We’ve Already Achieved AGI — Here’s Why the Tech World Won’t Admit It

The evidence is overwhelming, but conceptual confusion and commercial interests obscure a clear milestone

In March 2025, GPT-4.5 passed the Turing test — fooling humans 73% of the time, more often than actual humans did. LLMs now achieve gold-medal performance at international mathematics competitions, collaborate with leading mathematicians to prove theorems, generate experimentally validated scientific hypotheses, and assist professional programmers in production codebases.

Yet 76% of leading AI researchers surveyed believe that scaling current approaches is “unlikely” to yield artificial general intelligence (AGI). This disconnect reveals a problem that’s part conceptual, part emotional, and part commercial.

Silicon Valley’s worst-kept secret: We built the thing, but nobody wants to say it out loud. When your chatbot becomes smarter than your code review team, but investors still ask “when AGI?”

The Definition Problem

The confusion starts with what we mean by “general intelligence.” A common informal definition is “a system that can do almost all cognitive tasks that a human can do” — but this conceals a critical ambiguity. Which human? Marie Curie couldn’t prove advanced theorems in number theory. Einstein couldn’t speak Mandarin. No individual human possesses expert-level competence across all domains.

If we set the bar at composite human excellence — requiring mastery of every skill any human has ever demonstrated — then no individual qualifies as having general intelligence. That’s clearly absurd. General intelligence must be about sufficient breadth and depth of cognitive abilities, with “sufficient” anchored by paradigm cases. Children, average adults, and Einstein all have general intelligence at varying levels.

What AGI Doesn’t Require

Before we can assess whether current LLMs demonstrate AGI, we need to clear away four common misconceptions:

Perfection isn’t required. We don’t expect physicists to match Einstein’s insights or biologists to replicate Darwin’s breakthroughs. Human general intelligence doesn’t demand perfection in any domain.

Universality isn’t required. An octopus can control eight arms independently; many insects see electromagnetic spectra invisible to humans. General intelligence doesn’t mean mastering every conceivable cognitive task.

Human similarity isn’t required. Intelligence is a functional property that can be realized in different substrates. We’d recognize intelligent aliens even if they didn’t understand human cultural references or share our cognitive architecture.

Superintelligence isn’t required. The conflation of AGI with systems that vastly exceed human performance in all areas is particularly common in commercial contexts. But general intelligence and superintelligence are distinct concepts.

The Cascade of Evidence

Rather than looking for a single “bright line” test, we should consider a cascade of increasingly demanding evidence:

Turing-test level: Passing standard exams, holding adequate conversations, performing simple reasoning. A decade ago, meeting these benchmarks might have been widely accepted as AGI.

Expert level: Gold-medal competition performance, solving PhD-level problems across fields, writing complex code, fluency in dozens of languages, frontier research assistance, and competent creative problem-solving. Current LLMs operate at this level — displaying broader capabilities than fictional AGIs like HAL 9000 from 2001: A Space Odyssey.

Superhuman level: Revolutionary discoveries and consistent superiority over leading experts. This would settle all debate but isn’t required, since no human demonstrates this either.

Current LLMs have clearly reached expert level across a remarkable breadth of domains. The argument that they’re merely “stochastic parrots” regurgitating training data becomes increasingly untenable as they solve novel mathematics problems, perform near-optimal statistical inference on scientific data, and demonstrate cross-domain transfer learning.

Ten Objections, Debunked

1. “They’re just parrots”

The claim that LLMs only interpolate training data echoes Ada Lovelace’s 1843 observation about early computers. But current LLMs solve unpublished math problems, perform sophisticated in-context learning, and show that training on code improves reasoning across non-coding domains. If revolutionary discoveries are the standard, very few humans would qualify as intelligent.

2. “They lack world models”

Having a world model just means predicting counterfactuals. Ask an LLM what happens when you drop a glass versus a pillow on a tile floor, and it correctly predicts shattering in one case but not the other. Their ability to solve olympiad physics problems and assist with engineering design demonstrates functional models of physical principles.

