Google AI chief says scaling isn’t enough for true breakthroughs in AI

DeepMind CEO Demis Hassabis discusses the future of Artificial Intelligence (AI), arguing that while advancements in computing power and data have been crucial, they’re not the sole key to unlocking true breakthroughs like Artificial General Intelligence (AGI) and instead Google AI looks beyond Scaling as

“Scale Only Gets You So Far“ Demis Hassabis, DeepMind CEO

The Rise of Large Language Models and the Race for Scale

The past year has witnessed significant developments in the field of AI, particularly with the emergence of powerful Large Language Models (LLMs) like ChatGPT and Google’s own Gemini. These models have captured the public imagination with their ability to generate human-quality text, translate languages, and answer questions in an informative way. However, Hassabis cautions against solely relying on “scaling” – throwing more data and computing power at the problem – as the solution to achieving AGI.

Beyond Scaling With Google AI: The Need for Fundamental Research and New Approaches – Beyond Scaling

While acknowledging the importance of scale, Hassabis emphasises the crucial role of fundamental research in propelling AI forward. He highlights Google’s commitment to this area, emphasizing its long history of pioneering machine learning techniques. He argues that achieving AGI will likely require “several more innovations” beyond just scaling existing techniques.

Exploring New Frontiers: Agent-based AI and Safe Development

One potential avenue for future breakthroughs lies in exploring new approaches like agent-based AI. This concept involves developing AI systems that can actively learn, plan, and take actions in the real world, going beyond the limitations of current “passive Q&A systems.” While acknowledging the potential benefits and increased utility of such systems, Hassabis also stresses the importance of developing robust safety measures. He advocates for the use of “hardened simulation sandboxes” to test these advanced AI agents before deploying them in the real world.

The Future of AI: A Collaboration Between Research and Development

In conclusion, Demis Hassabis’s interview offers valuable insights into the current state and future trajectory of AI development. While acknowledging the significance of advancements in scaling and computational power, he emphasises the need to go beyond scaling with Google AI, and for a balanced approach that prioritises fundamental research, explores new avenues like agent-based AI, and prioritises the safe development and deployment of increasingly powerful AI systems. This collaborative effort between research and development will be crucial in shaping the future of AI and ensuring its responsible advancement for the benefit of society.

You may aslo like:

• Google Gemini vs ChatGPT-4
• Google’s VP on the Future of Advertising
• What is Generative AI (AGI)?
• Or learn more about Large Language Models (LLMs) at Stanford Online.