Nov 26, 2025 · 1h 5m · mad
What’s Next for AI? OpenAI’s Łukasz Kaiser (Transformer Co-Author)
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of The MAD Podcast, host Matt Turck interviews OpenAI Lead Research Scientist Łukasz Kaiser about the trajectory of artificial intelligence, the transition from pre-training compute scaling to reasoning models, the creation of the Transformer architecture, and the future of multimodal AI.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Matt holds 17.1% of the talking time here. How this is scored →
Matt as informed peer 3.0 Guest teaching 5.1 Guest disagreement 0.6 Matt pushing back 0.4
speaking balance: gold is Matt, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 40:20 Rejection of perceived leap from GPT-5 to 5.1
When Matt suggests GPT-5.1 represents a massive capability jump over GPT-5, Łukasz directly rejects the premise, stating 'I think less than you think' and explaining that pre-training was mostly about cost reduction.
Hardest push from Matt ▶ 53:17 Challenging transformer sufficiency with alternative architecturesMatt challenges the assumption that scaling transformers and reasoning is sufficient, explicitly pressing Łukasz on whether alternative paradigms like Yann LeCun's JEPA are required for true generalization.
Biggest teaching moment ▶ 48:34 Visual dot puzzle demonstrating frontier model limitationsŁukasz educates Matt on the jagged frontier of reasoning models by presenting a simple primary-school visual puzzle that stumps both Gemini 3 and GPT-5.1 despite their Olympiad-level math capabilities.
Matt holds his own ▶ 56:08 Detailed technical knowledge of Codex Max and context compactionMatt demonstrates deep technical fluency by asking a highly detailed question referencing GPT-5.1 Codex Max's long-running agentic execution and multi-context window compaction mechanisms.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Matt as informed peer | Guest teaching | Guest disagreement | Matt pushing back | Why |
|---|---|---|---|---|---|---|
| Debunking the AI Slowdown Narrative and GPU Scaling Laws | 3 | 4 | 1 | 1 | Matt asks an informed question citing recent model drops to challenge the AI slowdown narrative. Łukasz explains how pre-training scaling laws still hold log-linearly while the new reasoning paradigm offers superior compute-to-capability gains. | |
| Practical Capability Leap: Web Browsing and Practical Real-World LLMs | 1 | 5 | 1 | 0 | Łukasz details how reasoning models search the web dynamically (e.g., finding zoo hours) rather than hallucinating from outdated memory. Matt listens as Łukasz highlights the gap between public perception and actual current capabilities. | |
| Developer Adoption and the Shift in Coding Workflows | 3 | 4 | 1 | 1 | Matt asks Łukasz to detail obvious low-hanging fruit in frontier model development. Łukasz outlines non-obvious internal engineering bugs, data cleaning challenges, and multimodal improvements. | |
| Deep Dive into Reasoning Models and Chain of Thought | 2 | 6 | 0 | 0 | Matt prompts an educational breakdown of reasoning models. Łukasz explains how non-differentiable chain-of-thought tokens require reinforcement learning and verifiable rewards rather than simple next-token gradient descent. | |
| Reinforcement Learning Paradigms: From RLHF to Outcome Verifiers | 3 | 5 | 0 | 0 | Matt raises the distinction between pre-training and post-training RL. Łukasz explains the shift from easily hackable RLHF preference models to robust outcome verifiers in technical domains. | |
| Generalization of Reasoning and Future Visual Chain of Thought | 4 | 5 | 1 | 1 | Matt references OpenAI's GDPVal benchmark to ask about expanding RL across economic sectors. Łukasz emphasizes how messy raw internet pre-training data is and why visual reasoning remains undertrained. | |
| How Chain of Thought Operates and Summary Generation | 3 | 6 | 0 | 0 | Matt inquires whether the displayed chain of thought matches actual processing. Łukasz clarifies that user-facing thinking is a distinct summary of raw, messy model tokens and explains self-correction behavior. | |
| Łukasz Kaiser’s Academic Journey: From Mathematics to Google Brain | 2 | 3 | 0 | 0 | Matt asks about Łukasz's path into AI. Łukasz describes transitioning from theoretical computer science in Poland and Germany to Google, highlighting France's unique 10-year academic leave policy. | |
| Inside the Creation of the Transformer Architecture | 3 | 5 | 1 | 1 | Matt asks how the Transformer paper came together. Łukasz corrects the myth of all eight authors being in one room and recalls early pushback against building single models for multiple tasks. | |
| Moving from Google to OpenAI and Lab Cultures | 2 | 3 | 0 | 0 | Matt asks about Łukasz's move to OpenAI. Łukasz details Ilya Sutskever's repeated recruitments, COVID workplace dynamics, and scaling differences between Google Brain and early OpenAI. | |
| Organizational Structure and GPU Resource Allocation at OpenAI | 3 | 4 | 0 | 0 | Matt asks if researchers compete for GPU compute. Łukasz clarifies that resource distribution is primarily governed by technical requirements, with pre-training consuming the majority. | |
