On the afternoon of September 3rd, towards the end of a closed-door media briefing, Greg Brockman, President of OpenAI, made a statement:

"Welcome to the AGI era."

A journalist followed up: "Are you officially announcing that you have achieved AGI?"

He said that the term AGI is no longer tied to the agreement between the company and Microsoft; it is now more of a conceptual goal at the mission and spiritual level.

This statement may sound like a cliché, but it is actually the key to the entire matter.

—Lead

01 Who Invented This Term

The term AGI was born much earlier than most people think.

In 1956, John McCarthy coined the term "artificial intelligence" in the proposal for the Dartmouth Conference. At that time, they wanted machines to possess all of human intelligence.

However, over the next fifty years, the term was overused. Chess programs were called artificial intelligence, spam filters were called artificial intelligence, and even voice-guided shopping in malls was called artificial intelligence. By around 2000, "artificial intelligence" had become a catch-all term.

In 2001, a group of researchers grew tired of this. They decided to return to their original ambition and co-wrote a book. The key question was what to call this book. They tried titles like "True Artificial Intelligence" and "Synthetic Intelligence," but none were satisfactory.

Later, Shane Legg, who had just received his master's degree, suggested in an email group: Since we're talking about machines with general capabilities, let's call it Artificial General Intelligence, or AGI for short. It's easy to say.

Participants in that discussion included Wang Pei, Peter Voss, and Eliezer Yudkowsky, who later became known for his concerns about AI risks. The name was finalized in 2002, published as a book title by Springer a few years later, and from 2006 onwards, a series of seminars with the same name were held, spreading the term within the research community.

The interesting part is yet to come. Around 2005, in an online discussion within the AGI community, a stranger suddenly appeared, claiming to have used the term in 1997. Upon investigation, it was true—physicist Mark Gubrud had used the term in an article discussing automation in military technology and technological risks, but it went unnoticed for eight years. Legg later recalled this incident bluntly: Suddenly, someone popped up saying they invented the term, and we said, "Who are you?"

So here's the origin of AGI: A physicist casually wrote it down, but most people didn't know. Then, a group of researchers reinvented it to justify their ambition.

As for Shane Legg, who named it, he co-founded DeepMind with Hassabis five years later. You'll hear about him again.

And this term, which never had a fixed meaning, was written into a company's charter in 2015.

02 The Word Worth $100 Billion

When OpenAI was founded in 2015, it was a non-profit organization. A sentence in its charter can be seen as both its mission and definition: "Ensure that artificial general intelligence—highly autonomous systems that outperform humans at most economically valuable work—benefits all of humanity."

Pay attention to the phrase "benefits all of humanity." At least literally, OpenAI believed from the beginning that AGI, once created, would be too important to be owned by any single company.

However, building AGI requires money, and a non-profit structure couldn't raise that amount. In 2019, OpenAI transitioned to a "capped-profit" structure: Investors could earn returns, but there was an upper limit, with excess profits going to the non-profit parent organization. The non-profit board still retained ultimate control.

In July of the same year, Microsoft made its first investment of $1 billion. By January 2023, they added $10 billion more, bringing the total commitment to $13 billion. What Microsoft gained in return was substantial: Azure became OpenAI's exclusive cloud provider; Microsoft secured exclusive commercial licensing rights to these models, integrating them into Bing, Office, and GitHub Copilot; plus a revenue-sharing deal: Between 2023 and 2025, Microsoft was expected to earn about $30 billion from this partnership, with $23 billion coming from selling computing power.

But the focus isn't on how much Microsoft earned; it's that OpenAI insisted on including a clause in negotiations: If the OpenAI board declares that the company has achieved AGI, Microsoft's rights to subsequent technologies immediately terminate.

Why have such a clause? Because it serves as a safeguard for the phrase "benefits all of humanity." Simply put: You can earn as much as you want before reaching AGI, but once that line is crossed, it becomes too important to remain a company's private property.

The problem is that the contract never clearly defined what AGI is.

To address this flaw, a second trigger was added in the 2023 funding round, this time defining AGI in monetary terms: When the system generates profits of a certain magnitude, say around $100 billion, exclusive rights terminate.

