Rishabh Singh has spent his career building systems that predict things. He wrote C++ trading systems for global investment banks, and he has filed patents on AI architectures including what he calls his Dynamic Optimization and Forecasting Model. When a journalist asked him when artificial general intelligence would arrive, he declined to give her a year.

He explained why in a September essay, "AGI Timelines and Why I Don't Predict Them," which argues that AGI timeline predictions, "as a genre, deserve far more skepticism than they usually get." Singh is co-founder and CTO of the social app JoinMeIn, founder of the AI marketing platform Nova Reach and of Nova Infinity, a portfolio of vertical AI software products, and the founder of Helix Mind, where he works on language-model architectures and training systems. His position is not a dodge. It is a planning rule, and the record of AGI forecasts gives it a good deal of support.

What counts as AGI?

That depends on who is asked, which is the first problem with putting a date on it. OpenAI's Charter defines artificial general intelligence as "highly autonomous systems that outperform humans at most economically valuable work." That is a threshold of labor replacement, measured against what people are paid to do.

Researchers at Google DeepMind proposed a ladder instead. Their paper "Levels of AGI", first published in November 2023, sorts systems by how general they are and how well they perform, from "Emerging" (equal to or somewhat better than an unskilled human) through "Competent" (at least the 50th percentile of skilled adults), "Expert," "Exceptional" and "Superhuman." The authors placed ChatGPT, Bard and Llama 2 at the first rung, Emerging AGI. They wrote that the Competent level, which no public system had reached when they wrote it, corresponds best to many earlier conceptions of AGI, while OpenAI's labor-replacement threshold sits higher up the ladder.

Some of the people building the systems avoid the term altogether. Dario Amodei, chief executive of Anthropic, wrote in his October 2024 essay "Machines of Loving Grace" that "I dislike the term AGI," calling it "an imprecise term that has gathered a lot of sci-fi baggage and hype." He prefers "powerful AI," which he summarises as a "country of geniuses in a datacenter."

When the finish line itself is disputed, a date for crossing it carries less information than it appears to.

How far have AGI forecasts moved?

By years at a time, and in both directions. In October 2023 the research group AI Impacts surveyed 2,778 researchers who had published in top AI venues. Their aggregate forecast gave a 10 per cent chance that unaided machines would outperform human workers at every task, better and more cheaply, by 2027, and even odds by 2047. The same survey's 2022 round had put the 50 per cent date at 2060, so the estimate moved 13 years in a single round.

The same researchers were far more cautious when the question was framed as jobs rather than tasks. They gave a 10 per cent chance of full automation of all occupations by 2037 and 50 per cent by 2116, a date that had itself moved 48 years from 2164 a year earlier. The authors note that the gap between the two framings, for what is arguably the same achievement, has been consistently more than sixty years. In the same survey, 57.8 per cent of respondents put at least a 5 per cent chance on extremely bad outcomes such as human extinction.

Forecasting platforms produce their own numbers. On Metaculus, where about 1,900 forecasters predict when the first general AI system will be devised, tested and publicly announced, the community median stood at 7 November 2030 on 23 October 2026, with the middle half of forecasts running from March 2028 to October 2036.

The most detailed public scenarios have moved too. The authors of AI 2027, released by the AI Futures Project in April 2025, wrote that December that developments pointed to a slower path toward autonomous AI research than they had first laid out. By April 2026 they had moved back toward their original pace, citing progress in AI programming tools. That is a revision later and then a revision sooner, about four months apart.

The habit is older than the current boom. In 1960, and again in a 1965 book, Herbert Simon, a founding figure of artificial intelligence and later a Nobel laureate in economics, wrote that machines would be capable within twenty years of doing any work a man can do. That deadline passed more than four decades ago.

What have the lab leaders said?

Dates that cover most of a decade. Sam Altman, OpenAI's chief executive, wrote in his January 2025 post "Reflections" that "we are now confident we know how to build AGI as we have traditionally understood it," and that "in 2025, we may see the first AI agents 'join the workforce.'" Amodei wrote that powerful AI "could come as early as 2026, though there are also ways it could take much longer."

Demis Hassabis, chief executive of Google DeepMind, has been more measured. Speaking at a briefing in DeepMind's London offices in March 2025, he said that "over the next five to 10 years, a lot of those capabilities will start coming to the fore," while adding, "We're not quite there yet," and "we've still got quite a lot of research work to go before that," CNBC reported.

None of this makes the forecasters careless. They are serious people answering a hard question. The spread in their answers, and how often the answers change, is the heart of Singh's argument.

Why won't Singh name a year?

Because, in his view, no one has the evidence to. The analogy he gave the journalist is a volcano. Ask a group of geologists in the early stages of studying one to predict the exact date of its next major eruption, he wrote, and some might offer a confident-sounding date based on current patterns. A more careful geologist would say the system is complex enough, with enough interacting variables, that anyone offering a specific date "is expressing confidence the actual evidence doesn't support, however sincerely they believe their own estimate."

AGI forecasts, he argues, work the same way. People who say it is a few years off, and people who say it is decades away and overhyped, are "almost always extrapolating from the specific trend they happen to find most salient": recent capability jumps for the first group, recent failures for the second. He adds that he has spent enough time close to actual research to know how often confident trend lines from a year or two earlier turned out to be wrong, "in both directions."

What do the benchmarks show?

