GDM’s Pushmeet Kohli on solving science's biggest challenges with AI

Key Concepts:

• Transformative Impact (Scientific, Commercial, Social)
• AlphaFold (Protein Structure Prediction)
• Alpha Evolve (Code Optimization Agent)
• Alpha Earth (Geospatial Model)
• SynthID (Watermarking System for GenAI Content)
• AI Co-scientist (Multi-agent System for Scientific Discovery)
• IMO (International Mathematical Olympiad)
• Alpha Geometry & Alpha Proof (Domain-Specific Math Models)
• Gemini DeepThink (General Math Solving Model)
• AGI (Artificial General Intelligence)
• Data Democratization
• API for Science

1. Overview of DeepMind's Science Initiatives

• Pushmi Kohli, head of Science and Strategic Initiatives at DeepMind, discusses the rapid pace of scientific advancement within the organization.
• DeepMind focuses on goals with transformative impact, not incremental improvements.
• The team aims to leverage AI progress to solve "the next impossible thing for the benefit of humanity."

2. Recent Alpha Science Launches

• Alpha Evolve: A Gemini-powered code optimization agent that has significantly improved data center optimization and sped up Gemini training.
  • Saved 0.7% of the entire compute fleet, resulting in substantial monetary savings.
  • Found state-of-the-art solutions for 75% of open math problems and surpassed existing solutions for 20%.
• Alpha Genome: A model for deciphering the human genome.
• Alpha Earth: A geospatial model that combines remote sensing data into a single semantic representation.
  • Propagates information about specific attributes of a location to reason about where else similar attributes might be found (e.g., species habitats).
  • Builds upon the concept of Google Earth by integrating and making accessible vast amounts of planetary information.

3. Framework for Problem Selection

• DeepMind uses a specific algorithm to determine which problems to tackle:
  1. Transformative Impact and Feasibility: The problem must have the potential for transformative scientific, commercial, or social impact and be realistically achievable.
  2. Difficulty: There should be a consensus that the problem is unlikely to be solved within the next 5-10 years.
  3. Accelerated Timeline: DeepMind believes it can solve the problem in half or one-third of the time compared to the community's expectations.

4. Impact Categories and Examples

• Scientific Impact: AlphaFold is the prime example, revolutionizing protein structure prediction.
  • Solved a fundamental problem in science with applications in drug discovery and understanding human health.
  • Won the Nobel Prize for Demis Hassabis and John Jumper.
  • The AlphaFold Protein Structure Database democratizes access to protein structures for researchers worldwide.
• Commercial Impact: Alpha Evolve demonstrates commercial impact through optimizing data centers and speeding up Gemini training.
• Social Impact: SynthID, a state-of-the-art watermarking system, addresses the risks associated with generative AI.
  • Ensures that users can distinguish between synthetically generated and naturally occurring content.
  • All GenAI content across Google modalities (text, images, videos) is watermarked with an imperceptible signal.

5. The Role of AGI

• DeepMind anticipates the development of increasingly powerful and general AI models.
• The team's focus is on determining how to effectively utilize AGI to solve specific problems and benefit humanity.
• The key question is: "What will we use it for?"

6. Collaboration Between Science and Gemini Teams

• Close collaboration exists between the Science and Gemini teams at DeepMind.
• Collaboration areas include:
  • Base architecture development for improved scientific task performance.
  • Evaluation metrics for assessing progress.
  • Data selection for training Gemini on specific domains (biology, chemistry, materials, code, cybersecurity).
  • Joint projects like the IMO project.

7. IMO Project: Alpha Geometry, Alpha Proof, and Gemini DeepThink

• The IMO project involved developing AI models to solve mathematical problems.
• Alpha Geometry: A model specifically focused on solving geometry problems.
• Alpha Proof: A model that searches for valid proofs using an underlying LLM and a domain-specific language called Lean.
  • Lean provides formal proofs, ensuring the correctness of solutions.
• Gemini DeepThink: A general math-solving model built on top of Gemini 2.5 Pro.
  • Achieved a gold medal at the IMO, surpassing the previous silver medal achieved with domain-specific models.
  • Solves problems specified in natural language (English) rather than specialized mathematical languages.
  • The success of Alpha Proof provided training data for Gemini by generating problems and valid solutions.

8. Generalization of Math Skills

• The extent to which math skills transfer to general domains is an ongoing research question.
• Empirical analysis and ablation studies are used to determine the impact of math-related training data on various abilities of AI models.
• Instruction following and explainability in math problem-solving may generalize to other language model capabilities.

9. Accessibility and Deployment

• DeepMind emphasizes making its breakthroughs accessible to the world.
• AlphaFold's protein structure predictions are available through an API and the AlphaFold Protein Structure Database.
• Alpha Genome has a custom UI for researchers to analyze the effects of mutations in the human genome.
• Gemini DeepThink is accessible through the Gemini app.

10. AI Co-scientist

• AI Co-scientist is a multi-agent system that simulates the scientific process.
• Gemini plays multiple roles: hypothesis generator, reviewer, and critique.
• The system generates and critiques ideas, ranks them, and edits them.
• It has shown promise in generating novel insights into important problems, such as antimicrobial resistance.
• An anecdotal example involves a scientist being surprised that Co-scientist generated a hypothesis that their team had been working on for years.

11. API for Science

• The concept of an API for science is explored, drawing parallels to the increasing accessibility of software development.
• The key challenge is the specification question: how to define scientific problems effectively.
• Building interfaces that can naturally capture what scientists are trying to do is crucial.
• Feedback from scientists is essential for developing efficient communication channels between AGI and researchers.

12. Conclusion

• DeepMind is committed to using AI to drive transformative scientific breakthroughs and benefit humanity.
• The organization's science initiatives span a wide range of domains, from protein structure prediction to code optimization and geospatial modeling.
• Collaboration, accessibility, and a focus on solving difficult problems are key principles guiding DeepMind's approach.