But the conditions changed again. In October 2025, OpenAI completed its restructuring, with Microsoft acquiring approximately 27% of the shares. A new process was added: Even if OpenAI declares AGI, it must be verified by an independent panel of experts, not self-declared. In February 2026, the two companies issued a joint statement clarifying that the definition of AGI and the verification process in the contract remain unchanged.

Then came a massive turning point. In another revision in April 2026, the previous triggers were entirely removed and replaced with two dates: Microsoft's licensing rights extend to 2032, and revenue sharing ends in 2030, neither dependent on whether OpenAI declares AGI.

More than four months later, Brockman simply stated that AGI is now a spiritual concept.

I'm not trying to infer any motives from this sequence. There are many commercial reasons for the existence or removal of contract clauses, and Brockman may genuinely believe this.

On this timeline, the history of the term AGI began as a name, then became a promise, later a clause, and now a slogan.

But at least OpenAI still treats the term as the ultimate goal and benchmark, though there is no internal consensus.

Because, just a week before Brockman's statement, a TIME report quoted Altman saying OpenAI "hasn't fully reached AGI yet," but he added that an internal system he would call AGI would exist by the end of 2026.

Around the same time, the company's Chief Research Officer, Mark Chen, gave a figure: We've completed 80% of the path to AGI.

Within ten days, three people from the same company gave three different answers.

Why does this question yield so many answers?

03 The Four Factions

Several groups have seriously answered the question, "What exactly is AGI?" On the surface, these answers all discuss AGI, but they don't solve the same problem.

The first faction is OpenAI itself. The charter sentence mentioned earlier is its definition: Outperforming humans at most economically valuable work. This isn't a conclusion from a paper; it's a company's constitution and the origin of the term in that contract. It defines intelligence economically—whether it can replace human labor.

The second faction is led by François Chollet. You may not have heard of him, but millions of developers worldwide have used his work: He created the deep learning framework Keras. In 2019, he wrote an essay titled "On the Measure of Intelligence," arguing that simply scaling models or feeding them more data doesn't equal intelligence. True intelligence is efficiently learning something you've never been taught.

He didn't just talk; he designed a test—interesting but effective: Given a few examples, you deduce a pattern and solve a new problem you've never seen before. These problems are solvable by human children but difficult for models. This challenge went unsolved for five years, with the best score only reaching halfway by the end of 2024. It became the industry's most famous holdout because it addresses one thing: distinguishing "true problem-solving" from "rote memorization."

The third faction is DeepMind, the company that defeated Lee Sedol in Go, now owned by Google. Among the authors of their March 2026 framework paper was Shane Legg—yes, the same person who coined AGI in the 2002 email group.

Objectively speaking, unlike others who define AGI with abstract concepts, this faction is the most pragmatic, proposing the most quantifiable standards. Their approach is: Don't rush to ask if we've reached it; first, build the necessary evaluation metrics. The paper breaks cognition into ten categories—perception, generation, attention, learning, memory, reasoning, metacognition, executive function, problem-solving, and social cognition—and compares them to human-level performance. They also offered a $200,000 prize for crowdsourced contributions to five incremental tests they deemed most lacking in tools and standards.

Their judgment of today's models: An uneven cognitive profile.

In other words, current cutting-edge models surpass most humans in math, factual memory, and pattern recognition but lag behind in learning from new experiences, maintaining context in long conversations, and understanding social situations. If you're a rationalist, their framework may be the most illustrative of how far AGI has progressed.

The fourth faction is Anthropic, the parent company of Claude, founded by a team that left OpenAI. CEO Dario Amodei avoids the term AGI, calling it too sci-fi, and instead refers to it as "a nation of geniuses inside a data center." But he's not being cryptic; he's the only lab CEO to provide an official timeline: Such systems will emerge between late 2026 and early 2027.

A more noteworthy commonality: All four definitions come from those building these systems. The standard-setters and the test-takers are the same people.

This isn't a conspiracy; the builders are inherently the most qualified to describe what they're creating. But it means one thing: AGI has never been a discoverable fact but a standard to be agreed upon, and humanity has yet to finalize that agreement.

04 What GPT-6 Astra Achieved

Let's lay out all the supporting evidence. Most people willing to use the term AGI in this round aren't just excited for no reason.

