ET@Davos: Google DeepMind CEO Demis Hassabis on China's AI Leap & Limitations, the AGI Myth & More

Key Concepts

• AGI (Artificial General Intelligence): Hypothetical intelligence that matches or exceeds human cognitive abilities across a wide range of tasks.
• Transformers: A deep learning model architecture that has revolutionized natural language processing and is foundational to many modern AI systems.
• Deep Reinforcement Learning: A machine learning technique where an agent learns to make decisions by trial and error, receiving rewards or penalties for its actions.
• TPUs (Tensor Processing Units): Custom-designed hardware accelerators developed by Google specifically for machine learning workloads.
• AlphaFold: An AI system developed by DeepMind that predicts the 3D structure of proteins from their amino acid sequence.
• Gemini: Google’s latest and most capable multimodal AI model.
• World Models: AI systems that learn an internal representation of the world, allowing them to plan and reason about future events.
• Gemma: Google’s family of open-source AI models.
• NotebookLM: A Google tool utilizing AI for research, summarization, and exploration of new topics.

The Current State of AI & Google’s Position

Deis Hassabis, CEO of Google DeepMind, asserts that Google and DeepMind have been responsible for “almost all of the inventions that the modern AI industry relies on,” including transformers, deep reinforcement learning, and AlphaGo, over the past 10-15 years. While initially focused on research and open-source contributions (like with Transformers and AlphaFold, which is now used by over three million researchers), the landscape has shifted. The current era is characterized by significant private investment in large AI models, necessitating a degree of proprietary protection of architectural “secrets.” This change occurred because other companies were leveraging Google’s published work without reciprocal contribution, creating an unsustainable asymmetry.

Hassabis highlights a recent shift within Google, moving from a perceived position of “catching up” to being “at the forefront” of AI. This was achieved by translating research into faster product deployment, adopting a “startup mentality,” and integrating research directly into existing Google products. Gemini is cited as a key example of this success, demonstrating leading overall capabilities as validated by Apple’s rigorous evaluation process. The deal with Apple is viewed as a significant endorsement of Google’s progress.

Open Source vs. Proprietary Development

Hassabis expresses a strong belief in “open science and open source” as drivers of rapid scientific progress, citing AlphaFold as a successful example of this approach. However, he acknowledges the necessity of protecting some architectural details of current large models due to the substantial investment involved. He reiterates the issue of companies benefiting from Google’s open-source contributions without reciprocation, which prompted a shift towards a more guarded approach. Google continues to support open-source ecosystems with models like Gemma, focusing on smaller, accessible models.

Tools & Internal Usage of AI at Google

Hassabis personally utilizes NotebookLM, a Google tool, for research summarization and exploration. He emphasizes the significant leap in capability with Gemini 3, finding its deep research feature particularly useful in daily life. He anticipates full integration of these AI capabilities into Google Workspace, with a beta version of an AI inbox already launched, with the ultimate goal of automating email management.

Global AI Landscape & India’s Approach

Regarding the global AI landscape, Hassabis believes China is rapidly closing the gap with the US, potentially being only six to twelve months behind. However, he stresses that China has yet to demonstrate “innovating beyond the frontier” – creating the next foundational breakthroughs like Transformers or AlphaGo.

For countries like India, Hassabis advises focusing on applying existing AI technologies to revolutionize industries and foster startups, rather than attempting to build frontier models from scratch. He suggests leveraging the readily available models from established providers and focusing on capability overhangs and exploring existing technology’s full potential.

Quantum Computing & AGI

Hassabis acknowledges Google’s significant investment in quantum computing and its collaboration with the quantum computing team led by Hartmut Neven. Currently, AI is assisting quantum computing, particularly in quantum error correction. He believes that while quantum computers may eventually accelerate AI algorithm training, AGI is likely to arrive before fully mature quantum computers. He personally believes classical computing has more potential than currently appreciated, citing AlphaFold’s success in approximating solutions to problems governed by quantum effects.

The Path to AGI & Future Implications

Hassabis estimates AGI could be achieved within the next 5-10 years. He defines AGI as the ability to generate new scientific theories, not just prove existing ones. He envisions a future of “radical abundance” driven by AGI, with breakthroughs in medicine and other fields. He acknowledges Yan LeCun’s skepticism about the current LLM approach to AGI, agreeing that one or two breakthroughs may still be needed, potentially involving concepts like “world models.”

Hassabis and Sundar Pichai regularly discuss AI integration across Google’s products, viewing Google DeepMind as the “engine room” creating the core AI models (Gemini, world models, video models, image models) that are then deployed across Google’s services. This integration is accelerating, with a strong focus on rapid deployment.

Societal Impacts: Jobs & Energy

Addressing concerns about societal impact, Hassabis acknowledges potential job disruption but anticipates the creation of new, higher-level jobs as has historically occurred with technological advancements. Regarding energy consumption, he notes that AI models are becoming more efficient (10x annually), but demand continues to rise as the pursuit of AGI intensifies. He believes AI will ultimately contribute to energy solutions through new materials, grid optimization, and the development of technologies like fusion (where Google is collaborating with Commonwealth Fusion).

Balancing Science & Leadership

Hassabis describes himself as a scientist first and foremost, but also possessing an entrepreneurial spirit. He finds equal enjoyment in both research and leading a large organization, viewing his role at Google DeepMind as a unique opportunity to pursue both passions and impact billions of lives. He still finds time for chess, occasionally playing online to unwind.

Conclusion

Deis Hassabis presents a compelling vision of Google’s central role in the ongoing AI revolution. He emphasizes the company’s historical contributions, its recent advancements with Gemini and TPUs, and its commitment to both open-source initiatives and proprietary development. He articulates a clear path towards AGI, acknowledging the challenges and potential societal impacts while remaining optimistic about the transformative potential of AI to address global challenges and usher in an era of abundance. The interview highlights the complex interplay between scientific innovation, commercial strategy, and societal responsibility in the rapidly evolving field of artificial intelligence.