On September 6, Jensen Huang posted three words: “AGI has arrived.” Hours later, Gary Marcus answered that Huang had offered “no evidence and no definitions” and called it an attempt at “a takeover of a scientific question by corporate fiat.” They were referring to OpenAI’s GPT-6 Astra. Huang never defined AGI, so his claim that it had arrived on NVIDIA hardware sounded more like marketing than measurement. Marcus, meanwhile, helped develop a definition at agidefinition.ai but then shifted the debate to the 10 AGI requirements he names in a bet, several of which describe capabilities that look more like superintelligence than AGI.
This is the biggest mistake the independent parties keep making — that is, moving the goalposts of AGI out toward superintelligence instead of just calling it. Instead, we are calling it: AGI is here, but it’s not superintelligence; in fact, it is just barely competent.
We Wrote Our Criteria Down In August 2025
In The Quiet Roar Of Artificial General Intelligence, we defined AGI as software that can autonomously act in pursuit of goals across domains by learning new skills, collaborating with humans and machines, and building software tools. That’s three specific capabilities: self-learning, collaboration, and tool-building. We also said it would not arrive all at once as superintelligence. Instead, it is arriving in four stages: competent, independent, strategic, and perhaps superintelligent.
We set some rough timelines in the report, as well, with stage one landing anywhere from 2026 to 2030. We further said that stage one, competent AGI, will act effectively within a specific domain under close supervision, execute interrelated tasks over days, identify its own knowledge gaps (but not reliably fill them), and refine its actions using contextual memory and self-critique. Those words were carefully chosen as a forecast. And now it’s happening.
ARC-AGI-3 Went From Unsolvable To Solved In Six Months
Prior to Astra, the ARC-AGI tests — a series of questions that AI has historically struggled with that humans do not struggle with — was the Mount Everest of AI benchmarks. Frontier models were only scoring 1% on the third version of it. Astra scored 62.7% on ARC-AGI-3 using ARC Prize’s harness and wrote its own Python libraries midplay to solve unfamiliar games. Using OpenAI’s better harness that improved ASTRA’s memory, the model scored 99.9%. While Astra busted down the ARC-AGI-3 benchmark, it is just barely over the thresholds for two of our three AGI criteria.
So Where Are We? Competent, Not Independent
If we score Astra against Forrester’s stage one criteria, here is how we see it:
- Tool-building is here. Forrester’s stage one criteria says competent AGI combines existing tools into more complex ones. Astra skipped ahead and wrote new ones. Nobody specified those libraries. Astra decided it needed them.
- Self-learning is here for one session. Forrester’s stage one asks a system to spot its knowledge gaps and refine its actions using contextual memory and self-critique. That is exactly what happened in the ARC benchmark. But the learning dies with the run, and it needed a special memory-optimizing harness from OpenAI. This crosses our self-learning threshold but just barely.
- Collaboration is the thin leg. Forrester’s stage one asks only that a system clarify goals through back-and-forth that manages uncertainty. Today’s AI models already do that, if unreliably. They do not negotiate trade-offs, adapt to your preferences over months, or build the shared context a colleague does.
As AI gets deeper in Stage one and approaches stage two, these capabilities will improve dramatically — effective within a specific domain and under close supervision. That is the definition we published 13 months ago, and Astra now meets it. Rather than extend the AGI definition, as the industry has been doing for a few years now, we are, again, calling it: AGI is here, but it’s barely competent.
Change The Label, Change Your Plans
Put Astra, or Claude Fable, in the right harness and write “competent AGI” on the whiteboard instead of “frontier model.” Stage one of four arrived, in 2026, a year earlier than anybody thought remotely possible. Does that change what your next planning meeting is about? It should.
Forrester clients can read the full report, The Quiet Roar Of Artificial General Intelligence. Then ask your team one question: If independent AGI shows up next year, what do we do differently? Schedule a guidance session or inquiry with either of us, and we’ll help you get ready.
