What “we know how to build AGI” is meant to mean

On January 6, 2025, Sam Altman posted a short essay called Reflections to his personal site. Most of it was a plain account of OpenAI’s recent year, the launch of o1, the growth of ChatGPT past 300 million weekly active users, and the November 2023 board episode in retrospect. One sentence, partway through, did most of the work the essay was published to do. “We are now confident we know how to build AGI as we have traditionally understood it.”

That sentence, sixteen words long, appeared two weeks before President Trump’s inauguration, fifteen days before the $500 billion Stargate announcement, and four months before OpenAI announced, on May 5, 2025, that its operating company would become a public benefit corporation. It was neither a research finding nor a product release but a declarative statement, made by the chief executive of the company most plausibly positioned to define what “AGI” means in the period when commercial usage of the term still has financial meaning.

The sentence does three things at once. The first is technical. The claim that the lab now knows how to build AGI is, in its most defensible reading, a claim that the research direction is settled. The architectural questions are settled enough that further progress is a matter of execution rather than discovery. Altman has been building toward this position for years. “In three words: deep learning worked,” he wrote in The Intelligence Age on September 23, 2024. Three Observations, posted February 9, 2025, formalized the underlying scaling claim: “the intelligence of an AI model roughly equals the log of the resources used to train and run it.” The reasoning models that began with o1-preview’s release on September 12, 2024, and continued through o3 and o4-mini in April 2025, were presented by OpenAI as evidence that reinforcement learning at the post-training stage scales the way pretraining does. More compute, applied at inference time as well as training time, produces more capability.

Altman, 41, has run OpenAI since 2019. He co-founded the lab in 2015 as a nonprofit, presided over the 2019 capped-profit restructuring, and oversaw the 2025 conversion to a public benefit corporation. Before OpenAI he ran Y Combinator. The technical claim he is making in Reflections rests on research conducted by people he hires; he is not himself a research scientist. What he brings to the claim is the institutional position from which it can be made. He is the CEO of the lab that shipped GPT-3, GPT-4, ChatGPT, and the o-series reasoning models, and that began a staged release of GPT-6 Astra on September 3, 2026.

The second thing the sentence is doing is definitional. “AGI as we have traditionally understood it” is the operative phrase. The traditional understandings of AGI that have been in circulation since the 1950s are not consistent with one another. The Turing-style framing asks whether a machine can be distinguished from a human in conversation. The economic framing written into OpenAI’s 2018 charter asks for “highly autonomous systems that outperform humans at most economically valuable work.” The Bostrom-style framing asks whether systems can perform autonomous research at the level needed to produce further AI capability without human input. The 2020s benchmark framing asks whether systems pass particular evaluations. Each of these traditions defines AGI differently. None has been met by an external standard. By embedding the phrase “as we have traditionally understood it,” Altman invokes a tradition that does not exist as a single thing, and lets the reader supply the most flattering of the available definitions.

The definitional move is not necessarily disingenuous. The phrase has plausibly been chosen to acknowledge that the term AGI is used loosely and that OpenAI has reached the level of capability that most working AI researchers would have called AGI a decade ago. The benchmarks the o-series models have cleared are real. In OpenAI’s September 2024 evaluation of its full o1 research model, a step beyond the o1-preview released that day, o1 reached the 89th percentile on Codeforces. Taking a consensus answer across 64 samples per problem, it scored 83 percent on AIME 2024, where GPT-4o scored 13 percent by the same method, and 78 percent on GPQA Diamond. Those are not contestable claims. The contestable claim is whether reaching those benchmarks satisfies the traditional understandings the sentence invokes.

The third thing the sentence is doing is institutional. The capital deployed against AI in the eighteen months following the Reflections post, including the $500 billion Stargate commitment of January 21, 2025 and the $40 billion OpenAI funding round closed later that year at a $300 billion valuation, is underwritten by the assumption that the path Altman describes is real. If AGI is buildable on the trajectory the lab has described, the returns to capital deployed at the frontier are commensurate with the size of the bet. If it is not, they are not. The sentence in Reflections states the assumption underneath the financing. It is not a marketing claim.

The disagreements come from two directions. The first is architectural skepticism, argued most prominently by Yann LeCun, who left Meta at the end of 2025 to found AMI Labs and has argued for years that large language models are a dead end as a path to general intelligence. Demis Hassabis of Google DeepMind has made a more measured version of the argument: at the Axios AI+ Summit in December 2025 he said scaling should be pushed to the maximum but that one or two more breakthroughs are probably still needed. The disagreement is not about whether scaling produces capability gains. It is about whether the gains converge on the kind of intelligence the term has historically denoted.

The second is empirical critique, argued most consistently by Gary Marcus, who appeared as a witness alongside Altman at a Senate Judiciary subcommittee hearing on May 16, 2023. Marcus’s position then and since is that current systems are brittle, that they fail in characteristic ways under modest distribution shift, and that the impressive benchmark performance reflects the saturation of measurable tasks rather than progress toward general intelligence.

Altman’s response to both critiques, in the body of his own writing, has been to treat the disagreement as a question of timeline rather than direction. He does not argue that LeCun is wrong about the eventual need for architectural innovation. He argues that the current approach has more headroom than the skeptics estimate. The Gentle Singularity, posted June 10, 2025, made the case that agents “doing real cognitive work” were already in the field.

The likeliest reading of the Reflections sentence is that it is functionally a bet about audience. The phrase “AGI as we have traditionally understood it” can be redefined to fit whatever OpenAI ships next. The first test came on September 3, 2026, when OpenAI began a staged release of GPT-6 Astra. Its president, Greg Brockman, suggested to Axios that the model might mark AGI’s arrival, closing with “Welcome to the AGI era.” Whether the educated public reads it that way is contested. If the next several model generations produce systems that the educated public reads as AGI, the sentence will be remembered as a clear-eyed forecast. If they do not, the sentence will be remembered as an overstatement that capital was raised against. The claim does not need to be technically correct in any narrow sense to perform its institutional function. It needs to be read as correct by enough of the people writing checks. So far it has been.

The question the sentence leaves open is whether the term still has external content, or whether the lab has achieved enough definitional authority that whatever it ships will be called AGI by enough of the relevant audience for the institutional purposes of the claim to be served. Reflections is the best evidence for the view that the second has already happened. The sentence does not need to be falsifiable to do its work. It needs to be repeated in financing memos, government press releases, and industry coverage.