Demis Hassabis on shipping momentum, better evals and world models

In this episode of "Release Notes," Logan Kilpatrick interviews Demis Hassabis, CEO of Google DeepMind, about the recent advancements and releases from DeepMind. They discuss DeepThink, Genie 3, and the partnership with Kaggle to launch Game Arena, highlighting the rapid progress in AI and the challenges of creating consistent and reliable AGI systems. Hassabis emphasizes the importance of building world models, the convergence of different models into an omni-model, and the role of games as benchmarks for AI development. They also explore the potential applications of Genie 3 in training AI agents, interactive entertainment, and understanding the nature of reality, as well as the significance of tool use and reasoning in AI systems.

Outlines

Part 1: Introduction and Core Concepts

Part 2: Benchmarking and Evaluation

Part 3: Tool Use and Model Evolution

Part 4: Future Outlook

Sign in to continue reading, translating and more.

Open full episode in Podwise

Google for Developers

Part 1: Introduction and Core Concepts

Introduction to DeepMind's Recent Progress and DeepThink

Reinforcement Learning, Jagged Intelligence, and Genie 3

World Models, Genie 3 Applications, and Simulation Theory

Part 2: Benchmarking and Evaluation

Jagged Intelligence, Game Arena, and the Need for New Benchmarks

Game Arena as a Benchmark and the Challenges of Non-Game Environments

Reward Functions, Meta-Learning, and Scaling Dimensions

Part 3: Tool Use and Model Evolution

Tool Use as a Scaling Dimension and the Evolution of Models

Building Applications and the Convergence of Models

Part 4: Future Outlook

Future Game Development and Closing Remarks

Demis Hassabis on shipping momentum, better evals and world models

Google for Developers

Part 1: Introduction and Core Concepts

00:00Introduction to DeepMind's Recent Progress and DeepThink

Introduction to DeepMind's Recent Progress and DeepThink

04:54Reinforcement Learning, Jagged Intelligence, and Genie 3

Reinforcement Learning, Jagged Intelligence, and Genie 3

07:50World Models, Genie 3 Applications, and Simulation Theory

World Models, Genie 3 Applications, and Simulation Theory

Part 2: Benchmarking and Evaluation

13:01Jagged Intelligence, Game Arena, and the Need for New Benchmarks

Jagged Intelligence, Game Arena, and the Need for New Benchmarks

16:05Game Arena as a Benchmark and the Challenges of Non-Game Environments

Game Arena as a Benchmark and the Challenges of Non-Game Environments

19:59Reward Functions, Meta-Learning, and Scaling Dimensions

Reward Functions, Meta-Learning, and Scaling Dimensions

Part 3: Tool Use and Model Evolution

22:36Tool Use as a Scaling Dimension and the Evolution of Models

Tool Use as a Scaling Dimension and the Evolution of Models

25:47Building Applications and the Convergence of Models

Building Applications and the Convergence of Models

Part 4: Future Outlook

29:26Future Game Development and Closing Remarks

Future Game Development and Closing Remarks