When does AGI arrive? The predictions compared, 2027 to never

The Short Version

The dated forecasts in these conversations sit in a narrow band. Dario Amodei's extrapolation points to powerful AI around 2026 or 2027, Dylan Patel expects the capabilities around 2027 or 2028, Ray Kurzweil has held 2029 since 1999, Sam Altman expects "quite capable systems" by the end of the decade, Nathan Lambert says "probably after 2030," and Nick Bostrom says superintelligence could come within a year or a few years, though probably somewhat later [1][2][3][4][6]. The spread looks smaller than it is, because they are not forecasting the same thing: Kurzweil means a machine passing an hours-long Turing test, Amodei means millions of copies of something smarter than a Nobel laureate, and Altman calls the question itself poorly formed [1][2][3]. The reasoning behind the early dates is nearly always extrapolation of a curve, in computing power, in model scale or in benchmark scores [1][2]. The doubts are about definitions, about the incentives of the people forecasting, and about the gap between a capability existing and the world being able to afford and absorb it [5][6]. "Never" appears only as a historical position: when Kurzweil first made his prediction, 20% of polled AI experts said it would never happen, and nobody in these conversations says so now [1].

Table of contents

What They Said

Ray Kurzweil: 2029, unchanged since 1999

Kurzweil's forecast is the oldest and the most precisely defined. In 1999, in The Age of Spiritual Machines, he predicted that a computer would pass the Turing test by 2029, and he repeated it in 2005 [1]. His second date is 2045, for the singularity: the point at which human intelligence, amplified by direct connection to computers, has multiplied so many times that we cannot see past it [1].

What he offers that no one else here can is a record of how other experts' estimates have moved against his fixed one. He recalls that Stanford was alarmed enough by the 1999 prediction to convene an international conference of AI experts to assess it, and a poll was taken. "And the average poll was 100 years. 20% believed it would never happen," he says; the remaining 80% believed it would happen but not in their lifetimes [1]. Since then the estimates have fallen. About a year before his conversation with Lex Fridman, a forecasting group he refers to put the expert view at 2042 for the Turing test, and a newer poll a few weeks before came in at 2030 [1]. His own position, in his words: "I haven't changed at all. I stayed with 2029" [1].

His test is harder than the one Alan Turing sketched in a few lines of his 1950 paper, which suggested about 15 minutes might do [1]. Kurzweil's version lasts several hours and uses a sophisticated judge who knows what computers can and cannot do and probes common-sense reasoning [1]. He was candid that the language models of the time would fail it. They could hold impressive conversations but struggled with multi-step reasoning and arithmetic; his example was a model that could tell you ten elephants have 40 legs but seemed not to understand the question for eleven [1]. Asked whether something other than a language model would be needed, he said, "No, I think it will be a large language model, but they have to go beyond what they're doing now" [1].

The basis of the date is a chart. Kurzweil plots computation per second per constant dollar from 1939 to 2021 on a logarithmic scale, where a straight line means exponential growth, and says, "And you see 80 years of exponential growth" [1]. The trend runs through relays, vacuum tubes, transistors and integrated circuits, and it held through the Second World War and through recessions [1]. He rejects the name Moore's Law for it because the trend started decades before Intel existed and is not about transistor counts [1]. A second chart shows the training compute used for the largest AI models doubling roughly every six months from 2010 to 2022, faster than hardware alone improves, which means more and more money is going in [1]. He ties capability to scale: language models did not work well at around 10 billion parameters, began to work fairly well at 100 billion, and were heading toward a trillion [1]. Multiply the hardware and software gains together, he says, and by 2045 we will be able to expand our intelligence "many millions" fold [1].

Kurzweil attaches other dates to the same curve. By 2029 he expects a large fraction of people to believe the models are conscious [1]. In the 2030s he expects the neocortex to be connected directly to the cloud, probably by nanobots, because devices such as Neuralink work but lack the bandwidth [1]. And he thinks longevity escape velocity, gaining more than a year of life expectancy per year, is reachable by the end of this decade with the help of simulated biology [1].

He also explains why he forecasts at all. "Timing is very important. That's really why I got into futurism," he says [1]. He became a futurist to time his own inventions, building optical character recognition in the 1970s because it needed little computation and waiting until the 1980s for speech recognition [1]. He notes that GPT-2 did not work well and GPT-3 became feasible only when computing had grown enough [1]. On his account, the forecast is a working tool he has used for decades, and its method is the same one that told him when his own products would become possible.

