All Articles
AI5 min read20 September 2026

Every AI Company Says Its Model Is AGI Now. I Do Not Think They Mean The Same Thing.

TL;DR

Every few weeks someone announces AGI, and they do not seem to mean the same thing by it. Here is what the word is actually supposed to mean, the five levels nobody quotes correctly, and why nobody building this is going to slow down.

AIAGIOpinionPolicyAI Safety

Every few weeks now, someone announces AGI. A new model ships, and somewhere in the launch thread or the podcast appearance, the word gets used like it is settled. I keep waiting for the moment where it actually feels settled, and it never comes, because the companies saying it do not appear to agree on what they are claiming.

What AGI Is Actually Supposed To Mean

OpenAI's own definition is systems that outperform humans at most economically valuable work. That is a real bar, and by that bar nothing available today clears it. But OpenAI has also described five internal levels on the way there, and that framework is more useful than the word itself because it forces you to say which layer you actually mean.

  • 1.Chatbots, systems that hold a conversation and respond to what you type.
  • 2.Reasoners, systems that can work through a problem at something close to human level.
  • 3.Agents, systems that take multi step action on their own, using tools, with minimal supervision.
  • 4.Innovators, systems that produce genuinely new discoveries in science or mathematics, not just recombinations of what already exists.
  • 5.Organizations, systems that run the full scope of what an organisation does, coordinating, deciding, operating without ongoing human direction.
  • By most accounts we have cleared level two. Level three is where things get honest, because agents exist right now, I have built two of them myself this year, and the truthful description of the current state is that they work in a demo and fall over at scale. That is not a criticism of the people building them. It is just where the technology actually is, underneath the announcements.

    The Claims Do Not Match Each Other

    In March, Nvidia's Jensen Huang said on a podcast that he thought AGI had already been achieved. He was answering a specific question about whether an AI could start and grow a billion pound business, not making a general claim, but the clip travelled without that context. Sam Altman has said the opposite kind of thing in the same year, that OpenAI is not quite there yet, while also saying he expects an internal system he would call AGI before the end of the year.

    Around the same time both of those statements were being made, a new benchmark launched showing every frontier model, OpenAI's, Google's, Anthropic's, scoring under one percent on a set of tasks ordinary humans solve without difficulty. Both things are true at once. The models are extraordinary at some tasks and close to useless at others, and the industry has not agreed on which of those facts gets to define the word.

    Wonder Has A Shelf Life

    I remember the first time I connected an AI model to an app I was building and pointed my phone camera at my fridge. It named every item inside with something close to perfect accuracy. I sat there for a minute just looking at the screen. That was a genuine moment of wonder for me.

    It is not anymore. Not because the technology got worse, but because wonder does not last, no matter what causes it. The calculator did the same thing a generation earlier. It was extraordinary when it first existed, then it was just a tool, then every new feature bolted onto it was extraordinary again for a while, then that became normal too. AI is running through the exact same cycle, just much faster, and I think that speed is part of why the AGI announcements keep coming. The wonder fades so quickly now that companies need a bigger word every few months just to produce the same reaction the last announcement produced.

    The Part That Is Not About Marketing

    On the eighth of September, Jacob Coxon, who had worked at both OpenAI and Anthropic, resigned from Anthropic and wrote that the people actually building this technology privately believe it could kill everyone by the end of the decade. He was explicit that this was not a marketing stunt. Anthropic's own alignment lead, Evan Hubinger, backed him publicly and put a number on it, more than ten percent within the next decade, in his own estimate.

    I do not think everyone building this agrees with that number. But I do not think Coxon and Hubinger are lying about what they hear privately either, and that gap between what gets said in a boardroom and what gets said in a press release is the actual story, more than any single model launch is.

    Nobody Is Going To Slow Down

    A few days after Coxon's resignation went public, the response from the top of the American government was to call the safety concerns a hoax, and to say the only guardrail AI needs is a strong president. Whoever wins AI wins, was the actual phrase used, tied directly to staying ahead of China. I do not think that framing is going away, because once a government decides that AI capability is the same thing as national power, slowing down stops being a safety decision and starts looking like surrender.

    That is my honest read on where this goes. The people closest to the technology are frightened enough to resign over it, in public, on the record. The people with the authority to actually slow anything down have decided the competition matters more than the fear. Both of those things are happening at the same time, in the same month, and I do not think either side is going to move. The work will not stop. It will just keep moving forward under cover of a race nobody is willing to lose, whatever the people building it privately believe about where it ends.

    Have a follow-up?

    Explore more from the archive

    Just finished "Every AI Company Says Its Model Is AGI Now. I Do Not Think T…"? Ask anything related, I'll surface other articles that might help.

    Found this useful?

    Share it with someone who'd enjoy it.