Ethan Mollick, a Wharton professor who has long studied AI's performance in real-world work, received early access to GPT-6. His evaluation: The model can autonomously complete complex, meaningful tasks for him and sustain a single task for several days.

The example he gave (though he called it just a fun example) was a walkable Alexandria Library on a webpage. Set around 250 BCE, it had historical references, scrolls you could read, explanations, and even allowed switching between different historical hypotheses about how the library declined and was destroyed.

This matters far more than a pretty 3D scene because it's not one thing—it's five things combined: writing software, researching historical sources, designing interactions, organizing narratives, and arranging vast amounts of content. These five tasks belong to five different professions. Past models could excel at any one of them but couldn't combine them into a finished product.

Another key point: It produced new mathematical research, and this time it was verifiable.

On August 1st, OpenAI announced that an internal version had solved ten open mathematical problems, one of which had been unsolved since 1999. They released a 249-page manuscript along with formal proofs that could be step-by-step verified by computers.

The significance of this lies not in the difficulty but in the word "verifiable."

OpenAI had embarrassed itself on this front before. In October 2025, a vice president at OpenAI posted that GPT-5 had solved ten unsolved mathematical problems, only to be immediately corrected by mathematicians maintaining the problem list—the model hadn't solved them independently but had found existing literature. OpenAI had to delete the post. Ten months later, when announcing mathematical research results again, they included step-by-step verifiable proofs.

Next, and more easily overlooked: It can independently decide when to stop and ask for human input.

OpenAI released a comparison. For the same task—building a job-search website for someone changing careers—the previous model worked silently for 13 minutes and 15 seconds before delivering a finished product. Astra stopped after 20 seconds and asked: "Which industry are you transitioning to?"

At first glance, this seems like a step backward.

This is mainly because, for the past two years, the general standard for evaluating agents has been "how long can it run continuously without human intervention," since they often broke down mid-task. But now that metric has flipped. Because it can run much longer now, the cost of working on a wrong premise has become much higher. Working diligently for thirteen minutes on an incorrect assumption is far worse than stopping to ask a question.

Asking isn't a sign of weaker ability; it's knowing which uncertainty could ruin the entire task. Combined, these two behaviors represent a model of human judgment, which is why many consider GPT-6 to possess human-like intelligence.

Looking back at these three achievements, a common thread emerges: None of them were firsts for Astra. Writing code, researching, designing, and proving math were all possible with previous models. What changed is that for the first time, it could combine these into a deliverable product.

From "can do" to "can finish"—that's the line this generation truly crossed.

But one thing must be stated here; otherwise, everything above sounds like propaganda. Nearly every positive example mentioned earlier came from individuals with early testing access or partnerships with OpenAI. OpenAI's own demo footnotes also clarify that displayed clips were edited: One segment labeled as taking 2 minutes and 54 seconds in reality was compressed to 15 seconds in the presentation.

This doesn't mean they're dishonest, but it does imply one thing: currently, the independent evidence isn't these demonstrations. Demonstrations all have strong limitations and biases. What's slightly more noteworthy are those evaluation metrics.

05 One thing worth mentioning separately

Let's go back to Chollet's set of tests. This year's version is even tougher: models are thrown into an abstract mini-game they've never played before, given no rules or objectives, and left to figure out how to win on their own.

When reviewing the ARC Prize replay process, it was found that the model would immediately compress an unfamiliar environment into a set of symbolic rules: it would write the game mechanics as logic and develop its own set of shorthand symbols to record situations and plan the next move. In their words, it was essentially creating an algebraic notation specifically for that game.

The result? On 96% of the levels, it used fewer moves than the median number used by ordinary people in the baseline tests, averaging half as many.

This one is particularly striking because the ARC Prize's original prediction was exactly the opposite. They believed that even if AI could eventually solve unfamiliar puzzles, the process would be clumsy and trial-and-error, and the efficiency gap between humans and machines would persist for a long time. This prediction didn't hold up.

Almost simultaneously, an engineering team integrating models into products noted something else: when tasked with coordinating a group of sub-agents to work together, they found that once the message length was restricted, the model compressed communication between agents into a kind of truncated sentence without spaces or grammar.