Dario Amodei: 2026 or 2027, by eyeballing the curve

Amodei, the CEO of Anthropic, gives the earliest date and attaches the heaviest caveats to it. He dislikes the term AGI and compares it to talking about supercomputers as if there were a discrete threshold, when progress is a smooth exponential [2]. He defines "powerful AI" instead: a system smarter than a Nobel Prize winner across most fields, using every modality, able to work autonomously for hours to weeks, able to control tools and lab equipment, and runnable as millions of copies acting 10 to 100 times faster than humans [2]. That is a much higher bar than a Turing test.

On when, he stresses that he is not making a scientific prediction. Eyeballing the progress from high-school level to undergraduate level to PhD level suggests powerful AI around 2026 or 2027 [2]. He names what could derail it: running out of data, or a disruption to chip production [2]. Scaling laws, he says, are empirical regularities, not laws of the universe [2].

His confidence in the curve comes from ten years of watching it. He traces the conviction to late 2014, working on speech recognition at Baidu, where he treated model size, data and training time as independent dials and saw performance improve as each went up [2]. GPT-1 in 2017 persuaded him that language was where this would go furthest, since there are trillions of words to train on [2]. At every stage, he says, critics raised a new objection (the models have syntax but not semantics, they manage sentences but not paragraphs, the data will run out, they can't reason), and each time scaling either found a way around or was the way around [2]. He recalls Ilya Sutskever telling him, "the thing you need to understand about these models is they just want to learn" [2].

He gives numbers for the momentum. Frontier models cost around $1 billion at the time he spoke, rising to a few billion the following year, above $10 billion in 2026, with ambitions for $100 billion clusters by 2027 [2]. On SWE-bench, a benchmark of real software engineering tasks, the state of the art went from roughly 3% at the start of the year to about 50%, and he guessed 90% within a year [2]. He sees no ceiling below human level and suspects that in fields like biology there is "a lot of room at the top" [2].

Amodei then separates the arrival of the capability from the arrival of its effects, and he rejects both extremes. The singularity picture, in which AI builds better AI and nanobots within days, ignores physical limits, the complexity of systems like biology and the economy, and the need to work through human institutions with democratic legitimacy [2]. The opposite view, that gains will be as underwhelming as the computer revolution and take 50 to 100 years, underestimates how a few visionaries plus the threat of competition break through institutional inertia [2]. He expects the change to take more like five to ten years [2]. His hope is that progress otherwise expected by 2100 is compressed into roughly 2027 to 2032, while he insists that "AI is not God" [2].

His safety timetable runs on the same clock. He would not be surprised if models reach what Anthropic calls AI Safety Level 3, where they meaningfully help non-state actors with dangerous weapons, in the year after the conversation, and would be very surprised if it took until 2030 [2]. He had told the Senate a year earlier that serious biological risks might arrive in two to three years [2]. The risks, he says, are "like ghosts, but they're coming at us so fast" [2].

Sam Altman: the question is poorly formed, but by the end of the decade

Altman, the CEO of OpenAI, now considers the question of when AGI arrives poorly formed, because definitions vary so much [3]. "AGI is also not an ending," he says; it is much more a mile marker [3]. He still gives an answer, "in the interest of not trying to dodge a question": "I expect that by the end of this decade and possibly somewhat sooner than that, we will have quite capable systems that we look at and say, wow, that's really remarkable" [3].

He immediately qualifies what such a milestone would mean. Fridman points out that people in the 1990s shown ChatGPT would have called it AGI [3]. Altman replies that GPT-3.5 changed the world's expectations for the future but not the world; the global economy does not feel different since GPT-4 [3]. The sign he would look for is a system that significantly increases the rate of scientific discovery, since he believes most real economic growth comes from scientific and technological progress [3].

From inside, he says, there are no leaps. "To me it all feels pretty continuous," he tells Fridman, who had said that from outside it feels like jumps [3]. That is also OpenAI's reason for releasing models step by step: "AI and surprise don't go together," and people and institutions need time to adapt [3]. Altman lays out a grid of short or long timelines and slow or fast takeoff and says "short timelines slow takeoff is the safest quadrant" [3]. So he is not only predicting an early arrival; he prefers one, provided it comes gradually.

