Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
Lex Fridman (host), Jürgen Schmidhuber (guest), Lex Fridman (host)
In this episode of Lex Fridman Podcast, featuring Lex Fridman and Jürgen Schmidhuber, Juergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11 explores jurgen Schmidhuber on self-improving AI, curiosity, and universal intelligence Lex Fridman and Jürgen Schmidhuber discuss foundational ideas for building truly general, self-improving AI systems. Schmidhuber contrasts narrow, practical deep learning (like LSTMs and transfer learning) with his theoretical work on meta-learning, Gödel machines, and asymptotically optimal universal problem solvers. They explore curiosity, creativity, and compression as core principles of both human and machine intelligence, arguing that intelligence and even consciousness emerge as side effects of efficient problem solving and data compression. The conversation extends to evolution, the nature of physical reality, future robot learning, economic impact, and the possibility that humanity may be the first civilization poised to fill the universe with intelligence.
Jurgen Schmidhuber on self-improving AI, curiosity, and universal intelligence
Lex Fridman and Jürgen Schmidhuber discuss foundational ideas for building truly general, self-improving AI systems. Schmidhuber contrasts narrow, practical deep learning (like LSTMs and transfer learning) with his theoretical work on meta-learning, Gödel machines, and asymptotically optimal universal problem solvers. They explore curiosity, creativity, and compression as core principles of both human and machine intelligence, arguing that intelligence and even consciousness emerge as side effects of efficient problem solving and data compression. The conversation extends to evolution, the nature of physical reality, future robot learning, economic impact, and the possibility that humanity may be the first civilization poised to fill the universe with intelligence.
Key Takeaways
True meta-learning requires systems that can inspect and modify their own learning algorithms.
Unlike standard transfer learning, Schmidhuber’s notion of meta-learning opens the space of possible learning algorithms to the system itself, allowing recursive self-improvement where the AI learns not just tasks, but how to improve its own way of learning.
Get the full analysis with uListen AI
Theoretically optimal universal problem solvers exist, but their constant overhead makes them impractical for everyday tasks.
Methods like the Gödel machine and Marcus Hutter’s fastest problem solver are asymptotically optimal up to an additive constant, which becomes negligible for huge problems but is prohibitive for the smaller-scale problems humans typically care about, hence the dominance of heuristic methods like gradient descent.
Get the full analysis with uListen AI
Intelligence and scientific progress can be seen as improvements in data compression.
From Kepler’s ellipses to Newton and Einstein, better theories compress more observational data into shorter descriptions; Schmidhuber formalizes this as “compression progress” and ties it directly to the experience of insight, beauty, and scientific fun.
Get the full analysis with uListen AI
Curiosity and creativity can be formalized as intrinsic rewards for compression progress and for solving new, self-invented problems.
In his “artificial curiosity” and “power play” frameworks, agents are rewarded for discoveries that reduce description length or for inventing and solving the simplest problems just beyond their current competence, mirroring how human scientists generate and tackle their own questions.
Get the full analysis with uListen AI
Consciousness may emerge as a byproduct of self-modeling for prediction and control.
A recurrent network that learns to predict the world efficiently will also learn compact internal models of the agent itself (since the agent appears in all its data); when a controller uses this model to mentally simulate consequences of its own actions, the resulting self-referential processing closely resembles what we call consciousness.
Get the full analysis with uListen AI
Simple algorithms likely underlie both AGI and the physical universe.
Schmidhuber argues that the most powerful problem solvers and physical laws tend to have very short descriptions, suggesting that both AGI’s core code and even quantum phenomena might be generated by compact, deterministic programs rather than true randomness.
Get the full analysis with uListen AI
The next major AI wave will be model-based, curiosity-driven reinforcement learning in real robots, not just passive pattern recognition.
He foresees robots that learn like children—building their own world models from raw sensory data, using RL and intrinsic curiosity to explore, and learning skills (like assembly) from high-level human demonstration—transforming traditional industries and the nature of work.
Get the full analysis with uListen AI
Notable Quotes
“All of science is a history of compression progress.”
