Greg Brockman on OpenAI's Road to AGI

In this episode of the Latent Space podcast, Alessio and Swyx interview Greg Brockman from OpenAI about recent releases including GPT-5 and GPT-OSS. They discuss the evolution of reasoning in AI models, moving from offline pre-training to online learning loops, and the importance of reinforcement learning. Brockman emphasizes the need for more compute to unlock further advancements and the potential for general-purpose learning technology to transfer skills across domains, including biology and mathematics. The conversation covers the capabilities of GPT-5, its potential as an intellectual partner, and the importance of safety, security, and alignment with human values. They also touch on the practical engineering decisions behind GPT-OSS, the value of American open source, and the changing landscape of software engineering with AI.

Outlines

Part 1: GPT-5 Introduction and Reasoning Evolution

Part 2: Generalization and Real-World Applications

Part 3: Model Capabilities and Productization

Part 4: Efficiency, Architecture, and Open Source

Part 5: Engineering, Teamwork, and Future Outlook

Sign in to continue reading, translating and more.

Open full episode in Podwise

Latent Space

Part 1: GPT-5 Introduction and Reasoning Evolution

Introduction to GPT-5 and the Evolution of Reasoning at OpenAI

Reinforcement Learning, Online Learning, and the Value of Tokens

Compute as the Bottleneck and the Scaling of Reinforcement Learning

Part 2: Generalization and Real-World Applications

General-Purpose Learning and the Transferability of Skills

Compute Expenditure, Real-World Interaction, and Biological Neural Nets

Biological Language, Genetic Conditions, and the GPT-5 Era

Intellectual Leaps, Problem Difficulty, and Interactive Coding Applications

Part 3: Model Capabilities and Productization

Connecting Intelligence to Real-World Applications and Extracting Potential Energy

AI Productization, CodeX, and Agent Robustness

Model Spec, Psychohistory, and the Future of LLMs

Human Preferences, RL Rewards, and Model Personalization

Generalization, Model Switchers, and Adaptive Compute

Part 4: Efficiency, Architecture, and Open Source

Simplicity, Pricing, and Compute Limitations

Inference Efficiency, Self-Improving Agents, and Architectural Decisions

API Use vs. Single Machine Models and American Open Source

Part 5: Engineering, Teamwork, and Future Outlook

Engineering at OpenAI and the Impact of LLMs

Adapting Team Work and the Value of Engineers

The Big Picture and the State of AI Research

Prioritization and Exploring the Possibility Space

Angel Investing, Time Travel, and Advice for the Future

Greg Brockman on OpenAI's Road to AGI

Latent Space

Part 1: GPT-5 Introduction and Reasoning Evolution

00:04Introduction to GPT-5 and the Evolution of Reasoning at OpenAI

Introduction to GPT-5 and the Evolution of Reasoning at OpenAI

03:04Reinforcement Learning, Online Learning, and the Value of Tokens

Reinforcement Learning, Online Learning, and the Value of Tokens

06:45Compute as the Bottleneck and the Scaling of Reinforcement Learning

Compute as the Bottleneck and the Scaling of Reinforcement Learning

Part 2: Generalization and Real-World Applications

10:25General-Purpose Learning and the Transferability of Skills

General-Purpose Learning and the Transferability of Skills

13:57Compute Expenditure, Real-World Interaction, and Biological Neural Nets

Compute Expenditure, Real-World Interaction, and Biological Neural Nets

17:34Biological Language, Genetic Conditions, and the GPT-5 Era

Biological Language, Genetic Conditions, and the GPT-5 Era

20:51Intellectual Leaps, Problem Difficulty, and Interactive Coding Applications

Intellectual Leaps, Problem Difficulty, and Interactive Coding Applications

Part 3: Model Capabilities and Productization

24:21Connecting Intelligence to Real-World Applications and Extracting Potential Energy

Connecting Intelligence to Real-World Applications and Extracting Potential Energy

27:30AI Productization, CodeX, and Agent Robustness

AI Productization, CodeX, and Agent Robustness

31:19Model Spec, Psychohistory, and the Future of LLMs

Model Spec, Psychohistory, and the Future of LLMs

34:02Human Preferences, RL Rewards, and Model Personalization

Human Preferences, RL Rewards, and Model Personalization

37:21Generalization, Model Switchers, and Adaptive Compute

Generalization, Model Switchers, and Adaptive Compute

Part 4: Efficiency, Architecture, and Open Source

40:55Simplicity, Pricing, and Compute Limitations

Simplicity, Pricing, and Compute Limitations

44:04Inference Efficiency, Self-Improving Agents, and Architectural Decisions

Inference Efficiency, Self-Improving Agents, and Architectural Decisions

47:53API Use vs. Single Machine Models and American Open Source

API Use vs. Single Machine Models and American Open Source

Part 5: Engineering, Teamwork, and Future Outlook

51:32Engineering at OpenAI and the Impact of LLMs

Engineering at OpenAI and the Impact of LLMs

54:16Adapting Team Work and the Value of Engineers

Adapting Team Work and the Value of Engineers

57:24The Big Picture and the State of AI Research

The Big Picture and the State of AI Research

1:00:40Prioritization and Exploring the Possibility Space

Prioritization and Exploring the Possibility Space

1:04:04Angel Investing, Time Travel, and Advice for the Future

Angel Investing, Time Travel, and Advice for the Future