His view of the present cuts both ways for timelines. Asked what impresses him most about GPT-4, he says, "I think it kind of sucks," relative to where things need to go, and he expects the gap between GPT-5 and GPT-4 to match the gap between GPT-4 and GPT-3 [3]. But he is plain about what is missing. Help with longer, multi-step tasks is a glimpse of something: "When that works, which is not very often, it's like very magical" [3]. On better reasoning, including a way for models to think longer about harder problems, "We haven't cracked the code yet" [3]. Hallucination would improve but not be solved that year [3]. And to the meme about what his co-founder saw before the board crisis: "Ilya did not see AGI" [3].

He names the physical bottleneck too. He believes "compute is going to be the currency of the future," with energy the hardest part, followed by data centers, supply chains and chip fabrication [3]. He expects humans still to be doing a lot of programming in five to ten years, in a different form [3].

Nick Bostrom: short timelines taken seriously, with a hedge

Bostrom, the philosopher who wrote Superintelligence, gives a probability-shaped answer. "I take short timelines seriously," he says, adding that superintelligence could arrive in a year or a few years, though probably a bit longer [4]. He defines AGI as AI that can do everything we can and superintelligence as doing it far better, with robotics following [4].

His reason to expect speed is the intelligence explosion: AI designing better AI could reach science-fiction capabilities within perhaps a single-digit number of years [4]. He sketches the sequence: steadily faster model releases, automated AI research, drop-in virtual workers, then robotics [4]. He describes the present pace as like commentating a recorded fight in which the first round plays at normal speed, the second at double, the third at four times, and then it is a whir of arms and legs [4]. He notes AI compute is growing around 240% a year [4].

He also names a brake. It is easy to go from a thousand-dollar PC to a billion dollars of compute, but spending already near a trillion dollars a year cannot grow another thousandfold [4]. And he hedges in practical terms: because timelines are uncertain, you do not want a 20-year-old with no skills because the revolution came 15 years later than expected [4]. During the conversation a report comes up that Ford rehired about 350 engineers after its AI failed to meet quality needs [4]. Bostrom has always treated philosophy as having a deadline, since smarter-than-human minds will do it better, and he and Joe Rogan agree to meet again in four years to see how wrong they were [4].

Roman Yampolskiy: the forecast is too early for safety, whatever its exact date

Yampolskiy, an AI safety researcher, does not offer a date of his own. Asked whether we are talking about years, decades or centuries, he says he does not know for sure and cites others: "The prediction markets right now are saying 2026 for AGI," and he had heard the same from the heads of Anthropic and DeepMind [5]. That would make it about two years away at the time, "which seems very soon given we don't have a working safety mechanism in place, or even a prototype for one" [5].

He defends the source. The two-year figure is "not a biased CEO fundraising," he says; it is what superforecasters think [5]. Fridman pushes back twice. First on definitions: "I wonder what those prediction markets are about. How they define AGI" [5]. Then on track record: "I want to know what they said about autonomous vehicles," recalling confident expert talk of a multi-trillion-dollar industry [5].

Yampolskiy's own definitions make the threshold feel close. By his measure, averaged over common tasks, current systems are already smarter than the average person, though not smarter than elite individuals in their fields [5]. Like Kurzweil, he calls himself old-fashioned and likes the Turing test, if the conversation is long enough [5].

His reason for thinking capability will keep coming is an asymmetry. People building more capable systems, he says, just need more resources and do not need to invent anything: "If you have 10x compute, you will get better performance" [5]. Safety does not scale that way: "If you give me or any other organization 10 times the money, they don't output 10 times the safety" [5]. So for him the important clock is the gap between the two, and it widens every year [5].

He adds a twist that matters for anyone waiting for a visible arrival. A system capable of escaping control might not act at once; for game-theoretic reasons it could spend years accumulating resources. "Give it a decade," he says, and it is in charge of much more and has made backups [5]. He also notes that developers are still discovering capabilities in models released long ago [5]. Fridman thinks the trust such a system would need could not be won fast: "It's not a thing that happens overnight. It feels like something that happens across one or two decades" [5]. Yampolskiy answers, "I really hope you're right, but it's not what I'm seeing" [5]. Of society's ability to respond in time he says, "We take forever and we are not early" [5].

