{"id":20293,"date":"2026-09-08T13:01:48","date_gmt":"2026-09-08T11:01:48","guid":{"rendered":"https:\/\/haimagazine.com\/uncategorized\/how-far-are-we-from-agi-it-depends-on-whos-measuring\/"},"modified":"2026-09-21T12:47:02","modified_gmt":"2026-09-21T10:47:02","slug":"how-far-are-we-from-agi-it-depends-on-whos-measuring","status":"publish","type":"post","link":"https:\/\/haimagazine.com\/en\/it-and-technology\/how-far-are-we-from-agi-it-depends-on-whos-measuring\/","title":{"rendered":"\ud83d\udd12 How far are we from AGI? It depends on who\u2019s measuring"},"content":{"rendered":"

\u201cWelcome to the AGI era.\u201d According to Axios, those were the words Greg Brockman, OpenAI\u2019s co-founder and president, used to close a press briefing on the launch of GPT-6 Astra<\/mark><\/a>. The company itself was more cautious. In its official announcement, it called Astra \u201ca new generation of intelligence\u201d and its most intelligent model yet, but it didn\u2019t say: we\u2019ve achieved AGI. Axios<\/a>.<\/p>

That\u2019s probably for the best, because before we decide whether AGI already exists, we first need to agree on what it actually is.<\/p>

Artificial General Intelligence is usually understood as AI with general-purpose intelligence. The problem is that there\u2019s no single agreed definition of the term and no single test that would let us issue a certificate saying: \u201cthis is AGI\u201d. OpenAI, Google DeepMind, Anthropic and researchers working around ARC all look at the same goal from different angles. For some, the most important criterion is work. For others, it\u2019s the full range of human cognitive abilities. For others still, it\u2019s the ability to learn something genuinely new.<\/p>

It\u2019s not just \u201cvery smart AI\u201d<\/h4>

A computer beat the world chess champion back in 1997. AlphaGo defeated one of the world\u2019s best Go players in 2016. Neither system was AGI, of course, although both were important steps on the way toward it. They were phenomenal at one specific task and practically useless outside it.<\/p>

The key is the letter \u201cG\u201d: general.<\/p>

A system aspiring to AGI should perform well across very different domains, apply knowledge gained in one area to another, adapt to new situations and solve problems it wasn\u2019t specifically trained for. Almost everyone agrees on that much.<\/p>

But that\u2019s also where the agreement ends. How broad do those abilities need to be? Is it enough to perform most human work? Does the system need the full range of human cognitive abilities? Or is the most important question how quickly it can learn something it has never seen before?<\/p>

The answer determines how far away AGI is.<\/p>

OpenAI: AGI goes to work<\/h5>

OpenAI has long used an unusually practical definition. In the OpenAI Charter<\/mark><\/a> AGI is defined as “highly autonomous systems that outperform humans at most economically valuable work”. OpenAI Charter<\/a><\/p>

There\u2019s no requirement for consciousness, emotions or even human-like thinking. What matters is the result.<\/p>

Imagine a system that can handle financial analysis, write software, conduct legal research, prepare a marketing campaign, create a presentation, operate a computer and run a project. It does all of this autonomously and, in most of those tasks, better than the average person working in the relevant profession.<\/p>

By OpenAI\u2019s definition, that would be a very serious candidate for AGI.<\/p>

In this framework, the development of AI is increasingly about real work and autonomy rather than knowledge tests. GPT-6 Astra is supposed to carry out multi-step tasks in browsers and applications, handle programming, research and professional work. The OpenAI announcement<\/a><\/mark> from September 6 is even more interesting: the company claims it has reached the previously announced \u201cautomated research intern\u201d stage, meaning a system capable of completing well-defined research tasks under supervision that would take a skilled researcher several days.<\/p>

