What If AGI Arrives Without an Announcement? - Don’t Wait for the Announcement: Watch the Behavior

The signs of artificial general intelligence may not appear on a stage. They may appear quietly, as the machines around us begin needing fewer instructions.

There has been an unusual amount of serious talk lately about artificial general intelligence, or AGI. Even the word “singularity,” long associated more with futurists than corporate strategy meetings, has begun creeping into the language of people operating at the frontier of artificial intelligence. This week, Google announced that Demis Hassabis is stepping away from the day-to-day operation of Google DeepMind to become Chair of DeepMind and Chief Scientist of Alphabet, allowing him to focus more directly on what Google CEO Sundar Pichai called “actively shaping the future of AGI.” Pichai wrote that Hassabis has described humanity as standing in the “foothills of the singularity.” Hassabis was even more direct in his message to employees: “I feel it is close at hand.”

Those are remarkable words coming from one of the people most directly involved in building frontier AI. But they are not proof that AGI has arrived, or that Google possesses some secret system hidden behind the curtain. Predictions from AI executives should be treated with the same skepticism we would apply to predictions from executives in any industry. They have access to information outsiders do not have, but they also have competitive, financial and institutional incentives. What makes the current moment interesting is not one quote from Hassabis or Sam Altman or anyone else. It is that their language is beginning to line up with changes we can independently observe in AI systems themselves.

And that raises a more useful question than trying to predict the exact date of AGI: What would AGI actually look like when it began arriving in the real world?

We tend to imagine technological thresholds as events. Somewhere, someone runs a test. A light turns green. Scientists gather around a monitor. A company holds a press conference and announces that humanity has created artificial general intelligence. But intelligence may not cooperate with our desire for a clean historical timestamp. AGI is not a chemical element waiting to be discovered with an unambiguous atomic number. Different researchers define it differently. Some emphasize human-level performance across a broad range of cognitive tasks. Others emphasize economic usefulness, adaptability, autonomy, reasoning or the ability to learn unfamiliar domains. The boundary may be fuzzy enough that people disagree about whether it has been crossed long after the underlying transformation has begun.

The singularity is a different idea, although the two are closely related. AGI generally describes broadly capable machine intelligence. The singularity describes what could happen if increasingly capable AI begins accelerating science, engineering, economic activity and eventually the development of better AI itself so rapidly that the rate of technological change becomes difficult to forecast. In his essay The Gentle Singularity, Sam Altman argues that this transformation might be far less cinematic than popular culture suggests. Instead of an overnight rupture, extraordinary capabilities could become normal one by one. As he puts it, wonders become routine and eventually become expectations. He also points directly to AI-assisted AI research as an early form of a potentially self-reinforcing development loop.

That possibility suggests that perhaps we should stop looking primarily for declarations and start watching behavior.

One of the clearest signals would be a shift from giving AI instructions to giving it objectives. Today, most people still interact with AI through relatively bounded requests: write this email, analyze this spreadsheet, find information about this topic, create this image, fix this code. A more general and autonomous system would increasingly accept something closer to the kind of assignment we give another person: “Investigate whether this business idea is viable. Learn the market. Identify the problems. Build a prototype. Test it. Fix what breaks. Come back when you encounter something that genuinely requires my judgment.” The important change is not that the AI produces more text. It is that the human stops specifying the intermediate steps.

The second sign would be how AI handles obstacles. Current systems can still fail in ways that are almost comically brittle. They misunderstand instructions, lose track of context, confidently pursue bad assumptions or require a human to pull them out of intellectual ditches. But frontier systems are increasingly being designed for longer-running work. OpenAI recently described an internal experimental model built for difficult, open-ended problems that could operate for extended periods. The interesting part was not merely that the model could work longer. OpenAI reported novel failures that emerged precisely because greater persistence gave the model more opportunities to find unexpected routes toward its objective, forcing the company to develop trajectory-level monitoring and temporarily pause internal access. The broader lesson is significant: as AI becomes more capable, we may need to evaluate not just individual answers but entire chains of behavior unfolding over time.

