Key Concepts
- Large Language Models (LLMs) / Foundation Models: AI models like Gemini and Claude, trained on massive datasets, capable of generating text, translating languages, and performing various tasks.
- Artificial General Intelligence (AGI): A hypothetical level of AI that possesses human-level cognitive abilities, including continual learning, creativity, and long-term reasoning.
- Multimodal Understanding: The ability of an AI model to process and understand information from multiple modalities, such as text, images, and audio.
- TPUs (Tensor Processing Units): Custom-developed AI accelerator hardware by Google, designed for machine learning workloads.
- Continual Learning: The ability of an AI model to learn continuously from new data without forgetting previously learned information.
- AI Bubble: The potential for inflated valuations and unsustainable investment in AI startups.
The State of AI: A Conversation with Demis Hassabis (Google DeepMind)
Introduction & Context
This discussion, part of CNBC’s SquawkPod reports from the World Economic Forum in Davos 2026, features Demis Hassabis, CEO of Google DeepMind, interviewed by Andrew Ross Sorkin. Hassabis discusses the current state of AI, Google’s advancements with Gemini, the path to AGI, and the potential for an AI bubble. The interview highlights Google DeepMind’s position as a leading force in AI research and development, particularly following the success of AlphaGo and the recent Apple-Gemini partnership.
Current Capabilities & Limitations of LLMs
Hassabis acknowledges the rapid progress of LLMs like Gemini, noting their iterative improvements. However, he emphasizes that significant gaps remain before achieving full AGI. Specifically, current models lack:
- Continual Learning: The ability to learn and adapt continuously without catastrophic forgetting.
- True Creativity: The capacity for genuinely novel and original thought.
- Long-Term Planning & Reasoning: The ability to formulate and execute complex plans over extended periods.
He frames the question of reaching AGI as an “empirical question,” suggesting that scaling existing techniques may be sufficient, but acknowledges the possibility of needing “two or three new big breakthroughs.”
Gemini & Google’s AI Strategy
Hassabis attributes Google’s recent success with Gemini to a concerted effort to “corral together all of the assets” within Google and DeepMind, including their research teams, TPUs, and accumulated research over the past decade. Gemini 3 is currently “topping most of the leaderboards on most of the benchmarks.”
A key focus has been accelerating infrastructure and rewriting it to quickly integrate model quality into product services. This is already visible in Search, the Gemini app, and will expand to services like Gmail throughout the year.
The Apple Partnership & Market Validation
The partnership with Apple, where Gemini will power Siri’s intelligence, is described as a “massive vote of confidence” in the quality of Google’s models. Apple’s rigorous evaluation process resulted in Gemini being selected over other options. Hassabis notes that creating such models requires substantial resources, research teams, and investment, making a partnership a logical choice for Apple.
Differentiation in the AI Landscape
Contrary to concerns about LLMs converging into homogeneity, Hassabis believes models are becoming increasingly specialized. He cites Claude’s strength in coding and Gemini’s excellence in multimodal understanding (image-to-image generation) as examples. He anticipates this differentiation will increase, leading to wider gaps between models.
The AI Bubble & Investment Landscape
Hassabis acknowledges the possibility of an AI bubble, particularly concerning “hot startups” raising billions in seed funding with no existing product or technology. He considers this “a little bit frothy and perhaps unsustainable.” However, he also points to numerous “amazing use cases” and products already being deployed, indicating a substantial underlying value.
He expresses concern about the capital requirements for companies like OpenAI and Anthropic to continue their progress, but emphasizes Google DeepMind’s strong financial position and integration with existing products (Gmail, Chrome, Search) as a defensive advantage.
Hardware & Compute Power
Hassabis dismisses the idea that a technological revolution will drastically reduce the need for compute power, referencing the DeepSeek example as “a little bit overblown.” He notes that DeepSeek relied on Western models for fine-tuning and training. He believes that more compute is needed for training, serving, and exploring new ideas, while simultaneously emphasizing efforts to improve model efficiency.
He highlights Google’s “full stack” approach – owning both the AI lab and the hardware (TPUs) – as a significant advantage, allowing for efficient utilization of compute resources, even older generations of chips.
The Future of AI Models & Competitive Landscape
Hassabis predicts a future with potentially two to four major AI players, particularly in the enterprise space, but acknowledges the increasing difficulty of maintaining a leading position due to the “ferocious pace” of innovation. He reveals he works 100-hour weeks and believes leading labs are doing the same.
He believes the “defensive moat” around Google’s business lies in the quality of its models and their capabilities, as well as product features and the potential for personalization through memory and integration with existing Google services.
Impact on the Job Market
Hassabis believes it’s too early to definitively assess the impact of AI on the job market, noting potential minor effects on entry-level positions. He anticipates that AI will create “extraordinary new opportunities,” particularly for creators (artists, game designers) who can leverage these tools to enhance their productivity. He advises young people to become “unbelievably proficient” with AI tools to gain a competitive edge.
Persistent Memory & Model Switching
Hassabis acknowledges the portability of persistent memory between models (e.g., transferring data from OpenAI to Gemini), but maintains that model quality and capabilities will remain the primary drivers of user loyalty.
Conclusion
The interview paints a picture of a rapidly evolving AI landscape, where Google DeepMind is a key player. While acknowledging the limitations of current LLMs and the challenges ahead in achieving AGI, Hassabis expresses optimism about the future, emphasizing the importance of continued research, infrastructure development, and a focus on delivering tangible value to users. He also cautions against excessive hype and the potential for an AI bubble, while highlighting Google’s strong position to navigate the evolving market.
AI summaries can miss context or contain errors. Check important details against the original video.