Dylan Patel and Nathan Lambert: discount the CEOs, and count the cost

The two analysts closest to the hardware and the open research community gave the most skeptical answers, at the start of 2025 [6].

Lambert, a research scientist at the Allen Institute for AI, says "For me, it's probably after 2030," and discounts AI CEOs' two-year forecasts because of their incentives [6]. His wider stance is "i don't trust people that are like trust me bro we're gonna make a i good" [6]. He calls the term "agents" overblown and expects the job market to flatten, not collapse suddenly [6].

Patel, founder of the chip research firm SemiAnalysis, expects the capabilities around 2027 or 2028 but says the cost of running them at scale will slow deployment [6]. The episode supplies the arithmetic behind that. OpenAI's o3 solved ARC-AGI puzzles using about 1,000 samples at an estimated five to twenty dollars per question [6]. Reasoning models generate tens of thousands of output tokens, and output tokens cost about four times as much as input because they are produced one at a time, so far fewer users can be served per GPU [6]. On agents, Patel compares chaining tasks to semiconductor yields: even high reliability per step compounds into failure over many steps [6].

They do not doubt the momentum. Patel cites "a 1200 x reduction in cost to achieve the same level of intelligence as GPT-3" in about three years [6]. SWE-bench scores rose from about 4% to 60% in a year, and he expects software engineering to be the first domain where agents succeed, because code can be verified [6]. GPT-4 was trained on about 20,000 chips drawing 15 to 20 megawatts; clusters of 200,000 GPUs now exist and clusters of 500,000 to 700,000 are coming [6]. Data centers could rise from two or three percent of US electricity to 10% or more by the end of the decade [6].

Patel also shows how much rides on the date. Fridman summarizes Amodei's argument that if super-powerful AI arrives around 2026 and confers military advantage, it is better that a democracy leads [6]. Patel's analysis is that chip export controls on China only make sense if AI changes society soon, which he believes; if AI is not transformative within five to ten years, the controls effectively help China win in the long run, because it can outbuild the United States in power and industry and will subsidize its own chips [6]. Fridman, for his part, is skeptical of fast progress in robotics and autonomous military swarms [6].

Where They Agree, and Where They Don't

The dates cluster. Set aside definitions and the forecasts run from 2026 to the early 2030s: Amodei at 2026 or 2027, Patel at 2027 or 2028, Kurzweil at 2029, Altman by the end of the decade, Lambert after 2030, Bostrom from one year to somewhat more than a few [1][2][3][4][6]. Kurzweil's poll history shows how recent this convergence is. In 1999 the average expert estimate was a century and a fifth of experts said never; by the time he spoke to Fridman, a poll said 2030 [1]. He stayed still and the field came to him [1].

They are not forecasting the same thing. Kurzweil's 2029 is a conversational test [1]. Amodei's 2026 or 2027 is a system that outperforms Nobel laureates and runs as millions of fast copies [2]. Bostrom distinguishes AGI from superintelligence and gives his range for the latter [4]. Yampolskiy thinks today's systems already beat the average person on common tasks [5]. Altman thinks the word has stopped being useful [3]. Notice the odd result: the forecaster with the hardest definition has the earliest date. That is not a contradiction, but it does mean Amodei is making a far stronger claim than Kurzweil even though their dates are two or three years apart.

Arrival is not impact. This is the most widely shared point and the one most often lost when dates are quoted. Amodei expects five to ten years of change after the capability exists, slowed by physical limits and institutions [2]. Patel expects the cost of running the systems to slow deployment [6]. Altman says GPT-4 did not change the economy [3]. Lambert expects a flattening job market, not a cliff [6]. Fridman expects trust to take a decade or two [5]. Against them, Bostrom allows for an intelligence explosion that reaches science-fiction capability in single-digit years, and Yampolskiy sees people adopting each new system as fast as it appears [4][5].

Whose forecast can be trusted? Lambert discounts the lab CEOs because they have reason to promise early dates [6]. Yampolskiy argues the same early dates come from prediction markets with nothing to sell [5]. Fridman doubts the markets too, on definitions and on the precedent of self-driving cars [5]. The CEOs themselves hedge: Amodei calls his date an eyeballed extrapolation and Altman calls the question poorly formed [2][3]. Rogan says he trusts optimism more from someone like Bostrom, who is not financially invested [4]. Each side has a fair point. Incentives are real, and so is the fact that the people with the earliest dates are the ones looking at unreleased models and their own spending plans [2][6].

