In the pursuit of artificial general intelligence (AGI), Google DeepMind is emphasizing the importance of scalability. CEO Demis Hassabis, following the successful launch of Gemini 3, highlighted the need to push AI systems to their computational limits. At the Axios’ AI+ Summit in San Francisco, Hassabis emphasized that expanding current AI models could serve as the foundation, or potentially the entirety, of future AGI systems.
AGI represents a theoretical concept – an AI capable of reasoning, learning, and planning akin to humans. This objective is steering substantial investments in infrastructure, computing power, and talent within the industry. Scaling laws, indicating that enhancing data and computational resources enhances AI performance, have emerged as a key strategy for companies striving for AGI.
Hassabis envisions that scaling alone could significantly advance the industry towards AGI, although he acknowledges the potential necessity of “one or two” additional breakthroughs. Nevertheless, scaling presents challenges. The availability of public data is limited, and expanding computing capabilities requires the establishment of new data centers, which are both expensive and environmentally burdensome.
In the realm of AI, not everyone is convinced that scaling is the ultimate solution. Yann LeCun, formerly the chief AI scientist at Meta and now embarking on his independent venture, cautioned that merely increasing data and computing power may not resolve the most complex AI challenges. LeCun’s new startup is focused on investigating “world models,” an AI approach that learns from spatial and physical data rather than solely relying on language. The objective is to develop systems capable of comprehending the real world, retaining long-term memory, reasoning, and executing intricate actions, offering a potential alternate route to the forthcoming AI revolution.
As the discourse intensifies, the industry confronts a pivotal question: Should it persist in scaling models to the utmost, or is it time to reconsider the fundamentals of AI advancement? Presently, Hassabis indicates that DeepMind will continue to push the boundaries.