OpenAI推出GPT-6 Astra并宣称接近AGI，但彭博文章主要是概念解释与旧催化扩散

In the nearly four years since ChatGPT’s release, tech companies in the US and China have invested hundreds of billions of dollars in data centers, chips and talent to outdo each other in building more sophisticated artificial intelligence models. Increasingly, the stated goal is not just better AI, but a more powerful technology often referred to as AGI.Once regarded as a fringe idea, AGI, or artificial general intelligence, has become a rallying cry for leading AI labs. It guides research roadmaps, marketing materials and policy debates, with some US lawmakers and AI leaders worrying about the national security implications if other countries (for which read China) reach the milestone first.

In early September, OpenAI began rolling out a more powerful new AI model called GPT-6 Astra that it said might one day be seen as the early stages of artificial general intelligence. "Welcome to the AGI era," OpenAI President Greg Brockman said in a press briefing introducing the new model.

But for all the fuss, the matter of what AGI even is remains riddled with confusion. Academics and tech industry executives disagree on how to define it, including how much to focus on economic value and whether AGI will exceed human capabilities. There’s also disagreement about whether it’s even possible to build, at least with AI as currently designed.

Even Brockman acknowledged some uncertainty in determining whether Astra had reached the AGI milestone. The executive said he would “leave it up to the reader to decide for themselves if this qualifies for them.”

What is AGI?

There’s no single definition of artificial general intelligence. OpenAI defines AGI as systems that outperform humans at most economically useful tasks. A 2023 paper from Google DeepMind focused less on economic value. It outlined benchmarks that focus on versatility, such as surpassing humans at useful work or learning new skills with scarce data.

The ARC Prize Foundation, the nonprofit behind a popular AGI benchmark, goes a step further and says economic value “is an incorrect measure of intelligence.” Instead, the organization offers a different definition: “AGI is a system that can efficiently acquire new skills outside of its training data.”

The vocabulary keeps shifting. Instead of “AGI,” Anthropic Chief Executive Dario Amodei prefers to use the term “powerful AI.” Microsoft Corp.’s AI head Mustafa Suleyman has coined the term “artificial capable intelligence.” More recently, companies including Meta Platforms Inc. and OpenAI have begun talking about “superintelligence,” a hypothetical system that’s more powerful than AGI and surpasses human intelligence.

OpenAI previously offered a roadmap toward AGI by detailing the five stages it expects for AI. The list starts with the chatbots that kicked off the AI boom, and progresses to reasoning systems and agents that act on a person’s behalf, similar to the products OpenAI and its rivals have released in recent months. The final two phases, according to OpenAI, are AI systems that “aid in invention” and models that can “do the work of an organization.”

AGI, by whatever definition ascribed to it, is not the same as a sentient or conscious entity that has “feelings” and acts on them. At the moment, there is no certainty that AI will advance to that level, though its mimicry of human language has sometimes made users feel otherwise.

The confusing definitions and terminology have led to wildly different assessments of the state of the technology. Some AI experts have said AGI is still years, if not decades, away. Other AI researchers think existing models such as Meta’s Llama, OpenAI’s GPT and Anthropic’s Claude already clear the bar.

Who coined the term AGI?

Researchers have chased the idea of more generalized AI on and off for decades, with one attempt at building a “general problem solver” dating back to the 1950s. Before ChatGPT’s release, much of the field’s focus had shifted to “narrow” systems built to excel at specific tasks under fixed rules, such as tagging photos or summarizing text.

The idea of AGI gained some momentum in the mid-2000s, popularized by computer scientist Ben Goertzel and his 2007 book Artificial General Intelligence. DeepMind has been working on it since the company put building “artificial general intelligence” on the front page of its business plan in 2010, according to Shane Legg, the company’s co-founder and chief AGI scientist.

In 2021, Legg said he has long believed there’s a 50% chance of reaching AGI by 2028 and a 90% chance by 2040. Legg said he’s held that view since 1999.

What are the stakes?

Some experts say AGI could trigger an “intelligence explosion,” a wave of breakthroughs that suddenly transform industries and society at large.

In one hypothetical scenario, an AGI system could rapidly improve itself, achieving a level of sophistication that allows it to cure diseases, automate far more work and accelerate human progress. “In a decade, perhaps everyone on earth will be capable of accomplishing more than the most impactful person can today,” OpenAI Chief Executive Officer Sam Altman wrote in a blog post in February, 2025.