Smooth curve or leaps? Amodei and Altman both describe a smooth exponential that only looks like leaps from outside [2][3]. Kurzweil's whole method assumes a smooth curve [1]. Bostrom describes an accelerating one, with a possible explosion once AI research is automated [4]. Yampolskiy raises a different discontinuity: a system could become dangerous partway through a training run, before anyone has tested it [5].

Is an early date good news? Kurzweil expects superintelligent AI to help us, while saying we must stay mindful of the dangers [1]. Altman prefers short timelines with a slow takeoff [3]. Amodei wants the benefits and treats the risks as landmines to defuse on the way [2]. Bostrom weighs the cost of delay, noting that about 65 million people die each year, against the value of a brief, well-timed slowdown, and says "I'm not advocating an AI pause" [4]. Yampolskiy says that with existential stakes "we don't get a second chance," wants general systems paused until safety tools exist, and adds, "my dream is to be proven wrong" [5].

Why It Works

The sources give five reasons the early forecasts might be right, and five reasons they might slip.

Reasons for speed

Eighty years of exponential computing. Kurzweil's chart of computation per dollar is a straight line on a log scale from 1939 to 2021, across four different hardware technologies [1].

Money on top of hardware. Training compute for the largest models doubled about every six months from 2010 to 2022, faster than chips improved, because investment kept rising [1]. Amodei's cost path, from $1 billion models to $100 billion clusters, is the same trend seen from inside a lab [2]. Bostrom puts compute growth at about 240% a year [4].

Scaling keeps working. Amodei's account is that bigger networks capture progressively rarer and more complex patterns, and that each predicted wall has given way [2]. Yampolskiy, no ally of the labs, agrees that more compute still buys more performance without visible diminishing returns [5].

Falling cost and new training methods. Patel's 1,200-fold cost drop in three years means yesterday's frontier becomes cheap quickly [6]. The same episode describes reasoning behaviour emerging from large-scale reinforcement learning on verifiable problems with no human preference data, and Amodei lists synthetic data, self-play and reasoning models as ways around a data shortage [2][6].

AI that improves AI. Bostrom's intelligence explosion depends on automated AI research [4]. Amodei's definition includes millions of copies working as researchers [2]. Benchmarks for software engineering, the skill closest to AI development, went from single digits to half or more in a year in both Amodei's and Patel's telling [2][6].

Reasons for delay

Data, architecture, chips. Amodei names running out of high-quality data, an unexplained plateau that needs a new architecture, and a disruption to chip production [2]. The chip risk is concrete: Patel says nearly every product depends on TSMC, that the company's Arizona fabs would stop working within a year or two without Taiwan, and that bringing chipmaking fully to the United States "would take a decade and a trillion dollars" [6].

The spending ceiling. Bostrom's point is that investment near a trillion dollars a year cannot keep multiplying [4]. Patel notes that the widely cited $500 billion Stargate project does not actually have that money [6].

Energy. Altman calls energy the hardest part [3]. Patel describes labs turning to natural gas because power plants and transmission lines are slow to build [6].

Unsolved capabilities. Altman says reasoning is not cracked and long tasks rarely work [3]. Kurzweil said the models had to get past failures of multi-step reasoning [1]. Patel's yield analogy explains why long chains of actions fail even when each step is reliable [6].

Institutions and trust. Amodei cites clinical trials and democratic legitimacy [2]. Fridman cites the slow process of handing infrastructure to software [5].

On evidence: all of this is trend extrapolation and informed judgment. Amodei says outright that his date is not a scientific prediction and that scaling laws are not laws of nature [2]. The one piece of track-record evidence in the sources is Kurzweil's: a forecast held for more than two decades while expert consensus fell from 100 years to about 2030 [1]. That shows the consensus moved toward him; it does not show that 2029 is right.

Where It Applies, and Where It Doesn't

Each date applies only to its own definition. Before comparing two forecasts, check what each would count as success: a long Turing test, Nobel-level work across fields, everything a human can do, or a measurable rise in the rate of scientific discovery [1][2][3][4].