Fast progress and real limits, often in the same results, which is close to the picture Singh describes. The ARC Prize Foundation's tests are built to measure reasoning on problems a model has not seen. On its leaderboard, OpenAI's o3 model scored 6.5 per cent on the ARC-AGI-2 test in April 2025. GPT-6 Astra, released in September 2026, scored 95.0 per cent. ARC's human panel scored 100 per cent.

The foundation's newer test, ARC-AGI-3, asks AI agents to adapt on the fly to interactive environments they have never seen. There the same model scored 62.7 per cent with ARC's standard harness and 98.6 per cent with ARC's "Provider Adapter" harness, which preserves the model's reasoning state so it can reuse prior work. The model was identical. The scaffolding around it made the difference.

The resources behind that progress are growing steeply. Epoch AI, a research group that tracks the field, estimates that the computing power used to train frontier language models has grown about five-fold a year since 2020, and that the cost of training them has grown about 3.5-fold a year.

What does a timeline do to the way companies build?

It changes their priorities, and Singh thinks often for the worse. The sums are large. The four biggest U.S. technology companies' plans pointed to about $650 billion of capital spending in 2026, much of it on AI data centers, Bloomberg tallied in February. Meta alone spent $34.9 billion on capital projects in its third quarter, this publication reported this week. Bets that size rest on assumptions about how capable AI will become, and when.

Singh's concern is what those assumptions do further down, inside the companies building products. A founder who believes AGI is imminent, he writes, "might rush safety and evaluation work, assuming there's no time to be careful." A founder who believes it is decades away might skip careful design work, assuming today's more limited systems don't warrant the same rigor. "Both mistakes come from over-trusting a specific prediction that neither person actually had solid grounds to make with real confidence in the first place."

For anyone shipping software, that reframes the question from when to what. Singh would rather ask what a team should be building and testing for, regardless of when more general capability arrives. His answer is systems that are "honest about their current limitations and designed to fail safely regardless of how capable the underlying models eventually become." A founder who spends real energy trying to time the future is, in his experience, usually distracted from that work.

It is a view shaped by where he has worked. Singh writes on his site that years inside bank trading systems, where success was measured in nanoseconds, taught him that "whoever builds the better infrastructure wins." At Nova Infinity he architected one shared AI, identity and billing layer so that each new product launches on the same foundation.

Where does Singh think AI stands now?

Somewhere between the two camps, and he is open about the uncertainty. Pressed for a personal view, he writes that his position is closer to "I don't know, and I'm suspicious of anyone who claims to know with real confidence." He credits the field with genuine, meaningful progress on specific narrow capabilities, and believes it is still meaningfully far from the broad, independent, adaptive reasoning that would deserve the AGI label without significant caveats.

Where that threshold sits, and how quickly the remaining distance closes, he calls one of the hardest questions in the field, made harder by the many confident and contradictory answers offered in public.

What he gave the journalist instead of a date was a framework: build as if today's systems are genuinely powerful and genuinely limited at the same time, "because that's what the actual evidence, carefully examined, currently supports." She did not use it as her headline. A founder declining to predict the future makes a weak pull quote, he concedes. But she told him afterward that it was the first answer in a long day of interviews that had not tried to sound more certain than the evidence allowed.

What does the other side say?

That some number is better than none. Forecasters argue that a range with stated uncertainty is more useful than silence, and that investors, policymakers and hiring managers must plan around some expectation either way. Metaculus publishes exactly that kind of range, and the AI 2027 authors describe their scenarios as planning tools rather than literal forecasts. On that view, declining to give a timeline only hides the assumption a person is already making.

Singh's essay offers a reply. The assumption he wants teams to plan around is a condition rather than a date: today's tools can do real things and can fail in real ways, and both deserve engineering attention now. In a field that rewards the boldest number, it is a quieter position, and for the people building products this year, it may be the more useful one.

Sources

Rishabh Singh is a client of Media Morpheus, the PR and podcast booking firm owned by the publisher of The Cranberry Journal. Singh's views, the journalist anecdote and all quotations attributed to him come from his essay “AGI Timelines and Why I Don’t Predict Them,” published on rishabhsingh.novareach.digital on 11 September 2026; his roles, companies, patent filings and background come from the same site as read on 23 October 2026. The survey figures come from “Thousands of AI Authors on the Future of AI” (Grace et al., AI Impacts, January 2024). OpenAI's definition is from its Charter; the levels framework is from “Levels of AGI for Operationalizing Progress on the Path to AGI” (Morris et al., Google DeepMind, arXiv 2311.02462, revised September 2025). The Altman quotation is from his blog post “Reflections” of January 2025; the Amodei quotations are from his essay “Machines of Loving Grace” of October 2024; the Hassabis quotations are from CNBC's report of 17 March 2025 as republished by NBC News. The Metaculus median was read on its question page on 23 October 2026. ARC-AGI scores are from ARC Prize's leaderboard as read on 23 October 2026. Compute and cost trends are from Epoch AI's trends page, updated 5 February 2026. The $650 billion capital-spending tally is Bloomberg's, as republished by Data Center Knowledge on 6 February 2026; Meta's quarterly figure is from this publication's report on its 21 October earnings. The AI 2027 revisions are as summarised in the Wikipedia article on the AI Futures Project. Herbert Simon's prediction appears in The New Science of Management Decision (1960) and was reprinted in The Shape of Automation for Men and Management (1965), as traced by Quote Investigator. The analysis is our own.

Topics technologyaiartificial intelligenceforecastingstartups

Senior Writer

Cory Chamberlain

Cory Chamberlain covers corporate strategy, private markets and the economics of reputation, along with the state-capacity questions that sit underneath them.

All stories