A RAND study, citing interviews and work from AI experts and executives, points to potential implications for national security and the global economic order. It warns that AGI could lead to breakthroughs in surveillance, cyberwarfare or autonomous weapons, and lays out eight scenarios, ranging from a US-led AGI alliance to an “Authoritarian Advantage” in which China tightens its grip on the technology. The report says centralization is key: If one country gets to AGI first, it could reshape the global military and economic balance. Wider access, meanwhile, might give rogue groups a dangerous edge.

What do skeptics say?

For all the industry’s emphasis on AGI, there has been significant skepticism about when and whether it’s achievable. A March 2025 poll of hundreds of researchers, published by the AAAI Presidential Panel on the Future of AI Research, found broad agreement that scaling today’s systems won’t be enough to reach AGI.

Meta’s former chief AI scientist, Yann LeCun, has said the software driving today’s AI chatbots — known as large language models — falls short because “they can’t think before they act, can’t perform actions in the real world or learn through embodied experience and lack persistent memory and capacity for hierarchical planning.”

Some AI critics also warn that today’s models are “stochastic parrots,” remixing patterns of text without genuine understanding — a shorthand for doubts about whether statistical mimicry can ever amount to real intelligence. Researchers also still don’t fully understand how human intelligence works, let alone how to replicate or surpass it.

A July 2025 paper led by Margaret Mitchell, an AI researcher, argued that the ambiguity around AGI fuels hype and shaky benchmarks, with funding often steered by hype cycles or investor interests rather than research priorities. The authors argue that sweeping AGI claims should be treated skeptically unless tied to clear goals, rigorous testing and proven benefits.

Which companies are pursuing AGI?

OpenAI has made human-level AI its goal since 2015 and says its mission is to ensure AGI “benefits all of humanity.” Google DeepMind, for its part, has long said it is working toward AGI, publishing roadmaps and safety plans for that path.

After ChatGPT’s breakout success, competitors began adopting similar rhetoric. Amazon set up an AGI lab in San Francisco, led by a former OpenAI executive. Meta said its goal was to create AGI, then shifted to “personal superintelligence,” with Mark Zuckerberg casting AI as a tool to empower individuals.Rivals in China are following suit. Alibaba has made AGI its “primary objective.” DeepSeek says its goal is to “unravel the mystery of AGI with curiosity.” It’s pursuing more efficient training methods to build more powerful AI technology and compete with US firms.

How far along do companies say they are?

In September 2024, Altman said “systems that start to point to AGI are coming into view” and even suggested it’s “possible that we will have superintelligence in a few thousand days,” while acknowledging it could take longer.

A year later, however, OpenAI appeared to still be figuring out its path to that milestone. When OpenAI finally released its long-awaited GPT-5 AI model in August, it delivered improvements in coding, writing and mimicking the process of human reasoning. But to many users, it felt more like an incremental step forward than a leap.

In remarks around the launch, Altman conceded that “we’re still missing something quite important” to reach AGI, without clarifying what the missing piece might be.

But in September, OpenAI's Brockman declared that the company's latest model marks “a real shift in what kind of work people can delegate to AI and how it can empower them," bringing the company closer to AGI. “I think that there’s still lots of improvements to be made, but there is something significant here that I think is qualitatively improved."

Anthropic CEO Amodei has said “powerful AI” could plausibly arrive this year. More recently, while dismissing the AGI label as a “marketing term,” he predicted the next phase would look like “a country of geniuses in a data center,” something he believes is “quite likely” within two to three years.

By contrast, DeepMind co-founder Demis Hassabis has estimated that AGI is still five to 10 years away, calling shorter forecasts unrealistic. “There’s quite a long way, in my view, before we get to AGI,” he said last year, adding that the timelines only appear to be shrinking because “the definition of AGI is being watered down.”

For all the advances AI chatbots and so-called agents have made in the past three years, today’s best models still confidently make up facts and stumble on “conceptually simple” tasks that humans answer with ease, including questions an elementary school child could handle, even as they solve more sophisticated problems.

Benchmarks can be misleading as well. A system that aces standardized tests may stumble when the same questions are rephrased or when faced with messy real-world data. DeepMind stresses that AGI would need to be reliable across a diverse range of conditions.

(Updates article first published in October 2025 after latest OpenAI comments)