| The Resurgence of Pre-Training, Distillation, and Model Economics | 3 | 6 | 1 | 1 | Matt asks about the future of pre-training. Łukasz explains how consumer scale transformed model deployment economics, shifting focus to distillation and smaller, cheaper inference models. | |
| Interpretability and Model Inner Workings in Modern AI | 4 | 5 | 0 | 0 | Matt inquires whether modern hybrid models remain black boxes. Łukasz cites OpenAI's recent work on sparse weights while emphasizing fundamental limits in comprehending massive parallel systems. | |
| Evolution from GPT-4 to GPT-5.1 and Post-Training | 3 | 6 | 2 | 1 | Matt notes that GPT-5.1 feels like a massive leap over GPT-5. Łukasz reveals that underlying technical changes were smaller than perceived, with post-training refinement driving user experience. | |
| Post-Training Techniques for Tone and Persona Steering | 3 | 5 | 1 | 0 | Matt asks about tone customization in GPT-5.1. Łukasz explains post-training RL steering and discusses why model naming was decoupled from underlying technical pre-training runs. | |
| Controlling Thinking Time and the Jagged Nature of Reasoning | 3 | 7 | 1 | 0 | Matt asks how thinking time allocation works. Łukasz explains compute-scaling benefits while highlighting how reasoning models exhibit jagged capabilities that fail simple primary school puzzles. | |
| Visual Case Study: The Dot Counting Puzzle Benchmark | 3 | 7 | 1 | 1 | Łukasz presents a visual dot-counting puzzle where frontier models fail simple context reasoning. Matt asks what trips them up, and Łukasz points to undertrained multimodal reasoning. | |
| Alternative AI Architectures, ARC Prize, and AI Interns | 4 | 5 | 1 | 1 | Matt asks whether non-transformer architectures like Yann LeCun's JEPA are needed for true generalization. Łukasz evaluates ARC Prize progress and engineering constraints on experimental architectures. | |
| Agentic Coding with GPT-5.1 Codex Max and Context Compaction | 4 | 6 | 0 | 0 | Matt references the specs of GPT-5.1 Codex Max. Łukasz explains how quadratic attention memory limits long-running tasks and how context compaction enables multi-day agentic workflows. | |
| Future of Knowledge Work, Trust, and Economic Adaptation | 4 | 6 | 1 | 1 | Matt asks what value remains for product builders as general models expand. Łukasz uses the translation industry to argue that human trust and verification preserve demand for human oversight. | |
| Frontier Research Directions: Physical Robotics and Embodied AI | 3 | 4 | 0 | 0 | Matt asks about future research frontiers. Łukasz identifies general-data RL as his key focus and predicts embodied robotics will rapidly advance once multimodal reasoning matures. |
Statements from this episode (39)
Kaiser: AI progress has been a smooth exponential increase in capabilities
“Fundamentally, if you look at AI progress, it's been a very smooth exponential increase in capabilities.”
Kaiser: Reasoning models are the second major milestone after Transformers
“One point was, of course, the Transformers when it started, but the other point was reasoning models.”
Kaiser: OpenAI began working on reasoning models around three years ago
“So we started working on it maybe three years ago”
Kaiser: Pre-training scaling laws still hold across OpenAI and Google
“What scaling clause says is that your loss will log linearly decrease with your compute. We totally see that and clearly Google sees that and all other labs.”
Kaiser: Reasoning yields far greater AI capability gains per dollar than pre-training
“With the new paradigm of reasoning, you can get much more gains for the same amount of money because it's on this like lower and like, there are just discoveries to be made and these discoveries unlock insane capabilities.”
Kaiser: AI coding tools recently became how many programmers work
“But I think it's the recent few months when the transition happened from, you know, people using it sometimes, but rarely, to now basically this being how a lot of people work in coding.”
Kaiser: Multimodal AI capabilities lag behind text performance
“The multimodal part still lags behind the text part to a large extent.”
Łukasz Kaiser: Reasoning models require verifiable data, excelling in math and coding
“So currently, and current for at least the Most basic ways we use it currently, it needs to be fairly verifiable. So there is an, is your answer correct or not? You prepare data for that. You can do that in mathematics, coding very well. You can do this in sci…”
Łukasz Kaiser: Early RLHF was brittle but crucial for chatbot development
“So it was a bit of a brittle technique, but it was a bit of RL that was extremely crucial to making the models chat.”
Łukasz Kaiser: Next-gen reinforcement learning will operate on general data
“I do believe the era of tomorrow will be broader. It will work on general data and maybe then it will expand to like domains that, that go beyond where, where it shines today.”
Kaiser: Math reasoning in AI models transfers to generic web searching
“If you learn to think for math, you can, you will sometimes do some, you know, some strategies are the transfer very much like look up on the web and see what they say and use that information. So some of these things are very generic and they start to transfe…”
Kaiser: AI reasoning in visual domains is currently very undertrained
“I think, especially thinking in the visual domains is very under trained, I believe.”