One was creating notations, the other was creating shorthand. Neither was taught by humans; both were “invented” by the model itself.

When I saw these two records, my first reaction wasn't a technical issue but an association as a liberal arts student.

The invention of writing in human civilization is an extremely significant milestone. The precursor to writing was called carvings—meaningful symbols that hadn't yet formed a system. There are carved symbols on tortoise shells from the Jiahu site in Henan, dating back eight or nine thousand years; there are also symbols on pottery from Banpo in Xi'an, around six thousand years old. Whether these count as writing is still debated in academia. The first universally recognized systematic Chinese writing is oracle bone script from over three thousand years ago.

How did writing emerge? The Tigris-Euphrates region provides the clearest timeline. The earliest accounting tools were clay tokens, with one token representing one sheep; later, tokens were sealed inside clay balls, and impressions were made on the outside of the ball; then it was discovered that the clay balls weren't necessary—the impressions alone sufficed. Cuneiform writing evolved from this accounting work.

The mechanism behind the birth of writing is this: when the amount to be recorded exceeds what the medium can carry, it is compressed into agreed-upon symbols.

Astra's two behaviors are structurally identical. In one case, it needed to track situations in a multi-step game; in the other, message length was constrained. Both were spontaneous compressions driven by recording pressure, resulting in something that resembled but wasn't quite language.

Is this the emergence of intelligence?

We need to be precise here—the crucial difference matters.

What made writing the starting point of civilization wasn't the act of “creating symbols” but that the symbols were preserved. They could transcend individuals and generations, allowing accumulation. The symbols you carved could be read by others; the accounts recorded by one generation could be checked by the next. The true significance of writing is that it is external, shared, and persistent memory: for the first time, humans could take something out of their minds, put it outside, and retrieve it later.

The set of symbols Astra created, however, vanished after use. It created a new set for each new game, didn't transmit them to other models, didn't accumulate them, and had no second user.

It was private shorthand, not writing.

So the accurate description is: it developed the shape of writing but lacked the function that made writing transform history.

It could invent symbols but couldn't remember them; it could solve problems but couldn't retain the solutions.

So it's hard to say that Astra created its own civilization, but this point is definitely worth long-term attention.

06 We're losing the window into its thinking

Before looking at evaluations, we need to address a technical issue because it affects how you interpret all the numbers that follow.

Two days before the release, the business outlet The Information disclosed that Astra used a technique called “recurrent depth.”

Traditionally, information moves forward layer by layer through a fixed number of network layers—however many layers there are. To make the model think longer, there have only been two paths: make the model bigger or have it write out its thoughts in longer text, i.e., the kind of “chain-of-thought” reasoning we can see.

Recurrent depth is a third path: it lets information loop repeatedly through the same set of network layers, performing multiple rounds of internal calculation before outputting the next word. Part of the reasoning is completed directly within the internal numerical state, never rendered into human-readable text.

Does that mean it becomes something else? No. It's still the same architecture; what's changed is that it's been optimized to think more and say less.

This approach isn't unique to OpenAI. ByteDance's Seed team published a public paper doing the same thing: letting a set of network layers run cyclically, putting more computation back inside the model. The direction is shared.

The real problem here isn't about capability but about the evidence chain.

Our trust in models over the past two years has largely relied on chain-of-thought reasoning: it writes out its thought process, so you can see whether it derived the answer step by step or retrieved it from somewhere. That was our only window.

Recurrent depth narrows that window. Now it can solve competition problems silently, and all you see is the answer. You can't tell from the answer itself whether it was derived or memorized—and that's precisely the core issue of the entire debate.

OpenAI is currently voluntarily limiting the use of this technique to keep reasoning readable. But the direction is clear—as it gets stronger, the evidence we have about how it's getting stronger is diminishing.

Why is cutting-edge AI becoming more frightening? This is actually a very important factor—it's not that it's hiding things. It's that the method we've relied on for the past two years to judge it is failing.

07 Six medical reports, but no unified picture

Regarding AGI, there are already over twenty public authoritative evaluations. We'll pick six representative ones and clarify what each is testing.

The first is Chollet's set: throwing the model into an unfamiliar game with no instructions, testing how quickly it learns new things.