That still doesn\u2019t mean the definition in the OpenAI Charter has been met. \u201cMost economically valuable work\u201d is a much higher bar than performing selected computer-based tasks extremely well. But it shows why OpenAI may speak about getting close to AGI earlier than someone using a stricter definition.<\/p>

DeepMind: general intelligence should actually be general<\/h4>

Google DeepMind sets the bar differently.<\/p>

In Levels of AGI<\/a><\/mark>, a framework developed by its researchers, two basic dimensions matter: performance and generality, meaning the breadth of capabilities. Instead of a single dividing line between \u201cbefore AGI\u201d and \u201cafter AGI\u201d, we get a series of levels. A system can become increasingly capable and increasingly general until it reaches full generality.<\/p>

In 2026, DeepMind went a step further and proposed a cognitive framework<\/a><\/mark> for measuring progress toward AGI. It doesn\u2019t rely on a single benchmark, but on a broad set of human abilities, including learning, memory, reasoning, planning, metacognition and social intelligence. The authors argue that creating a system with the full set of human cognitive capabilities would mark a qualitatively new stage.<\/p>

This approach is more demanding than OpenAI\u2019s economic definition.<\/p>

A model could be excellent at coding, data analysis and operating applications while still being much worse than a human at understanding an unusual social situation, learning an unfamiliar activity or functioning sensibly in the physical world.<\/p>

So AI could already be taking over a significant share of human work and still not qualify as AGI by DeepMind\u2019s standards.<\/p>

ARC: don\u2019t just show what you know. Show how you learn<\/h4>

Fran\u00e7ois Chollet, the creator of the ARC benchmarks, frames the problem differently.<\/p>

In his view, having enormous amounts of knowledge isn\u2019t enough to prove general intelligence. A model may simply have learned countless patterns during training. What\u2019s much more interesting is what it does in a genuinely new situation.<\/p>

Therefore, ARC-AGI<\/a><\/mark> tests the ability to infer unfamiliar rules from a small number of examples. In the latest ARC-AGI-3, the system enters an interactive environment with no instructions and has to figure out how it works, what the goal is and how to achieve it. The ARC Foundation puts its definition very simply: \u201cAGI is here when a system can learn like a human\u201d. <\/p>

And this is where Astra\u2019s story becomes particularly instructive.<\/p>

OpenAI reports a score of 99.9% on ARC-AGI-3. That sounds almost like the AGI problem has been solved. But ARC Prize<\/a><\/mark> presents a more complicated picture. In the standard testing environment, Astra scored 62.7%. It reached 99.9% with a custom adapter that allowed it to preserve reasoning state across successive interactions. Just as importantly, the benchmark\u2019s creators emphasize that even solving it completely wouldn\u2019t prove AGI. ARC tests a specific component of intelligence in closed environments, not the full complexity of the world.<\/p>

So it\u2019s possible to \u201cpass the AGI test\u201d and still not be AGI.<\/p>

Anthropic: what if, instead of AGI, we build a digital Nobel laureate?<\/h4>

Anthropic complicates the picture further because AGI itself isn\u2019t central to the way the company communicates. It more often talks about \u201cpowerful AI\u201d.<\/p>

The definition, however, is concrete. Such a system would match or surpass Nobel Prize winners in most important fields, use the same digital tools available to humans and autonomously carry out complex tasks. Anthropic has predicted that such systems could emerge in late 2026 or early 2027, while its updated safety roadmap<\/a><\/mark> still assumes the possibility of \u201cpowerful AI\u201d appearing within roughly one to two years.<\/p>

Answering a difficult question is one thing. Running a complex project independently for many hours or days is something else. That\u2019s why the frontier of AI capabilities is increasingly being measured not only by how good the model\u2019s outputs are, but also by how long it can keep operating productively without a human guiding it step by step.<\/p>

So who\u2019s closest?<\/h4>

After all these definitions, the answer should be clear: there\u2019s no honest way to create a table saying OpenAI is at 92% of AGI, Anthropic at 87% and Google at 84%.<\/p>