— Jürgen Schmidhuber
“True meta-learning is about having the learning algorithm itself open to introspection and modification.”
— Jürgen Schmidhuber
“We never have a program called creativity. It’s just a side effect of what our problem solvers do.”
— Jürgen Schmidhuber
“It would be awful and ugly if the universe needed an almost infinite number of extra bits to describe all these random events.”
— Jürgen Schmidhuber
“I’d be surprised if we humans were the last step in the evolution of the universe.”
— Jürgen Schmidhuber
Questions Answered in This Episode
How could we practically implement self-modifying learning algorithms while keeping them safe and verifiable?
Lex Fridman and Jürgen Schmidhuber discuss foundational ideas for building truly general, self-improving AI systems. ...
Get the full analysis with uListen AI
What metrics best capture “compression progress” or “depth of insight” in real-world AI systems?
Get the full analysis with uListen AI
How far can model-based reinforcement learning and curiosity alone take us toward human-level general intelligence?
Get the full analysis with uListen AI
If consciousness is just a byproduct of self-modeling, does that change how we should morally treat advanced AI systems?
Get the full analysis with uListen AI
What economic and educational changes are needed to adapt to a world where robots learn most skills through self-exploration and imitation?
Get the full analysis with uListen AI
Transcript Preview
The following is a conversation with Jurgen Schmidhuber. He's the co-director of atia Swiss AI lab and the co-creator of long short-term memory networks. LSTMs are used in billions of devices today for speech recognition, translation, and much more. Over 30 years, he has proposed a lot of interesting out of the box ideas on meta-learning, adversarial networks, computer vision, and even a formal theory of, quote, "Creativity, curiosity, and fun." This conversation is part of the MIT course on Artificial General Intelligence and the Artificial Intelligence Podcast. If you enjoy it, subscribe on YouTube, iTunes, or simply connect with me on Twitter @lexfridspelled F-R-I-D. And now here's my conversation with Jurgen Schmidhuber. Early on, you dreamed of AI systems that self-improve recursively. When was that dream born?
When I was a baby. No, that's not true.
(laughs)
When I was a teenager.
And what was the catalyst for that birth? What was the thing that first inspired you?
When I was a boy, I'm, I was thinking about what to do in my life and then I thought the most exciting thing is to solve the riddles of the universe and, and that means you have to become a physicist. However, then I realized that there's something even grander, you can try to build a machine that isn't really a machine any longer that learns to become a much better physicist than I could ever hope to be. And that's how I thought maybe I can multiply my tiny little bit of creativity into infinity.
But ultimately, that creativity will be multiplied to understand the universe around us? That's, that's the, the curiosity for that mystery that, that drove you?
Yes. Uh, so if you can build a machine that learns to solve more and more complex problems and more and more general problem solver, then you basically ha-have, um, solved all the problems. At least all the solvable problems.
So how do you think, what is the mechanism for that kind of general solver look like? Because obviously we don't quite yet have one or know how to build one, but we have ideas and you have had throughout your career several ideas about it. So how do you think about that mechanism?
So in the '80s, I thought about how to build this machine that learns to solve all these problems that I cannot solve myself and I thought it is clear it has to be a machine that not only learns to solve this problem here and this problem here, but it also has to learn to improve the learning algorithm itself.
Right.
So it has to have the learning algorithm in, um, representation that allows it to inspect it and modify it, such that it can come up with a better learning algorithm. So I call that meta-learning, learning to learn and recursive self-improvement, um, that is really the pinnacle of that where you then not only learn, um, how to improve on that problem and on that, but you also improve the way the machine improves, and you also improve the way it improves the way it improves itself. And that was my 1987 diploma thesis which was all about that, hierarchy of meta-learners that have no computational limits except for the well-known limits, uh, that Goedel identified in 1931, and, uh, for the limits are physics.
Install uListen to search the full transcript and get AI-powered insights
Get Full TranscriptGet more from every podcast
AI summaries, searchable transcripts, and fact-checking. Free forever.
Add to Chrome