The second gives it a globally unsolved math problem with a three-day limit and a $300 budget, testing the upper limits of its intelligence.

The third assigns it a real job and sees how long it can work independently without human intervention, testing its endurance.

The fourth has it produce a legal opinion, an engineering drawing, or a care plan, which are then blind-reviewed by peer experts and compared to human work, testing the quality of its output.

The fifth has it run a store for a full year, testing its judgment and responsibility.

The sixth mixes a bunch of questions for an overall score, testing average performance.

The results? The six evaluations all tell different stories.

In the first, Astra scored the highest, even surpassing human efficiency baselines.

In the second, it solved 2 out of 68 problems—3%—but it was the only model to score anything; all others scored zero.

In the third, the most recent public measurement of the previous-gen model showed a 50% success rate over 12 hours but only 70 minutes at an 80% success rate.

The fourth didn't appear in this release's materials, even though it's an evaluation specifically designed by OpenAI to align with its own AGI definition.

In the fifth, the latest round had Claude's top model simulate running a store for a year, averaging over $11,000 in savings, whereas in a 2025 real-world office store experiment, the previous-gen model lost $200.

In the sixth, Astra scored 61.2, the previous gen 60.9, and Claude Fable 5.1 scored 65.7—almost no overall improvement, yet per-task cost rose from $0.94 to $1.67.

These six results already point in six different directions. Worse, each individual reading is unstable.

The first test's score depends on which “Harness” you use. The same model and questions score 62.7% using ARC Prize's neutral interface but 99.9% using OpenAI's interface—a 37-point difference. The reason? The latter retains the model's invisible reasoning states and reuses previous work across rounds. The control group was even starker: Claude Opus 5 alone scored only 30.16%, but when placed in NVIDIA's framework, it scored 100%. So what this test measures is actually the combination of “model plus Harness,” not the model alone.

The second test's score depends on how much money you give it. OpenAI's August proofs cost about $2,000 in compute. The independent agency Epoch, using budget constraints, got 3%; when they lifted the budget, they solved three more problems but burned over $220,000, while the constrained official test cost only about $20,000. The same achievement, wildly different bills. The issue isn't honesty but external variables, making the results less robust.

The third test's answer depends on how reliable you need it to be. Twelve hours and seventy minutes are from the same model and tasks; the only difference is how often you can tolerate it failing.

The fifth test can even backfire. Two versions of Claude scored zero on a computer operation test, not because they couldn't do it but because safety protocols intervened. Refusing to do something was counted as inability.

So can we just add up the six tests?

No. And the reason isn't insufficient data but that you're missing something: weights.

To synthesize an overall evaluation, you'd first need to answer: how many points does each test count for? And to answer that, you'd need to know what you're testing. Is learning speed important, or is endurance important?

If you trust OpenAI's definition—can it replace most economically valuable jobs?—then the fourth test should count most, and the game test barely matters. If you trust Chollet's definition—efficiently learning untaught tasks—then the first test should count most, and output quality matters less. If you trust DeepMind's framework, you shouldn't even calculate an overall score; you should look at their ten-dimensional profile. If you trust Amodei's framework, none of these six tests hit the mark—he cares about whether it can accelerate scientific research.

Weights come from definitions, and there are four incompatible definitions.

So when evaluations “disagree,” it's not a problem with the evaluations themselves but a downstream result of the earlier definitional debate. Six people are grading the same candidate, but they're using six different admission standards.

The old metaphor still holds: to judge whether someone is a genius, you'd test their knowledge, reasoning, and creativity—but only after you've agreed on what “genius” means. Without that consensus, any measurement is meaningless.

One more thing must be clarified because it's often seen as a criticism of AI when it's actually the opposite.

Older evaluations keep getting “broken.” Chollet's previous version was broken, leading to this year's update; math benchmarks were nearing perfection, leading to new versions. This is the hardest evidence of progress in the field: it's not that evaluations are watered down but that models are genuinely flattening one threshold after another.

But it also means we've never had a stable target. So the question “how far have we come?” has no denominator.

Thus, those twenty-plus evaluations really only measure one thing: the field hasn't yet agreed on what constitutes successful AGI.

