Welcome to the AGI Era? Not So Fast!

OpenAI has launched GPT-6 Astra, and company president Greg Brockman has supplied the headline that the artificial-intelligence industry has been waiting years to print: "Welcome to the AGI era." Asked whether artificial general intelligence had finally arrived, Brockman went further, saying that personally he thought "we're there." Perhaps, he suggested, people looking back several years from now will identify this period, or even this particular model, as the point at which AGI arrived.

For those of us who have remained sceptical about the extravagant claims surrounding artificial intelligence, this is precisely the moment not to be carried away by the rhetoric. A company that has spent enormous sums developing a new product has announced that its new product represents a historic transformation of intelligence itself. That is interesting testimony, but it is hardly independent confirmation. Coca-Cola declaring that it has perfected refreshment does not settle the philosophy of thirst.

There is no universally accepted scientific test for AGI, which makes the term extraordinarily useful for marketing. Artificial general intelligence once suggested something comparatively clear: an artificial mind possessing the flexible, domain-general intelligence of a human being, capable of encountering genuinely unfamiliar situations, understanding them, learning what was necessary and acting competently without having its environment carefully constructed around it. As AI systems became increasingly capable, however, the boundary became harder to locate. AGI has gradually changed from an identifiable destination into something resembling a judgement call.

GPT-6 Astra is nevertheless an impressive system. Its reported results on reasoning, mathematics, coding, computer use and agentic tasks represent substantial advances. The spectacular headline is its performance on ARC-AGI-3, a benchmark intended to test adaptation to unfamiliar environments rather than memorisation of familiar examination material. OpenAI reports performance approaching saturation under its evaluation setup.

That deserves attention. It does not deserve metaphysics. A benchmark is an artificial environment with rules, scoring procedures, interfaces and success conditions. Passing increasingly difficult benchmarks demonstrates increasingly impressive capabilities, but there is no logical bridge from "this machine performed extremely well on our test of generalisation" to "therefore we have manufactured a general intelligence." The history of AI is littered with benchmarks that supposedly represented profound aspects of intelligence until machines became good at them, after which we discovered that intelligence was considerably larger than the benchmark.

Chess was once treated almost mystically. Surely a machine capable of defeating a world chess champion would have to think. Then Deep Blue defeated Garry Kasparov and civilisation continued. Computers became superhuman at Go. They became astonishingly good at recognising images, translating languages and generating prose. Large language models passed examinations that would defeat most humans. Each achievement was real. What repeatedly failed was the inference that success at the latest supposedly definitive task demonstrated that the remaining mysteries of intelligence had been solved.

Astra therefore confronts us with the old problem in a more impressive form. The benchmark numbers themselves also require caution because modern AI performance increasingly belongs not simply to a model but to an engineered system surrounding it. Memory, tools, context management, repeated attempts, computer access and the particular evaluation harness can substantially affect results. This is not cheating; humans also rely upon tools and external memory. But it complicates breathless claims that some percentage score demonstrates that a machine has crossed an ontological boundary from specialised computation into general intelligence.

Strip away the terminology and ask the elementary question: what has actually been demonstrated? Astra appears to be an exceptionally capable artificial system that can perform a widening range of intellectual and computer-based tasks. It can reason successfully across many constructed problems, write and analyse code, operate computer interfaces and carry out multistep activities with substantially greater autonomy than earlier systems. It can sometimes outperform human beings spectacularly.

None of this demonstrates consciousness. It does not demonstrate subjective experience, understanding in the human phenomenological sense, selfhood or free will. More importantly for the narrower AGI claim, it does not establish uniformly reliable competence across the open-ended mess of ordinary reality. A system can be superhuman in several dimensions while remaining surprisingly brittle in others.

This matters because "intelligence" is being quietly redefined around whatever machines happen to do well. If intelligence means high performance on mathematics, coding, formal reasoning and computer manipulation, modern AI looks increasingly super-intelligent. If intelligence also involves embodied knowledge, common sense, long-term autonomous judgement, social understanding, knowing when one's own reasoning has gone wrong, coping reliably with radically unfamiliar circumstances and maintaining coherent purposes in an uncontrolled world, the picture becomes considerably murkier.

The danger is that the industry defines the examination and then announces that its machine has passed. There is also a commercial incentive for the AGI boundary to move towards us. The phrase "we have produced a significantly better software model" is not worth trillions of dollars in imagined future economic transformation. "We have entered the AGI era" might be. Investors, governments and journalists understand the second sentence instantly. It implies that history has divided into before and after. The incentives therefore point overwhelmingly towards declaring victory early.