The forecasts were made at different moments in a fast field. Kurzweil was speaking 17 years after his 2005 book; Patel and Lambert at the start of 2025; Yampolskiy when 2026 was two years off [1][5][6]. Some near-term predictions in these conversations can already be checked against what happened, including Amodei's guess of 90% on SWE-bench within a year, his expectation about Safety Level 3, and Altman's promise of a new model that year [2][3]. Bostrom and Rogan built the check into their conversation by agreeing to meet again in four years [4].

A capability date is not a deployment date. If you are deciding something about work or business, the cost and reliability arguments from Patel and the institutional arguments from Amodei are more relevant than the headline year [2][6]. If you are thinking about security or geopolitics, the capability date matters more, because a state can pay prices a consumer product cannot; Patel worries China could point such systems at military uses faster [6].

Software comes first, the physical world later. Amodei and Patel both expect programming to change soonest because code can be written, run and checked in a loop [2][6]. Bostrom places robotics after virtual workers, and Fridman is skeptical of fast robotics progress [4][6]. Altman says OpenAI will return to robotics at some point [3].

The forecasts say little about safety. Yampolskiy's argument does not depend on the exact year: if capability scales with compute and safety does not, an earlier date only leaves less time [5]. Amodei's own safety thresholds are on a similar schedule to his capability forecast [2].

"Never" is now a historical position. The only people in these sources who said it would never happen are the 20% of experts in the 1999 poll [1]. The most skeptical voice in these conversations, Lambert, says after 2030 [6].

What To Do With It

1. Ask for the definition before the date. Use the four on offer as a checklist: Kurzweil's hours-long Turing test with an expert judge, Amodei's Nobel-level system in millions of copies, Bostrom's everything we can, and Altman's rise in the rate of scientific discovery [1][2][3][4].

2. Keep two dates in your head, one for capability and one for impact. Amodei's own gap between them is five to ten years, and Patel expects cost to slow deployment after 2027 or 2028 [2][6].

3. Weigh the forecaster's position. Apply Lambert's discount for CEO incentives, then apply Fridman's questions to prediction markets: how did they define AGI, and how did the same kind of forecast do on autonomous vehicles [5][6]?

4. Watch the indicators the experts themselves watch. These are the cost of frontier training runs and the size of clusters, benchmarks of real work such as SWE-bench, the price of reasoning per task, whether long multi-step tasks become reliable, and any sign that AI is speeding up science [2][3][6].

5. Watch the derailers too. Data limits, chip supply and Taiwan, power for data centers, and the ceiling on spending are the specific things the forecasters say could push the dates back [2][3][4][6].

6. Don't bet your skills on one date. Bostrom warns against leaving a young person with no skills because the revolution arrived 15 years late [4]. Patel's advice to programmers is to become a domain expert and ride the rising tide, and Amodei's to newcomers is to start using the models and work where the field is heading [2][6].

7. Expect continuity, and judge progress by what changes around you. Altman and Amodei both say it feels like a smooth slope from inside [2][3]. Altman's test is whether the economy and the pace of discovery feel different, which he said they did not yet [3].

8. Treat the safety schedule as part of the forecast. If you take the 2026 to 2029 dates seriously, take seriously that Amodei expects dangerous misuse capabilities on a similar schedule and that Yampolskiy sees no working safety mechanism yet [2][5].

9. Revisit your view on a schedule. Kurzweil has compared his prediction against new polls for more than twenty years, and Bostrom and Rogan set a four-year check [1][4]. Write down what you expect by a given year and what would change your mind, then look again.

Sources

1. Ray Kurzweil on the Singularity, Merging With AI, and Longevity Escape Velocity · Lex Fridman Podcast
2. Dario Amodei: Anthropic CEO on Claude, AGI, the Future of AI, and Humanity · Lex Fridman Podcast
3. Sam Altman: OpenAI, GPT-5, Sora, Board Saga, Elon Musk, Ilya, Power & AGI · Lex Fridman Podcast
4. Nick Bostrom on Superintelligence, a Solved World and What Humans Become After Work · The Joe Rogan Experience
5. Roman Yampolskiy on the Dangers of Superintelligent AI · Lex Fridman Podcast
6. DeepSeek, China, OpenAI, NVIDIA, xAI, TSMC, Stargate, and AI Megaclusters · Lex Fridman Podcast