Jad
Tarifi

In December 2025, a Tokyo startup with fourteen people said it had done the thing the trillion-dollar labs only promise. Integral AI announced TOWA, which it described as the world's first AGI-capable model. The internet did what the internet does. Headlines split down the middle - breakthrough or bravado - and Jad Tarifi, the company's co-founder and CEO, walked straight into the crossfire and kept talking.

Tarifi runs Integral AI as both chief executive and Chief AI Scientist, which tells you where his heart is. The company builds energy-efficient, interactive world models - software that learns how reality behaves and lets digital and physical agents plan inside it. The grander framing on the company's own site is less engineering and more incantation: "Give humankind a true magic wand." The product names read like a pulp sci-fi index - TOWA, Genesis, Stream, and a robot chef called Savor.

What makes him worth a second look is the discipline under the dreaming. Ask him to define AGI and he refuses the usual fog. "We define AGI pragmatically: a system that can learn any skill autonomously, reliably, safely, and efficiently." No singularity poetry, no goalpost-moving - four adjectives you can argue with. He insists the architecture behind TOWA is not a large language model at all, and that it can scale toward general intelligence at something close to human-level energy efficiency. That last word, efficiency, is the part he keeps circling back to.

Start with the detail that explains the rest: he grew up in Lebanon during years marked by conflict. A childhood where the ground can shift overnight tends to produce one of two adults - the cynic or the builder. Tarifi became the builder, with a near-utopian streak about what technology owes people. The "magic wand" line isn't marketing varnish bolted on later. It's the whole reason he got into this.

The credentials are an unusual stack. An education that runs through the American University of Beirut and the University of Waterloo, a master's in quantum computing, and a PhD in computer science and AI from the University of Florida - the dissertation aimed squarely at brain-inspired, mathematically rigorous theories of intelligence. Quantum mechanics in one hand, neuroscience in the other. Most people pick a lane. He collected lanes.

Then came Google, and the part of the story that makes recruiters lean in. For roughly nine years he was inside Google Research, where he founded and led the company's first generative AI team and pushed on neocortex-inspired learning and getting models to learn from limited data. This was generative AI before the acronym was a dinner-table word, before the transformer reordered the field. He was in the room early.

And then he left. In 2021 he co-founded Integral AI - and made a choice that still surprises people. He didn't plant the flag in Silicon Valley. He planted it in Tokyo. The logic was almost stubbornly simple: if the destination is embodied intelligence, you build where the world's best robots already live. Japan's manufacturing and robotics depth, he wagered, beats a bigger Bay Area cheque.

The obvious question follows him into every interview: how does a small team in Tokyo expect to compete with labs spending billions on compute? His answer flips the premise. The giants, he argues, are winning a race he doesn't want to run. Pour enough money and electricity into a large language model and you get a bigger large language model. Tarifi's wager is that the path to general intelligence isn't more of the same - it's a different architecture entirely, one inspired by how a biological brain learns, and one that doesn't need a power plant to think.

That's why "energy-efficient" shows up in nearly every sentence he speaks about Integral AI. It isn't a sustainability footnote. It's the strategy. If intelligence can be made cheap enough to run at roughly human-level efficiency, the advantage of the deepest-pocketed lab shrinks, and the moat built from raw spending starts to leak. He frames Integral AI's whole reason for existing around that bet: build the interactive world models that let agents learn, plan, and act without the astronomical cost - and the incumbents' lead stops mattering as much.

The same instinct shows up in the work he did before the startup. At Google he chased learning from limited data - getting models to generalize without an ocean of examples - and neocortex-inspired approaches to learning. It's a through-line that connects the quantum-computing master's, the brain-inspired doctorate, the Google research, and the company. The thesis has been consistent for over a decade: intelligence is a problem of structure and efficiency, not just scale.