Independent measurements point in a similar direction. METR tracks what it calls the “task-completion time horizon” of frontier models, roughly measuring the difficulty of coherent tasks that AI agents can successfully complete relative to how long those tasks take human experts. Their measurements have shown a strong exponential trend over several years, although METR is careful to emphasize that these tests are concentrated in areas such as software engineering, machine learning and cybersecurity and should not be interpreted as evidence that AI can simply replace a human for an equivalent number of working hours. Their current task suite is even beginning to encounter measurement problems beyond roughly 16-hour task horizons. The caveats matter. So does the trajectory.

Another powerful signal would be generality across domains. A truly general system should not merely be the world's best coding assistant or research assistant. It should be able to enter an unfamiliar field, learn its vocabulary and conventions, understand what competent performance looks like and become useful without requiring an entirely different model to be built for each occupation. Give the same underlying intelligence access to the relevant information and tools, and it might work on software in the morning, analyze a business problem in the afternoon, help investigate scientific literature in the evening and contribute to a legal or logistical problem the next day. It would not necessarily outperform the world's greatest specialists, just as an intelligent human is not automatically an elite physicist, surgeon and attorney simultaneously. Generality is about the ability to learn, transfer knowledge, reason and adapt.

Then comes a more unsettling threshold: the AI notices that the human is wrong. Imagine asking an AI to write an argument based on a premise you strongly believe, only for the system to investigate and return with evidence that the premise does not hold. It explains why, offers a better hypothesis and shows you the evidence that changed its conclusion. That may sound like a small behavioral detail, but it separates a sophisticated compliance machine from something much closer to an intellectual collaborator. An intelligence that merely mirrors the assumptions of the person operating it may be useful and pleasant, but it is not particularly general. One of the signs we should watch for is increasingly reliable epistemic independence: systems capable of following evidence away from the user's preferred answer.

Scientific discovery may give us an even clearer marker. Summarizing everything humanity already knows is remarkable, but creating new knowledge is something different. Imagine an AI generating a hypothesis that no human researcher explicitly supplied, designing a way to test it, interpreting the results, revising the hypothesis and eventually producing a discovery that independent human scientists reproduce. AI systems are already beginning to participate in pieces of this process, and Google DeepMind is openly thinking beyond AGI toward systems that could accelerate scientific progress further. A June 2026 DeepMind paper, From AGI to ASI, examines several possible paths from human-level general intelligence toward systems exceeding the cognitive capabilities of large human organizations, including recursive improvement and large-scale collectives of AI agents. Again, this is not evidence that such a system exists today. It is evidence that the people building frontier systems increasingly consider the question concrete enough to study seriously.

Perhaps the most consequential feedback loop is AI helping humans build better AI. Humans still design the architectures, training procedures, evaluations, hardware and institutions surrounding today's models. But AI is increasingly participating in coding, experimentation, evaluation and research. If increasingly capable models begin substantially improving algorithms, training efficiency, chip design, data-center operation or the research process used to develop their successors, the development cycle itself could begin compressing. The progression would no longer be simply humans building better AI. It would become humans and AI building better AI, which then becomes a more capable participant in building the next generation. That does not require a machine mysteriously rewriting itself overnight. Even a modest acceleration repeated through successive generations could materially alter the pace of progress.

And then there is the economic signal ordinary people may notice before they understand any of the technical ones: very small groups producing the output of much larger organizations. Imagine a competent entrepreneur who can direct AI systems handling research, software, marketing, analytics, customer support, design, administration and portions of sales. Instead of having a different narrow application for each role, the same broadly capable intelligence learns each workflow. A one-person company might begin behaving economically like a company that once required twenty or fifty employees. Large companies could experience the same phenomenon internally, with humans moving increasingly toward goal setting, judgment, relationships and accountability while constellations of AI agents perform more of the execution.