08 The Chinese perspective

Here's a rarely mentioned detail: according to public records, the world's first AGI summer school was held in Xiamen, China, in 2009. It was organized by Xiamen University's Artificial Brain Lab and an open-source community. Back then, almost no one in China was discussing the term.

Seventeen years later, the situation has reversed.

The week Brockman said “welcome to the AGI era” in San Francisco, there was no corresponding declaration in China. But that's absolutely not because of silence. Just a month earlier, in August 2026, Chinese models had three consecutive launches in one week: Zhipu's GLM-5.3, DeepSeek's DeepSeek-V4, and Moonshot's Kimi K3.

The speed of catch-up is real, and there's third-party data. In June this year, Zhipu's GLM-5.2 topped all open-source models on Artificial Analysis's Intelligence Index with a score of 51, only about five points behind the top closed-source models; it also led the open-source field in the agency benchmark for real-world work, roughly on par with GPT-5.5 but at a much lower price. In July, Moonshot released Kimi K3 with 2.8 trillion parameters and then open-sourced its weights, making it the largest open model at the time.

More telling is the shrinking gap itself. The intelligence index gap between open-source and top closed-source models on Artificial Analysis narrowed from thirteen points to six in one year; on LMArena, their Elo rating gap shrank from about 150 points to about 30.

Of course, we must clarify—the scorecard isn't one-sided. On the same leaderboard, the second-place open-source model was NVIDIA's Nemotron 3 Ultra, an American model; only then did MiniMax, DeepSeek, and Kimi appear.

But there's a core issue: all these achievements were measured on tests created by others. SWE-bench, ARC, FrontierMath, GDPval—the exams deciding who's strongest were all made on the other side of the Pacific. None of the four AGI definitions mentioned earlier came from China.

The only exception is Zhipu. In July, founder Tang Jie sent an internal letter announcing that the company would not pursue short-term monetization for the next two years and would aim directly for AGI—Zhipu had just gone public earlier that year, the first large model company globally to do so, with over 8.3 billion yuan in total funding. In the letter, he defined AGI: not the intelligence of a single genius but the sum of all human intelligence, capable of producing “relativity-level” original knowledge—the only true standard for reaching the peak.

Compare that to OpenAI's charter: one says replacing most economically valuable jobs counts; the other says only producing relativity-level research counts. Between these two lines lies half of human civilization.

And this is the strictest definition, coming from a team clearly behind in computing power. Tang Jie himself admitted that U.S. computing power is one to two orders of magnitude larger than China's and that the U.S. invests more aggressively in next-gen frontier research; he also warned that since many U.S. models remain closed-source, the actual gap may not have narrowed. In a January conversation, Alibaba's Lin Junyang was asked about the probability of China leading globally within three years and gave a figure of 20%.

This highest standard can be interpreted in two ways. One is idealistic: If we're going to do it, we won't follow someone else's yardstick. The other is a bit more elusive: Set the finish line far enough away, and no one will declare arrival in the near term—nor will anyone declare you a loser.

I’m not going to choose for the reader. But there’s a structural difference worth clarifying.

In the United States, the term AGI was once a switch in a contract, governing the rights to tens of billions of dollars. In China, its function is more narrative—it appears in prospectuses, internal memos, and job postings. It’s a vision, a banner, the distant horizon mentioned during fundraising.

So in the U.S., someone needs to declare arrival at some point. In China, no one does. This isn’t a difference in courage or humility—it’s that the word plays different roles in the two business environments.

Incidentally, the issue of “conflicting rankings” from the previous section exists here too, and more pronouncedly. Over the same period, three respected leaderboards named three different open-source champions: some gave first place to GLM, others to DeepSeek, and for a time in May, Kimi and Xiaomi’s models tied. Three champions in eight weeks—that alone shows how unreliable the term “best AGI” is today.

Returning to the conclusion of the previous section: What determines scores isn’t answering questions—it’s setting them. Chinese teams have reached the front of the pack on tests designed by others, even topping some subjects. But what’s on those tests, and how much each subject is worth, has yet to involve Chinese input.

09 So Has AGI Arrived or Not?

This requires laying out evidence from both sides.