Kaiser: ChatGPT uses a secondary model to summarize raw reasoning steps
“So in the current chat GPT, you will see a summary of the chain of thought on the side. So there is another model that takes the full chain of thought and shows you a summary because the full ones are usually not very nice to read.”
Kaiser: Reinforcement learning causes AI models to self-correct mistakes
“Even for math and coding, you start seeing that the models start correcting their own mistakes, right? Earlier, if the model made a mistake, it generally just tell you what it did and insist that the mistake was right or something like that. With the thinking,…”
Łukasz Kaiser says Ray Kurzweil was his first manager at Google
“I came to Ray Kurzweil's group. He was my first manager.”
Kaiser: The eight Transformer paper co-authors were never in one room
“I don't think all eight of us were ever in the same physical room.”
Kaiser: Making deep learning ideas work is harder than generating them
“In deep learning, people laugh that ideas are cheap. Making them work is, is the hard part.”
Kaiser: AI tech labs are more similar than people think
“I think in, in general, the tech Labs are more similar to each other than people think. There are some differences, but I think if I look at it from the world, you know, from the university in France, the difference between this university and any of the tech …”
Kaiser: Projects, not individual researchers, compete for GPU access at OpenAI
“I don't think it's so much people that compete. I think it's more projects that Compete for GPU access.”
Kaiser: Pre-training consumes the most GPUs of any AI development stage
“Currently, pre-training just uses the most GPUs of all the parts, so it needs the most GPUs, right?”
Kaiser: Pre-training science is plateauing, but compute scaling still improves loss
“Pre-training, as I said, I think it has reached this upper level of the S-curve in terms of science, but it can scale smoothly. Meaning if you put More compute. You will get better losses if you do things right, which is extremely hard, and that's valuable.”
Kaiser: OpenAI's frontier AI models do not yet have 100 trillion parameters
“Our models don't have a hundred trillion parameters yet.”
Kaiser: Reinforcement learning reasoning works better on larger pre-trained models
“Pre-training has always worked. And the beautiful thing is it even stacks with RL. So if you run this thinking RL process on top of a better model, it works even better. Than if you run it on top of a smaller model.”
Kaiser: Interpretability of large AI models faces fundamental complexity limits
“So the understanding of what the models are doing on a higher level has progressed a lot, but then it's still an understanding of what smaller models do, not the biggest ones. But it's not so much that these patterns don't apply to bigger models. They do. It's…”
Kaiser: Pre-training between GPT-4 and GPT-5 focused on reducing costs
“The pre-training part in that timeframe was mostly about making things cheaper. Not making things better.”
Kaiser: Model hallucinations are dramatically lower than two years ago
“There was these things called hallucinations. It's still with us to some extent, but dramatically less than two years ago.”
Kaiser: RL is a major component in post-training tone steering
“I don't work on post-training and it certainly has a lot of quirks, but I think the main part is, is indeed RL where you say, okay, is this response cynical? Is this response like that? And you say, okay, if you were told to be cynical, this is how you should …”
Kaiser: OpenAI model names are now detached from technical milestones
“Now the naming is by capability, right? GPT-Five is a capable model. 5.1 is a more capable model. Mini is the smaller model that's slightly less capable, but faster and cheaper. And the thinking models are the ones that do more research, right? In that sense, …”
Kaiser: Distillation lets OpenAI combine research projects without long pre-training runs
“With distillation, you have the ability to put a number of projects into one model. It's kind of nice that you don't need to wait on all of them to complete at the same time. You can try to periodically put this together, actually make sure that as a product i…”
Kaiser: Test-time compute increases AI capabilities faster than pre-training
“Using more tokens to think increases your capability, and it increases it, given the computation, way faster than pre-training, right?”
Kaiser: No frontier AI model can solve a specific first-grade math exercise
“I took one exercise from this math book and none of the frontier models is able to solve it.”
Kaiser: GPT-V Pro solves dot puzzles by running Python code loops
“The GPT-V Pro will run Python code to extract these dots from an image, and then it will count them in a loop.”
Łukasz Kaiser: AI models still struggle with multimodal and sequential reasoning
“The models are just, they're starting, like you see the first example they manage, so they've clearly made some progress, but they have not yet learned to do good reasoning in multimodal domains, and they have not yet learned to use one reasoning in context to…”
Łukasz Kaiser: AI pre-training expands stored knowledge rather than generalization
“Pre-training is a little different, right? Because it increases the data together with your increase in model size. So it doesn't necessarily increase generalization. It just uses more knowledge.”
Kaiser: Engineering complexity is the main bottleneck for experimental AI research
“The engineering part is the biggest bottleneck. I mean, GPUs are a bottleneck too, when you scale really up. But implementing something that's larger than one machine, it's an experimental research project, so you don't have a team to do that.”
Kaiser: OpenAI aims to create an AI intern by late 2026
“I think that's what OpenAI says is they say, you know, we say we'd like an AI intern by the end of next year.”
Kaiser: Unlimited model connection to the real world carries severe safety risks
“This thing, like how do models connect with the external world? It's a fundamentally very hard problem because, you know, when you connect in an unlimited way, you can break things in the real world.”