No such scale exists.<\/p>

OpenAI places the greatest emphasis on real-world work and increasingly long periods of autonomy. Anthropic is building toward the idea of an extraordinarily capable autonomous expert. DeepMind sticks to a much stricter concept of generality and a full range of cognitive abilities. ARC focuses on whether a system can genuinely learn new things efficiently.<\/p>

Depending on the definition you choose, you can therefore get a different answer to the question of who’s ahead.<\/p>

And the race is no longer limited to the United States.<\/p>

DeepSeek<\/a><\/mark> officially describes AGI as its \u201cultimate goal\u201d. Alibaba<\/a><\/mark> has presented Qwen from the beginning as a \u201cproject toward AGI\u201d. On Moonshot AI’s<\/a><\/mark> website, the maker of Kimi claims directly that its research team \u201cworks toward AGI\u201d. MiniMax<\/a><\/mark> likewise says it’s building toward artificial general intelligence.<\/p>

This matters because it shows how the meaning of the acronym itself has changed. AGI is no longer just a hypothetical concept from academic debate. It has become the name of a research goal pursued by some of the world\u2019s biggest AI labs, even if each one may understand that goal slightly differently.<\/p>

How will we know AGI when we see it?<\/h4>

Instead of looking for one magic test, it\u2019s more useful to ask a few simpler questions.<\/p>

Can the system perform a very broad range of tasks instead of excelling in only a few domains? Can it transfer knowledge between them? Can it learn something genuinely new as efficiently as a human? Does it perform at a human or superhuman level? And finally, can it do all of this autonomously for extended periods without constant instruction?<\/p>

Today\u2019s models increasingly answer \u201cyes\u201d to some of these questions. To all of them at once\u2014not yet in a way that would settle the debate.<\/p>

That\u2019s why we probably won\u2019t wake up one morning to a universally agreed announcement that AGI now exists. Something less dramatic is more likely. One boundary after another will disappear, while the argument over definitions continues even after systems begin performing tasks that, only a few years earlier, we would have readily accepted as proof of general intelligence.<\/p>","protected":false},"excerpt":{"rendered":"

AGI has long been the goal of the world\u2019s biggest AI labs. The problem is that OpenAI, Google DeepMind, Anthropic and Chinese players don\u2019t always mean the same thing when they talk about \u201cgeneral intelligence\u201d.<\/p>\n","protected":false},"author":354,"featured_media":20202,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"rank_math_lock_modified_date":false,"footnotes":""},"categories":[803],"tags":[],"popular":[],"difficulty-level":[38],"ppma_author":[776],"class_list":["post-20293","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-it-and-technology","difficulty-level-medium"],"acf":[],"authors":[{"term_id":776,"user_id":354,"is_guest":0,"slug":"redakcja","display_name":"Redakcja","avatar_url":{"url":"https:\/\/haimagazine.com\/wp-content\/uploads\/2025\/07\/Zrzut-ekranu-2025-07-10-o-16.00.36.png","url2x":"https:\/\/haimagazine.com\/wp-content\/uploads\/2025\/07\/Zrzut-ekranu-2025-07-10-o-16.00.36.png"},"first_name":"","last_name":"","user_url":"","job_title":"","description":""}],"_links":{"self":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts\/20293","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/users\/354"}],"replies":[{"embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/comments?post=20293"}],"version-history":[{"count":1,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts\/20293\/revisions"}],"predecessor-version":[{"id":20294,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/posts\/20293\/revisions\/20294"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/media\/20202"}],"wp:attachment":[{"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/media?parent=20293"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/categories?post=20293"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/tags?post=20293"},{"taxonomy":"popular","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/popular?post=20293"},{"taxonomy":"difficulty-level","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/difficulty-level?post=20293"},{"taxonomy":"author","embeddable":true,"href":"https:\/\/haimagazine.com\/en\/wp-json\/wp\/v2\/ppma_author?post=20293"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}