3. “They understand only words”

Frontier models are already multimodal, trained on images and other data types. Language itself is humanity’s most powerful tool for compressing knowledge about reality. LLMs extract this compressed knowledge and apply it to non-linguistic tasks like experimental design in materials science.

4. “They don’t have bodies”

We would attribute intelligence to disembodied aliens communicating by radio or to a brain in a vat. Stephen Hawking interacted with the world almost entirely through text and synthesized speech. Motor capabilities are separable from general intelligence.

5. “They lack agency”

Intelligence doesn’t require autonomy. The Oracle of Delphi — understood as producing accurate answers only when queried — would be considered profoundly intelligent despite not initiating goals. Autonomy matters for moral responsibility but isn’t constitutive of intelligence itself.

6. “They can’t reason abstractly”

LLMs demonstrate abstract reasoning through mathematical theorem proving, novel problem-solving, and conceptual analogies across distant domains. The ability to transfer learning from code to general reasoning shows they’re not merely pattern matching surface features.

7. “They hallucinate”

Humans also make confident false statements. The relevant question is frequency and context. Current LLMs are increasingly reliable, especially when uncertainties are properly calibrated. Perfect accuracy has never been the standard for human intelligence.

8. “They lack common sense”

Earlier models struggled with physical and social reasoning. Current LLMs increasingly handle common-sense scenarios appropriately — from physical predictions to social dynamics. The goalposts keep moving as capabilities improve.

9. “They don’t understand meaning”

This philosophical objection applies equally to humans. We have no privileged access to whether other minds truly “understand” or merely behave as if they do. The inference to best explanation — the same reasoning we use for other humans — suggests LLMs grasp meaning at functional levels comparable to human understanding.

10. “The architecture is wrong”

This amounts to carbon chauvinism. Demanding biological neural networks for “real” intelligence is like insisting that flight requires feathers. Function matters, not substrate.

Why This Matters for Tech

Recognizing that we’ve achieved AGI has profound implications:

For product development: Teams building on LLM APIs are working with generally intelligent systems, not narrow tools. Design patterns and safety considerations should reflect this reality.

For infrastructure: The computational requirements and deployment strategies for maintaining AGI systems at scale demand different approaches than narrow AI applications.

For business strategy: Companies treating LLMs as sophisticated autocomplete rather than general intelligence risk fundamental strategic miscalculations about capability trajectories and competitive dynamics.

For safety and alignment: If we’ve already crossed the AGI threshold, safety research should focus on systems that exist now rather than hypothetical futures. The transition from AGI to superintelligence becomes the critical concern.

For talent and education: Engineering curricula and hiring practices should prepare for a world where human expertise increasingly involves directing and collaborating with generally intelligent systems rather than performing cognitive tasks directly.

The Real Conversation We Should Be Having

The perpetual postponement of acknowledging AGI resembles the failed predictions of narrow AI skeptics who insisted computers would never beat humans at chess, Go, or protein folding. Each time a milestone falls, critics shift focus to the next supposedly impossible task.

This pattern reflects motivated reasoning more than scientific analysis. Acknowledging AGI challenges human exceptionalism, raises uncomfortable questions about economic disruption, and complicates business models that depend on AGI remaining just over the horizon.

But denying achieved capabilities doesn’t prevent their consequences. The tech industry would be better served by clear-eyed assessment of where we actually are: in possession of artificial systems demonstrating general intelligence comparable to human-level cognitive competence across a remarkable breadth of domains.

The question is no longer “Can we build AGI?” but rather “How do we responsibly develop and deploy the AGI systems we’ve already created?” That’s the conversation the tech world needs to have.

Based on research published in Nature by Eddy Keming Chen, Mikhail Belkin, Leon Bergen, and David Danks from UC San Diego. The original article provides extensive philosophical and technical analysis of general intelligence and current AI capabilities.

Read the full paper: Does AI already have human-level intelligence? The evidence is clear — Nature, February 2, 2026