Demis Hassabis on shipping momentum, better evals and world models

Key Concepts

• Deep Think: Thinking models that incorporate planning and reasoning capabilities.
• Genie 3: A world model capable of generating consistent and interactive simulated environments.
• World Model: An AI model that understands the physics, structure, and behaviors of the physical world.
• Omni Model: A converged model that combines the capabilities of separate models like Genie, Veo, and Gemini.
• Game Arena: A platform for benchmarking AI models through competitive gameplay.
• Jagged Intelligence: AI systems that excel in some areas but exhibit weaknesses in others.
• Tool Use: The ability of AI systems to leverage external tools (including other AI systems) to enhance their capabilities.
• RL (Reinforcement Learning): A type of machine learning where an agent learns to make decisions by interacting with an environment to maximize a reward.
• AGI (Artificial General Intelligence): A hypothetical level of AI that can perform any intellectual task that a human being can.

1. Unprecedented Momentum and Progress

• Google DeepMind is releasing new advancements at a rapid pace, almost daily.
• Recent releases include Deep Think, the IMO gold medal-winning model, and Genie 3, which has received significant positive attention.
• The current pace is a result of efforts built up over the past few years.

2. Deep Think and Thinking Models

• Deep Think models are inspired by DeepMind's earlier work on AlphaGo and AlphaZero, focusing on agent-based systems that can complete tasks with objectives.
• These models incorporate thinking, planning, and reasoning capabilities on top of powerful multimodal models.
• Deep Think enables parallel planning and refinement of thought processes, which is crucial for tasks like mathematics, coding, and scientific problem-solving.
• A version of the IMO gold model is available for Gemini app subscribers.

3. Deja Vu: Scaling RL and Data Bottlenecks

• DeepMind's early bet on RL, demonstrated by the Atari work, proved the potential of scaling these techniques.
• Similar data bottlenecks experienced with AlphaFold are now appearing in other domains, such as coding, highlighting the need for human expert data.

4. Jagged Intelligence and the Need for Better Benchmarks

• Current AI systems exhibit "jagged intelligence," excelling in some areas (e.g., IMO gold medal) but making simple mistakes in others (e.g., high school math).
• This indicates missing capabilities in reasoning, planning, and memory.
• Existing benchmarks are becoming saturated, necessitating new, harder, and broader benchmarks that assess capabilities like understanding world physics, intuitive physics, physical intelligence, and safety.

5. Genie 3: Building a World Model

• Genie 3 represents the convergence of multiple research branches and ideas.
• It aims to build a "world model" that understands the physics of the world, including physical structures, materials, liquids, and behaviors of living objects.
• A world model is crucial for AGI to operate in the physical world, enabling advancements in robotics and applications like Project Astra.
• Genie 3's ability to generate consistent worlds demonstrates a strong underlying model of how the world works.

6. Potential Uses of Genie 3

• Genie 3 is being used for internal training, such as training the SIMA (Simulated Agent) to play games within the generated environments.
• It has potential applications in interactive entertainment, enabling new types of games and entertainment experiences.
• Genie 3 raises questions about the nature of reality and physics, potentially offering insights into simulation theory.

7. Game Arena: A New Benchmark for AGI

• Game Arena is a partnership with Kaggle to benchmark AI models through competitive gameplay.
• It provides objective measures of performance (Elos, scores) and automatically scales with the capabilities of the systems.
• The platform will start with chess and expand to thousands of games, including computer games and board games.
• Eventually, AI systems may invent their own games, requiring other AI systems to learn them.
• Game Arena aims to drive progress by pitting the best models against each other and creating harder tests as they improve.

8. The Challenge of Evaluation in Non-Game Environments

• Most of life's problems are evaluation problems, but defining the source of truth in non-game environments is challenging.
• Humans have multi-objective functions that are continually weighted differently based on various factors.
• AI systems need to learn to interpret human users' goals and translate them into useful reward functions.
• Research is exploring metacognition and meta RL, where a system tries to determine the reward functions for another system to optimize against.

9. Tools as a New Scaling Dimension

• Tool use is a crucial capability for AI systems, enabling them to leverage external resources during thinking and planning.
• The question arises of what capabilities should be integrated into the main model versus used as tools.
• The decision depends on whether adding a capability to the main model helps or harms other capabilities.

10. The Model as a System

• Models are transitioning from being just weights to becoming entire systems, doing more out of the box.
• This requires product managers and designers to anticipate technological advancements and design products that can adapt to rapidly evolving models.
• The web ecosystem and app functionality may change as agents use these systems as tools.

11. The Future: Omni Models and AGI

• The goal is to create an "omni model" that combines the capabilities of separate models like Genie, Veo, and Gemini.
• This omni model should handle all different aspects to the same quality level as specialized models.
• The ultimate dream is to use these tools to create the greatest game ever after achieving AGI safely.

12. Token Milestone

• Google DeepMind has crossed the quadrillion token mark on a monthly basis.

Synthesis/Conclusion

The interview highlights Google DeepMind's rapid progress in AI, particularly with the development of thinking models, world models, and the exploration of new benchmarking methods. Key takeaways include the importance of reasoning and planning capabilities, the potential of simulated environments for training AI, the challenges of evaluating AI in real-world scenarios, and the vision of creating omni models that can handle diverse tasks. The discussion also emphasizes the need for better benchmarks to address the "jagged intelligence" of current AI systems and the transformative potential of tool use. Ultimately, the goal is to achieve AGI and leverage these advancements for scientific discovery and innovative applications.