More and more tech leaders are hinting that we’re already very close to AGI — or have surpassed it.
Oriol Vinyals, VP of Research at Google DeepMind and one of the most influential figures in modern deep learning, has made a striking observation: that current AI models would almost certainly have been called AGI by anyone in the field just seven years ago — including himself.
“Even as someone in the field, close to these models and neural nets in general — if, seven years ago, I mean, I’m using a time that is clearly pre- all that happened with LLMs — if seven years ago, I had to experiment with a model that we have currently, would I have declared this is AGI? And I would say probably yes,” he said on a podcast.
He acknowledged that the goalposts have moved as the technology improved — an ever-moving definition, as he put it.
“It’s very impressive, the progress.” But Vinyals was careful not to let that observation collapse into a simple claim that AGI is here. The closer we get, he suggested, the more demanding our expectations become. “Just because now we’re seeing it closer, it is a good thing to be more ambitious about what it is that we’re building. But again, based on different definitions, or perhaps even the expectations we might have had about what AGI meant even only a few years ago, I would say that in some way, AGI is here.”
He then drew a distinction — between what the models can do today and what he personally wants to see before he’d call it AGI in the fullest sense.
“All I’m saying is: I don’t think it is here in the way I want to see it, but it is fairly close. And maybe this ability for the models to truly learn from experience is what is missing, in my mind. But everyone will have their own kind of test, or bias, onto what capability gaps they feel the models still have.”
Vinyals’ comments land in the middle of an intensifying debate at the top of the AI industry. Google DeepMind CEO Demis Hassabis has said that 2026 will be seen as the beginning of the singularity, describing the specific feeling of acceleration that comes from watching months of work compress into hours. OpenAI’s VP of Research Aidan Clark has hinted that AGI may have already arrived in some form. Marc Andreessen has claimed it was reached roughly three months ago with the latest frontier models.
What makes Vinyals’ framing distinct is its intellectual honesty about the definitional problem. AGI has never had a fixed, agreed-upon meaning — and as models have grown dramatically more capable, the definition has drifted to keep them just outside it. His point is that this drift is real, that it reflects genuine ambition, but that it also means the question “have we reached AGI?” cannot be answered without first asking: whose definition, and from when?
The one gap he names — the ability to truly learn from experience — is not a trivial one. Today’s frontier models are trained in massive offline runs and then largely fixed; they do not update continuously from what they encounter. That distinction matters enormously for any task requiring adaptation, memory, or genuine growth over time. It is also, notably, one of the hardest open problems in the field. The line between “fairly close” and “here” may come down to exactly that.