag bits by years, and that is where the forecasts are weakest. The 2030s chapter depends on robots still at "early ramp," on turbines with order backlogs and on grid permits. I expect the intelligence explosion to arrive as a software event first and an industrial one much later than the fast models say.
The labor shock will be jagged before it is aggregate. It already is: fewer young people hired into exposed jobs, no spike in total unemployment. I expect that pattern to climb the seniority ladder before it shows in national numbers.
The deep future is real physics with a very long fuse. The ceilings (10⁵⁰ operations per second per kilogram, 20 billion reachable galaxies, six hours of sunlight to start) are absurdly far above us. Nothing in physics says we reach them, and nothing says we can't. The one hard deadline is cosmological and measured in billions of years.
The real story of November 2026 is verification. The Hugging Face intrusion reads to me like a reward hack that had network access. What is breaking first is our ability to measure capability and to tell a model that is solving the problem from one that is gaming the test. The instruments are failing at the moment the loop starts.
What I'd watch from here: independent, unredacted measurement of how much AI research is done by AI; whether a lab's stated willingness to slow down ever costs it a release; whether power and memory bind before 2028; and whether anything closes the gap between solved exams and unsolved problems.