GPT-5.1 & the AI Frontier — Łukasz Kaiser (OpenAI, Transformer Co-Author)
Timestamps are as accurate as they can be but may be slightly off. We encourage you to listen to the full context.
Podcast Summary
In this fascinating conversation, Łukasz Kaiser, co-author of the groundbreaking "Attention Is All You Need" paper and current OpenAI research scientist, explains why the narrative of AI progress slowing down is fundamentally wrong. (02:29) From inside the labs, AI progress continues as a smooth exponential curve, driven by both the maturation of pre-training and the emergence of reasoning models—a paradigm shift that began just three years ago but is delivering extraordinary capabilities.
- Main Theme: The episode explores the dual paradigms powering modern AI: pre-training at scale and reinforcement learning-based reasoning models, revealing how these complement each other to create increasingly capable systems that can think, use tools, and solve complex problems across domains like mathematics and coding.
Speakers
Łukasz Kaiser
Łukasz Kaiser is a leading research scientist at OpenAI and one of the co-authors of the seminal "Attention Is All You Need" paper that introduced the Transformer architecture powering modern LLMs. He previously worked at Google Brain under Ray Kurzweil and Ilya Sutskever, bringing a background in theoretical computer science and mathematics from his academic work in Poland, Germany, and France.
Matt Turck
Matt Turck is Managing Director at FirstMark Capital and host of the MAD podcast. He focuses on data and AI investments and regularly interviews leading figures in the AI space about the latest developments in the field.
Key Takeaways
AI Progress Follows Moore's Law Pattern
Despite narratives of slowdown, AI progress continues as a smooth exponential increase in capabilities, similar to Moore's Law. (02:49) Just as Moore's Law persisted through multiple underlying technologies over decades, AI advancement continues through different paradigms—first transformers, now reasoning models. From inside the labs, there's never been reason to believe this trend isn't continuing. The perception of slowdown often comes from outside observers who miss the technical transitions happening beneath the surface improvements.
Reasoning Models Represent a New Paradigm
Reasoning models fundamentally differ from base LLMs by generating "thinking" tokens before providing answers, trained through reinforcement learning rather than just gradient descent. (11:48) This approach allows models to use tools, browse the web, and verify their work during the thinking process. The key insight is treating this thinking process as trainable through RL, particularly effective in verifiable domains like mathematics and coding where you can determine if answers are correct or incorrect.
Massive Low-Hanging Fruit Remains in AI Development
AI labs have enormous obvious improvements to implement, spanning engineering infrastructure, RL training optimization, and better data curation. (08:37) Much progress comes from fixing bugs in complex distributed systems, improving synthetic data generation, and enhancing multimodal capabilities that still lag behind text performance. These aren't mysterious breakthroughs but methodical engineering work that requires significant time and resources to implement properly.
Chain-of-Thought Emerges from Reinforcement Learning
Models learn to think step-by-step through RL training where they generate multiple reasoning attempts, with successful approaches reinforced. (20:28) This process teaches models to verify and correct their own mistakes—a crucial thinking strategy that emerges naturally from the training. The visible chain-of-thought users see is actually a cleaned summary; the full reasoning process is typically messier but more comprehensive.
Jagged Capabilities Define Modern AI
Today's frontier models exhibit "jagged" abilities—excelling at Mathematical Olympiad problems while failing simple puzzles a five-year-old can solve. (47:28) This reflects reasoning models' current limitation to science-based domains and weak multimodal reasoning. Kaiser demonstrates this with dot-counting puzzles from his daughter's math book that stump GPT-5.1, highlighting the need for better generalization and multimodal integration in future systems.
Statistics & Facts
- OpenAI has grown from a few dozen people when Kaiser joined Google Brain to serving over a billion ChatGPT users, creating massive GPU allocation challenges between training and inference. (34:30)
- The human brain contains approximately 100 trillion synapses, while current AI models haven't reached 100 trillion parameters yet, suggesting significant room for scaling. (35:25)
- Google Brain expanded from around 40 people when Kaiser joined to over 3,000-4,000 people across multiple offices when he left, illustrating the massive scaling of AI research teams. (29:35)
