In Silicon Valley, there is a discussion about how far scaling regulations can advance technology.
Demis Hassabis, CEO of Google DeepMind, has stated his position on the matter. The firm just published Gemini 3 to great acclaim.
We must maximize the scaling of present systems since it will, at the very least, be a critical component of the eventual AGI system, he stated at last week’s Axios AI+ Summit in San Francisco. It might refer to the complete AGI system.
Artificial general intelligence, or AGI, is a theorized kind of AI that is just as capable of reasoning as humans. It’s the objective that all of the top AI firms are vying for, driving enormous investments in people and infrastructure.
According to AI scaling rules, an AI model will become smarter the more data and computation it is given.
Scaling alone, according to Hassabis, would probably lead the industry to AGI, but he believes “one or two” further breakthroughs will be necessary.
The issue with scaling on its own is that the amount of data that is publicly available is limited, and increasing computing requires the construction of data centers, which is costly and environmentally damaging.
The fact that the AI firms behind the top large-language models are starting to show declining returns on their enormous efforts in scaling worries some AI observers as well.
Researchers like Yann LeCun, Meta’s top AI scientist who recently announced he was leaving to lead his own business, think the industry should take a different approach.
He stated at the National University of Singapore in April that the majority of intriguing issues scale really poorly. AI cannot simply be assumed to be smarter with additional data and processing power.
As an alternative to large-language models that focus on gathering spatial information instead of language-based data, LeCun is departing Meta to concentrate on creating world models.
According to what he said on LinkedIn in November, the business aims to bring about the next great breakthrough in AI: systems that comprehend the physical world, have persistent memory, can reason, and can plan intricate action sequences.
