Gemini 2.0 and the evolution of agentic AI with Oriol Vinyals

This interview podcast features Professor Hannah Fry interviewing Oriol Vinyals, Vice President of Drastic Research and co-tech lead of Gemini, about the evolution of AI agents. The conversation begins by revisiting Vinyals' previous work on AI agents capable of playing StarCraft, then delves into the advancements in large language models and multimodal AI. Vinyals explains the two-phase training process (pre-training/imitation learning and reinforcement learning) used to develop these agents, highlighting the challenges of scaling and achieving truly general AI. He discusses Gemini's new features, including agentic capabilities that allow the AI to interact with the real world through a browser or search engine, and explores the potential for AI to perform complex tasks like planning and reasoning. The discussion concludes by considering the possibility of achieving Artificial General Intelligence (AGI) and the role of scaling and specialized applications in reaching that goal.

Outlines

Sign in to continue reading, translating and more.

Continue

Google DeepMind: The Podcast

Introduction and Guest Introduction

What is Drastic Research?

Evolution of AI Agents: From StarCraft to Multimodal AI

The Two-Phase Training Process of AI Models

Architectural Differences and Advancements in AI Models

Achieving Agentic Behavior in AI: From Core to Digital Body

Diminishing Returns of Scaling and Alternative Approaches

Beyond Scaling: Improving AI Models Through Data and Algorithms

The Potential and Pitfalls of Self-Generated Data for Training

Reinforcement Learning and the Challenge of Defining "Good"

Building the Digital Body: Capabilities and Safety Considerations

Reasoning and Planning in AI Agents

Memory in AI: From Long-Term to Working Memory

Gemini 2.0 and New Agentic Capabilities

Approaching Artificial General Intelligence (AGI)

Conclusion: Generality and the Path to AGI

Gemini 2.0 and the evolution of agentic AI with Oriol Vinyals

Google DeepMind: The Podcast

00:07Introduction and Guest Introduction

Introduction and Guest Introduction

01:21What is Drastic Research?

What is Drastic Research?

01:43Evolution of AI Agents: From StarCraft to Multimodal AI

Evolution of AI Agents: From StarCraft to Multimodal AI

03:23The Two-Phase Training Process of AI Models

The Two-Phase Training Process of AI Models

08:54Architectural Differences and Advancements in AI Models

Architectural Differences and Advancements in AI Models

09:55Achieving Agentic Behavior in AI: From Core to Digital Body

Achieving Agentic Behavior in AI: From Core to Digital Body

13:39Diminishing Returns of Scaling and Alternative Approaches

Diminishing Returns of Scaling and Alternative Approaches

16:06Beyond Scaling: Improving AI Models Through Data and Algorithms

Beyond Scaling: Improving AI Models Through Data and Algorithms

20:01The Potential and Pitfalls of Self-Generated Data for Training

The Potential and Pitfalls of Self-Generated Data for Training

23:40Reinforcement Learning and the Challenge of Defining "Good"

Reinforcement Learning and the Challenge of Defining "Good"

27:26Building the Digital Body: Capabilities and Safety Considerations

Building the Digital Body: Capabilities and Safety Considerations

31:11Reasoning and Planning in AI Agents

Reasoning and Planning in AI Agents

34:15Memory in AI: From Long-Term to Working Memory

Memory in AI: From Long-Term to Working Memory

38:59Gemini 2.0 and New Agentic Capabilities

Gemini 2.0 and New Agentic Capabilities

43:44Approaching Artificial General Intelligence (AGI)

Approaching Artificial General Intelligence (AGI)

47:26Conclusion: Generality and the Path to AGI

Conclusion: Generality and the Path to AGI