What’s Next for AI? OpenAI’s Łukasz Kaiser (Transformer Co-Author)

In The MAD Podcast, Matt Turck interviews Lukasz Kaiser, a key figure in modern AI and one of the inventors of the transformer architecture. They discuss the exponential progress of AI, the shift towards reasoning models, and the misconception that AI development is slowing down. Lukasz explains that while pre-training is still valuable, the new paradigm of reasoning models offers more significant gains for the same investment. He highlights the importance of reinforcement learning in training these models and addresses the challenges and opportunities in multimodal AI. The conversation also covers Lukasz's journey into AI research, the story behind the Transformer paper, the differences between research cultures at Google and OpenAI, and the future of pre-training and generalization in AI. They delve into the specifics of GPT-5.1, the role of post-training, and the balance between model capabilities and real-world applications, including the potential for AI in robotics and the evolving job market.

Outlines

Sign in to continue reading, translating and more.

Continue

The MAD Podcast with Matt Turck

The Exponential Growth of AI Capabilities

Reasoning Models as a New Paradigm in AI

The Rapid and Often Missed Progress in AI Functionality

The Education Gap and Low-Hanging Fruit in AI Improvement

Data Quality, Synthetic Data, and Multimodal Advancements

Deep Dive into Reasoning Models: How They Work and Why They Matter

Reinforcement Learning and Its Role in Training Reasoning Models

The Evolution of Reinforcement Learning and Its Future Generalization

Generalizing Reinforcement Learning and the Importance of Data Quality

Chain of Thoughts and Model Training

Lukasz Kaiser's Journey into AI Research

Joining Google and the Genesis of the Transformer Paper

The Novelty and Challenges of the Transformer Architecture

Transition from Google to OpenAI

Research Team Organization and GPU Access at OpenAI

The Future of Pre-training and the Impact of Economic Realities

Rediscovering Distillation and the Resurgence of Pre-training

Interpretability and the Black Box Nature of Complex AI Systems

Deep Dive on GPT-5.1: Reasoning, Synthetic Data, and Post-Training

Post-Training and Model Styles

Model Thinking Time and Mathematical Olympiad

Limitations of Reasoning Models and Multimodal Challenges

Generalization and the Future of AI

Architectural Changes and the Role of GPUs

Codex and Long-Running Workflows

Compaction and Security

The Future of Models and the Role of Humans

Interesting Research Areas and Robotics

Conclusion

What’s Next for AI? OpenAI’s Łukasz Kaiser (Transformer Co-Author)

The MAD Podcast with Matt Turck

00:00The Exponential Growth of AI Capabilities

The Exponential Growth of AI Capabilities

02:41Reasoning Models as a New Paradigm in AI

Reasoning Models as a New Paradigm in AI

05:22The Rapid and Often Missed Progress in AI Functionality

The Rapid and Often Missed Progress in AI Functionality

07:19The Education Gap and Low-Hanging Fruit in AI Improvement

The Education Gap and Low-Hanging Fruit in AI Improvement

09:31Data Quality, Synthetic Data, and Multimodal Advancements

Data Quality, Synthetic Data, and Multimodal Advancements

11:16Deep Dive into Reasoning Models: How They Work and Why They Matter

Deep Dive into Reasoning Models: How They Work and Why They Matter

13:18Reinforcement Learning and Its Role in Training Reasoning Models

Reinforcement Learning and Its Role in Training Reasoning Models

15:00The Evolution of Reinforcement Learning and Its Future Generalization

The Evolution of Reinforcement Learning and Its Future Generalization

17:02Generalizing Reinforcement Learning and the Importance of Data Quality

Generalizing Reinforcement Learning and the Importance of Data Quality

19:23Chain of Thoughts and Model Training

Chain of Thoughts and Model Training

21:39Lukasz Kaiser's Journey into AI Research

Lukasz Kaiser's Journey into AI Research

23:41Joining Google and the Genesis of the Transformer Paper

Joining Google and the Genesis of the Transformer Paper

26:08The Novelty and Challenges of the Transformer Architecture

The Novelty and Challenges of the Transformer Architecture

28:30Transition from Google to OpenAI

Transition from Google to OpenAI

30:22Research Team Organization and GPU Access at OpenAI

Research Team Organization and GPU Access at OpenAI

32:49The Future of Pre-training and the Impact of Economic Realities

The Future of Pre-training and the Impact of Economic Realities

35:38Rediscovering Distillation and the Resurgence of Pre-training

Rediscovering Distillation and the Resurgence of Pre-training

37:28Interpretability and the Black Box Nature of Complex AI Systems

Interpretability and the Black Box Nature of Complex AI Systems

39:42Deep Dive on GPT-5.1: Reasoning, Synthetic Data, and Post-Training

Deep Dive on GPT-5.1: Reasoning, Synthetic Data, and Post-Training

42:42Post-Training and Model Styles

Post-Training and Model Styles

45:12Model Thinking Time and Mathematical Olympiad

Model Thinking Time and Mathematical Olympiad

47:21Limitations of Reasoning Models and Multimodal Challenges

Limitations of Reasoning Models and Multimodal Challenges

50:14Generalization and the Future of AI

Generalization and the Future of AI

52:53Architectural Changes and the Role of GPUs

Architectural Changes and the Role of GPUs

55:01Codex and Long-Running Workflows

Codex and Long-Running Workflows

57:24Compaction and Security

Compaction and Security

59:35The Future of Models and the Role of Humans

The Future of Models and the Role of Humans

1:02:29Interesting Research Areas and Robotics

Interesting Research Areas and Robotics

1:04:54Conclusion

Conclusion