Integral AI

Here is a fact about the artificial intelligence business that is both obvious and slightly insane: nobody agrees on what the product is. Everyone agrees they are building toward "AGI," artificial general intelligence, and everyone agrees AGI would be worth an unfathomable amount of money, and almost nobody agrees on how you would know if you had one. This is a strange way to run a trillion-dollar industry. It is a bit like a gold rush where the prospectors have not settled on what gold looks like.

Integral AI, a startup of about 14 people with offices in Palo Alto and Tokyo, has a solution to this problem, which is to define AGI itself. In December 2025 the company announced that it had tested what it called "the world's first AGI-capable model," a system named TOWA. The announcement did the thing these announcements do: it got written up, it got the founders on podcasts, and it got a chorus of outside experts pointing out that there were no independently audited results and no agreed-upon definition of the thing being claimed. Both of these can be true at once. That is, roughly, the entire genre.

Integral's answer to "what is AGI" is unusually specific - and includes a criterion almost nobody else uses: energy.

What makes Integral more interesting than the average AGI press release is that it actually answered the question it was being criticized for dodging. The company proposed three tests. One: a system has to teach itself genuinely new skills, in domains it has never seen, without a pre-existing dataset and without a human standing over it. Two: it has to do this safely and reliably, without catastrophic mistakes or nasty side effects. Three - and this is the unusual one - it has to do it at human-level energy efficiency. The total energy cost of learning a skill should be no more than what a person would burn learning the same thing. Most of the industry treats energy as a bill you pay later. Integral treats it as part of the definition of intelligence.

You do not have to believe TOWA has passed these tests to find the framing clarifying. It is a way of saying: general intelligence is not a bigger chatbot. It is a system that learns fast, learns locally, and does not need a data center's worth of power to pick up a new trick. The company's stated mission, which sounds like marketing until you sit with it, is to "give humankind a true magic wand." The less magical translation is that Integral wants to build interactive world models and sell them as infrastructure - the intelligence layer underneath robots, factories, and software, the way electricity sits underneath everything else.

The people making this bet are not randos. Jad Tarifi, the co-founder, CEO, and chief AI scientist, started Google's first generative AI team back in 2011, which is to say he was working on this a full decade before it became the only thing anyone in tech wanted to talk about. His co-founder and CTO, Nima Asgharbeygi, spent years at Stanford building the kind of intelligent-agent architectures that learn and make decisions. The résumés are the reason investors like SoftBank's Deepcore and Samsung Next wrote checks. They are not, on their own, evidence that the AGI claim is true. Résumés rarely are.

So the honest way to hold Integral AI is with both hands. In one hand: a serious technical team, real investors, a coherent and even elegant thesis about what intelligence should cost. In the other: a very big claim, made by the company that benefits from it, without the independent verification that would settle the argument. The useful question is not "did they build AGI." The useful question is the one Integral, to its credit, actually put on the table - what would prove it - and then watching to see whether anyone, including Integral, can.