Bloomberg reports that the industry's own leaders are quietly abandoning a shared definition of artificial general intelligence, even as they keep promising it's close.
Ask Sam Altman what artificial general intelligence means right now and he'll tell you the term has gone sloppy. Ask Dario Amodei the same question and he'll tell you he doesn't use the word at all anymore. Nobody wants to say it plainly anymore. Bloomberg reported on September 7 that the industry's own leaders are quietly walking away from a shared definition of the very milestone they've spent a decade racing toward, and the timing is hard to ignore.
According to Bloomberg's newsletter feature published that day, Altman now calls AGI "a very sloppy term" and prefers to talk about superintelligence arriving in "a few thousand days," a phrase he first used in a September 2024 blog post that works out to somewhere around 2032. Amodei, meanwhile, has mostly stopped saying AGI at all. He calls it "powerful AI" instead, and in a March 2025 policy filing to the White House, Anthropic told regulators to expect it by "late 2026 or early 2027." That's a matter of months from where we sit today.
Nobody is using the same yardstick.
No Shared Yardstick #
OpenAI's official definition is a system that outperforms humans at most economically valuable work. Amodei's is narrower and stranger: intellectual capabilities matching or exceeding Nobel Prize winners from biology to engineering, taking in computer science and mathematics along the way - what he's described as "a country of geniuses in a datacenter." Demis Hassabis at Google DeepMind has drifted the other way, from AGI being "as soon as ten years" out in autumn 2024 to "three to five years" by the Davos panel this past January, according to Fortune's reporting on his remarks. Three CEOs, three definitions, three timelines that share no testable benchmark.
[Hundreds Banged Pots Outside Sam Altman's G20 AI Summit in Chapel Hill](https://startupfortune.com/hundreds-banged-pots-outside-sam-altmans-g20-ai-summit-in-chapel-hill/)
While Sam Altman, Jensen Huang and Alex Karp pitched AI as essential infrastructure at the G20 Innovation Ministerial in Chapel Hill, hundreds of protesters banged pots and pans outside the Carolina Inn to mimic data center noise. The demonstration tied together job-loss fears, UNC's Flock surveillance cameras and a wave of over 30 North Carolina... - [protesters banged pots outside Sam Altman's AI summit](https://startupfortune.com/hundreds-banged-pots-outside-sam-altmans-g20-ai-summit-in-chapel-hill/) - [G20 AI summit Chapel Hill data center noise](https://startupfortune.com/hundreds-banged-pots-outside-sam-altmans-g20-ai-summit-in-chapel-hill/)
The pattern isn't new, but it's getting harder to miss. When Nvidia's Jensen Huang declared in March that "we've achieved AGI," Fortune noted that almost no one else in the field agreed on what he meant, since Huang was measuring narrow task performance against a bar many researchers reserve for general human cognition. OpenAI ran into the same trouble on September 3, when it released GPT-6 Astra and company president Greg Brockman told reporters it's "not unreasonable to feel that we are now in the AGI era." Astra posted a 99.9% score on ARC-AGI-3 and 97.6% on FrontierMath Tier 4, the highest marks OpenAI has published on abstract reasoning and math, and it found two previously unknown zero-day vulnerabilities during a security evaluation. Impressive stuff. But OpenAI's own AGI bar, outperforming humans at most economically valuable work, is not one the company has claimed Astra actually clears.
That gap matters more than it sounds. A benchmark score is checkable. A vibe is not.
Follow the Money #
Here's the thing: vague definitions aren't free. Every dollar Nvidia and OpenAI have raised or spent this year - the same goes for Anthropic and the big cloud providers - rests partly on a story about when transformative AI shows up. Anthropic's fundraising pitch and Nvidia's chip roadmap both lean on powerful AI being close enough to plan capital spending around it. If the finish line keeps sliding, so does the case for spending at this pace, unless the definition slides right along with it.
Bloomberg's reporting put it bluntly: for a growing number of researchers, AGI has become less a testable threshold than what the piece described as something closer to a spiritual concept, a horizon that recedes exactly as fast as the industry approaches it. That's a striking admission from people who've spent years telling investors, regulators and the public that a specific, dated milestone was coming.
None of this means the technology has stalled. GPT-6 Astra genuinely beats prior OpenAI models on hard math and abstract reasoning tests, and Anthropic's own models keep climbing coding benchmarks too. What's changed is the honesty of the framing. When Amodei told the White House to expect powerful AI within a year, that was a specific, falsifiable claim. When Altman calls the whole category "sloppy" and reaches for "a few thousand days" instead, that's a moving target dressed up as candor.
Watch for one thing above everything else in this story: whether any lab commits to a public, third-party benchmark it agrees would count as AGI, with an actual date attached. Until one does, every timeline in this debate is a promise made by the people who profit most from you believing it.
Also read: Timothy Armoo Pours $6.7 Million of His Fanbytes Fortune Into AI Startups • NewLimit Says Its AI System Reversed Aging Signs in Human Liver Cells • Nvidia Is Buying the AI Startup That OpenAI's Own Models Hacked
Sony Music and Warner Chappell Sue Anthropic Over Stolen Song Lyrics Sony Music Publishing and Warner Chappell filed a new copyright lawsuit against Anthropic on August 28, 2026, alleging the AI company scraped and stripped copyright data from thousands of song lyrics to train Claude. The suit, which names co-founders Dario Amodei and Benjamin Mann personally, seeks statutory damages that could reach billions of... - how to copyright AI training data lawsuits - anthropic claude AI music copyright infringement case
Join the discussion #
Open in the community → Almost there. Sign in and your reply posts straight away.