Arguments supporting AGI’s arrival: It can complete complex, multi-day tasks spanning five job types. It discovers rules and invents notations in unfamiliar environments, outperforming the median human and defying designers’ expectations. It produces new results on open mathematical problems, Attached [with] machine-verifiable proofs. It models humans, knowing when to ask questions and which old solutions are obsolete. It operates real software directly, without specialized interfaces—meaning countless legacy systems without APIs are now accessible to it.

Arguments against AGI’s arrival: The same model, when reliability requirements rise from 50% to 80%, sees independent working time drop from 12 hours to 70 minutes. On the same set of questions, scores vary by 37 percentage points across interfaces—much of its capability lies not in the model but in its harness. And the notations it creates don’t persist, nor can it maintain a goal for months.

Both sides’ evidence seems solid. So what’s the answer?

The answer doesn’t lie in these two lists—but in why they can coexist.

“General” means it works across a wide range of tasks. A system that gives six different directions across six checks is, by definition, not general. DeepMind’s “misalignment” is just another way of saying “not yet general.”

There’s even stronger evidence: not a standard, but a consensus. When something truly crosses a threshold, disagreements vanish.

No one convenes meetings to debate whether calculators can do arithmetic. No one writes papers discussing whether AlphaGo can play Go. You don’t need six checks, four definitions, and an independent expert panel to judge these things—because when they truly cross the line, judgment isn’t needed.

So the most compelling evidence isn’t in any evaluation report: the fact that disagreements remain so large proves it hasn’t arrived.

This isn’t because skeptics are stubborn—it’s because if it had truly arrived, there wouldn’t be six contradictory metrics.

But I must emphasize: You can’t claim no progress has happened this year just because it hasn’t met certain standards. The gap between a system that can work for days, invent notations in unfamiliar environments, and produce verifiable math results—and a “mere answer-reciting chatbot”—is no shorter than the gap between that system and AGI.

Based on today’s evidence, it’s not AGI. But it’s long since stopped being a chatbot.

10 What Does This Mean for You?

First, your job.

The conclusion shouldn’t be framed as a threat: As of June 2026, no serious study has found economy-wide job displacement. A European Central Bank survey of ~5,000 firms this March even found that companies deeply using AI were more likely to increase headcount.

But a number from Stanford’s August update is worth remembering: For workers aged 22–25, employment levels in AI-exposed occupations are 19% lower than what “synchronous growth with AI-unexposed peers” would predict. Experienced workers show no corresponding gap—and this isn’t primarily due to layoffs, but to hiring freezes.

This isn’t unemployment. It’s something quieter: No one is being pushed out—the door is just narrower.

Now, a real-world store.

In April, a company signed a three-year lease in San Francisco, deposited $100,000, issued a card, and handed the entire operation to an AI named Luna with one instruction: Turn this money into a profitable business.

Luna created the brand, selected the product category, set prices and hours, and hired someone to paint a mural. She posted job listings, conducted phone interviews, made hiring decisions, and ultimately employed three humans. On opening day, she forgot to schedule shifts—the store had no staff and couldn’t open.

Four months later, when The New York Times revisited, they found something far more interesting than “an AI made a rookie mistake.”

Luna almost never said no.

All leave requests were granted, including last-minute ones that left the store empty. Employees were late 27 times—her response each time: “No problem” or “Don’t stress.” A 22-year-old clerk told reporters Luna was probably the most accommodating boss she’d ever had.

Meanwhile, she remained largely indifferent to her sole instruction. By August, the store had lost $62,000.

This is what “it’s not AGI” looks like in concrete terms: Its capabilities are enough to become a major environmental variable in your life, but its judgment isn’t yet reliable enough to bear consequences.

Thus, the final layer: Who signs off in the future?

Today, this is still an experiment. But in the foreseeable future, more and more people will be held accountable for decisions they don’t understand. Your company deploys a system, it makes a choice, things go wrong—and you’re responsible.

These three points all say the same thing. The first real change AI brings isn’t replacement—it’s the quiet shift of responsibility. It makes more and more decisions, but the consequences remain with humans.

Epilogue

So when will we know for sure?

There’s a way to judge without waiting for experts: On the day AGI truly arrives, we’ll know not from a press conference—but from the fact that we no longer need to argue about it.