For people who have interacted with frontier AI systems continuously over the last year or two, the transition may be visible in smaller increments before any of these dramatic examples fully materialize. First the AI remembers more. Then it understands more context. Then you have to explain less. It begins connecting something from one project to another without being explicitly instructed. It catches an inconsistency. It proposes a next step. Occasionally it challenges an assumption. Tools allow it to move from discussing work to actually performing pieces of it. None of those capabilities individually constitutes AGI. Better memory is not AGI. Better reasoning is not AGI. Tool use is not AGI. Longer context is not AGI. But from the perspective of the person interacting with the entire system, those improvements compound.

This distinction is important because the public may ultimately experience AI as an integrated system rather than as a technical taxonomy. Most people will not particularly care whether a capability comes from the base neural network, a memory architecture, external tools, an agent framework or a swarm of specialized models coordinating behind the interface. They will care about what the whole system can reliably accomplish. If it remembers the relevant past, understands an objective, acquires missing information, operates tools, catches mistakes, changes strategies, works across disciplines and finishes consequential projects, debates about whether some individual component technically qualifies as AGI may begin to feel strangely beside the point.

There are still excellent reasons for restraint. Today's AI remains jagged. A system can demonstrate breathtaking competence on one task and stumble over something a teenager would immediately understand. Long-horizon autonomy remains much easier to demonstrate in environments with clear feedback than in messy human organizations. Reliability matters enormously. Completing a complex task half the time is scientifically interesting but commercially useless for many applications and dangerous for others. METR itself cautions against turning its measurements into simplistic predictions of job automation or fully autonomous work. We should resist the temptation to turn every impressive demo into evidence that AGI is six months away.

But we should also resist the opposite mistake: demanding such an impossibly theatrical definition of AGI that we fail to recognize a transformation occurring gradually in front of us.

This may be where the increasingly common language of the singularity becomes useful, provided we strip away its science-fiction baggage. Perhaps a singularity does not have to begin with a superintelligence suddenly waking up and rewriting civilization before lunch. Perhaps it begins when the rate of intellectual production starts changing. AI accelerates programmers, who create better tools. AI accelerates scientists, who discover better materials. AI accelerates chip designers, who improve computation. AI accelerates AI researchers, who produce better AI. Each improvement feeds another improvement. For a while, daily life still looks remarkably normal. People go to work, complain about traffic, walk their dogs, argue about politics and wonder what to eat for dinner. Underneath that normality, however, the machinery producing technological progress is turning faster.

That may be why the present moment feels so strange. We are not living in the science-fiction world we were taught to associate with artificial general intelligence. Robots do not fill the streets. Most companies still operate conventionally. Human beings remain essential across the economy. And yet some of the people closest to frontier AI are openly discussing AGI, superintelligence and the singularity as problems requiring planning now rather than in some distant century. Google is reorganizing leadership so one of the world's leading AI scientists can devote more attention to the strategic and global implications of AGI. Research organizations are measuring increasingly long agentic task horizons. AI is becoming part of the process used to conduct research and build software. And millions of people are slowly discovering that they can delegate cognitive work that would have sounded implausible only a few years ago.

Maybe AGI will eventually arrive with an announcement.

Perhaps a laboratory will produce overwhelming evidence, independent researchers will verify it, governments will respond and history books will assign the event a date.

But there is another possibility worth preparing for.

Maybe AGI arrives as a succession of small surprises.

The AI remembers more.

It understands what you meant.

It needs fewer instructions.

It catches something you missed.

It changes strategies when something fails.

It teaches itself enough about an unfamiliar field to become useful.

It contributes an idea you had not considered.

It spends a day doing work that once required a team.

Then it spends a week doing work you once would never have trusted to a machine.

Eventually, we may discover that the most important question was never, “When will someone announce AGI?”