Demis Hassabis: AI scaling must be pushed to the limit — it is the key to AGI

Debates continue in Silicon Valley over whether simply “inflating” models with more data and computing power will be enough to reach artificial general intelligence (AGI). Google DeepMind CEO Demis Hassabis, whose team has just introduced Gemini 3, has taken a firm stance: scaling is not just an instrument but the foundation of AI’s future.

According to Business Insider, at the Axios AI+ summit in San Francisco he stated that scaling current systems had to be pushed to the maximum, since he believed this would be a key component of any final AGI system and possibly even its core in its entirety.

Why Hassabis Is So Confident in Scaling

The scaling law is an empirical observation that the more parameters a model has and the more data and compute it receives, the better it performs. Gemini 3, Claude 3.5, Grok-3 — all the latest breakthroughs rely on this principle. Hassabis acknowledged that one or two additional breakthroughs would probably still be needed, but he emphasized that the foundation would remain enormous scale.

The problems are clear: public data will soon be insufficient, and building data centers consumes gigawatts of energy. But the head of DeepMind believes the effort is justified.

Is There an Alternative?

Not everyone agrees. Yann LeCun, who is leaving Meta to start his own company, has argued that most interesting tasks scale very poorly and that it is misguided to assume that more data and compute automatically result in a smarter AI. He instead focuses on “world models” that learn from spatial representations rather than text.

Other critics point to diminishing returns: each additional trillion parameters brings a smaller improvement in quality while costs grow exponentially.

In Brief

Demis Hassabis of DeepMind views scaling of current models as the primary — if not the only — path to AGI and urges that it be pursued to its limits. Yann LeCun and part of the research community are seeking alternatives, such as models that understand the physical world rather than simply processing terabytes of text. The race for general intelligence continues, and no one yet knows who is right: those building ever larger “engines” or those searching for an entirely different recipe.