This does not mean Astra is harmless. Indeed, the genuinely disturbing parts of the announcement have remarkably little to do with whether it deserves the title AGI. OpenAI itself assesses Astra as having reached its highest, "Critical", cybersecurity capability threshold. That deserves far more attention than arguments over whether the machine possesses something sufficiently similar to human general intelligence to justify three fashionable letters. A system capable of sophisticated computer operation, coding and increasingly autonomous cyber activity can be strategically important without possessing consciousness, common sense or anything philosophers would recognise as a mind. A cruise missile does not need consciousness to be dangerous.Neither does an AI system.

This is where some critics of AI hype make the opposite mistake. They correctly observe that a language model is not a human mind and then conclude that there is therefore little to worry about. But those propositions do not follow. Industrial machinery transformed employment without becoming conscious. Nuclear weapons transformed international relations without understanding geopolitics. Algorithmic trading transformed financial markets without knowing what money is. An artificial system need not "understand" cyberwar in the human sense to become an extremely powerful cyberwar instrument.

That is also where the more sensational commentators go too far. The launch of Astra has already attracted claims that humanity now possesses an almost supernatural strategic weapon, that governments must secretly possess vastly more powerful unrestricted versions and that the balance of power in a future world war has suddenly been transformed. These claims run well ahead of the evidence.

There almost certainly will be intense military and intelligence interest in systems with these capabilities. Governments have every incentive to adapt frontier AI to intelligence analysis, cybersecurity, autonomous systems, logistics and other military applications. It is also reasonable to suppose that classified applications will not necessarily resemble the consumer versions available to the public. But "reasonable to suppose" is not evidence that a secret superintelligence is sitting beneath the Pentagon.

We should therefore reject two competing fantasies. The first is Silicon Valley's temptation to transform every major benchmark improvement into evidence that machine intelligence has crossed some historic threshold. The second is the apocalyptic temptation to transform every such announcement into Skynet.

There is a more interesting possibility between them. Perhaps AGI will never arrive as the single artificial mind imagined in science fiction. Perhaps there will be no morning on which humanity wakes to discover that somebody has switched on an electronic Einstein capable of doing everything a person can do, only better. What may emerge instead is an expanding technological assemblage: model, memory, computer access, specialised tools, autonomous agents, enormous processing power and networks of other models.

No component needs to be generally intelligent in the strong philosophical sense. The system as a whole merely needs to be useful enough to replace human functions. That possibility should concern AI sceptics considerably more than Brockman's slogan. The economic consequences of artificial intelligence do not depend upon philosophers agreeing that AGI exists. If an imperfect and unconscious machine can perform 70 or 80 per cent of a particular office worker's tasks at negligible marginal cost, arguing that it isn't "really intelligent" will provide little comfort to the displaced employee.

Likewise, a military does not care whether its cyber system experiences the thrill of discovering a vulnerability. It cares whether the vulnerability is discovered.

This produces the great irony of the AGI debate. The question everybody asks may turn out to be the wrong one. "Is GPT-6 Astra really AGI?" invites endless disputes over definitions of intelligence, consciousness, understanding and autonomy. There may never be universal agreement because these concepts were philosophically disputed long before computers entered the argument. The practical question is much colder: what can the machine actually do?

On that question Astra deserves serious attention. Its capabilities appear to have advanced considerably, particularly in autonomous computer work and cybersecurity. But impressive capability is not proof that OpenAI has solved general intelligence, still less consciousness or the nature of mind.

We should therefore take Brockman's declaration for what it is: the judgement of a man intimately involved with the company that built the technology, delivered during the launch of that technology. Perhaps history will vindicate him. Perhaps future generations really will point to September 2026 as the beginning of the AGI era. Or perhaps they will look back on the phrase much as we now look upon earlier predictions that thinking machines were only a few years away.

For the moment, "Welcome to the AGI era" belongs more comfortably on a Silicon Valley billboard than in a scientific textbook. GPT-6 Astra may be extraordinary software. It may become economically disruptive, militarily important and socially destructive in ways that should concern even those who reject the grandest claims made for artificial intelligence. None of those possibilities requires us to pretend that a benchmark score has answered one of the oldest questions in philosophy: what is a mind?

The sceptical position is therefore not that nothing happened. Something important plainly did. It is simply that extraordinary software remains a long way from proving an extraordinary metaphysical claim, over marketing propaganda.

https://www.youtube.com/watch?v=4H8YW3Wf3ls