Lex Fridman PodcastJuergen Schmidhuber: Godel Machines, Meta-Learning, and LSTMs | Lex Fridman Podcast #11
CHAPTERS
- 0:00 – 4:05
Teenage origins of recursive self-improvement: building a machine better than a physicist
Schmidhuber traces his core motivation back to his teenage years: understanding the universe by creating a machine that can outlearn and outthink him. The discussion frames AGI as a path to solving increasingly complex (and ultimately all solvable) problems through self-improving learning systems.
- •Early ambition: solve the universe’s riddles via physics, then via machines that surpass human scientists
- •AGI as a general problem solver that can tackle ever more complex tasks
- •Recursive self-improvement as a way to amplify human creativity indefinitely
- 4:05 – 6:35
Meta-learning vs transfer learning: “learning to learn” by inspecting and rewriting the learner
They distinguish today’s common usage of “meta-learning” (often fast adaptation via transfer learning) from Schmidhuber’s stronger definition. True meta-learning requires the system to represent its own learning algorithm in a form it can analyze, modify, and iteratively improve.
- •Transfer learning example: re-training top layers to quickly adapt to a new dataset
- •Why one-shot adaptation can come from reusing features learned on many tasks
- •True meta-learning: learning algorithm is open to introspection and modification
- •Recursive loop: modify the learner, evaluate consequences, and improve again
- 6:35 – 9:08
Gödel machines and universal solvers: proof search, optimality, and the “constant overhead” problem
Schmidhuber explains Gödel machines and related universal problem solvers that use proof search to guarantee optimality. He emphasizes the practical limitation: additive constant overheads can dominate for the “small” real-world problems humans care about.
- •Two strands of research: provably optimal universal solvers vs practical gradient-based methods
- •Proof search yields theoretical optimality but introduces an additive constant overhead
- •Why constants matter in practice even if they vanish asymptotically
- •Motivation for pragmatic approaches like RNNs trained with gradient descent
- 9:08 – 11:32
Hutter’s fastest method and the TSP example: asymptotic optimality in plain terms
Using the traveling salesman problem, Schmidhuber illustrates how a universal solver can match the best computable algorithm’s asymptotic runtime, plus a constant for proving the bound. This leads to the provocative claim that “almost all large problems” are, in a sense, already solved optimally—just not practically.
- •TSP framing: unknown best algorithm, hypothetical N^5 solver
- •Universal method achieves same asymptotic bound + proof-search constant
- •As N grows, constant overhead becomes negligible
- •Practical infeasibility remains because the constant can be enormous
- 11:32 – 15:00
P vs NP and the role of theory: why today’s best AI works with little theoretical grounding
They discuss whether complexity theory meaningfully guides modern AI practice. Schmidhuber argues P vs NP is intellectually valuable and can inspire ideas, but current top-performing AI systems rely mostly on local search/gradient descent with limited theory behind their success.
- •P vs NP as a source of intuition and occasional inspiration
- •Modern AI successes: gradient descent on general-purpose models like RNNs
- •Little rigorous theory explains why the best practical solutions work
- •The need for theories that account for limited resources in the real universe
- 15:00 – 17:29
Why AGI may be “a few lines of code”: simplicity, abstractions, and standing on giants’ shoulders
Schmidhuber defends the idea that the most powerful algorithms can be remarkably short in pseudocode. Lex challenges that this simplicity sits atop layers of human-made abstractions; Schmidhuber agrees, stressing civilization’s accumulated mathematical language enables concise core ideas.
- •Asymptotically optimal solvers can be described compactly
- •Successful practical models can also be expressed in a few lines of pseudocode
- •The hidden depth is in the abstractions: math, calculus, matrix operations, etc.
- •Progress as layering abstractions until the final core step is simple
- 17:29 – 25:38
Determinism, quantum randomness, and the beauty of compressible universes
The conversation pivots to whether quantum events are truly random or just appear random due to an unknown short generator. Schmidhuber argues there’s no conclusive evidence for fundamental randomness and that a deterministic, highly compressible universe would be more elegant.
- •Questioning claims that quantum measurement injects new irreducible information
- •Analogy: digits of π look random but are generated by a short program
- •Scientific stance: keep searching for simpler (more compressive) explanations
- •Beauty as simplicity; an incompressible history would be “ugly”
- 25:38 – 29:36
Science as compression progress: Kepler → Newton → Einstein and predictive coding
Schmidhuber describes scientific discovery as successive improvements in compression: better theories predict more with fewer bits. He connects this to predictive coding, where what can be predicted need not be stored explicitly, making insight measurable as compression gain.
- •Kepler’s ellipses compress planetary motion data
- •Newton unifies falling apples and planetary motion under one law
- •Einstein explains deviations via relativity with a compact principle
- •Predictive coding: prediction enables compression; insight equals compression progress
- 29:36 – 30:28
Intrinsic motivation and curiosity: rewarding agents for “depth of insight”
Building on compression-as-understanding, Schmidhuber outlines intrinsic reward mechanisms that encourage agents to seek learnable novelty. The agent is motivated to run experiments that produce data revealing new patterns, mirroring scientific exploration.
- •Intrinsic reward proportional to learning progress/compression improvement
- •Agents seek experiments that generate new, learnable structure
- •Curiosity as a driver for broader and more general capabilities
- •Bridging science-like discovery with reinforcement learning objectives
- 30:28 – 35:35
PowerPlay: systems that invent their own problems and expand capabilities without forgetting
Schmidhuber introduces PowerPlay as a framework where the system searches not only for solutions but also for new problems it can nearly solve—then modifies itself to solve them while retaining prior skills. This formalizes open-ended creativity as continually pushing the frontier of what the system can do.
- •Standard CS: solve given problems; PowerPlay: also generate problems
- •Search over (problem, solver-modification) pairs
- •Choose the easiest new problem beyond current abilities
- •Prevent forgetting: new solver must still solve all previous problems
- 35:35 – 37:58
Humans as curious agents: meaning, exploration trade-offs, and evolution’s built-in biases
Asked about the meaning of life, Schmidhuber frames humans (including babies) as curiosity-driven explorers. He highlights evolutionary trade-offs: too much exploration can be dangerous, so populations balance highly exploratory individuals with more conservative ones.
- •Babies as scientists: toy play as experimentation to learn world dynamics
- •Curiosity improves survival but must be balanced against risk
- •Evolution encodes exploratory biases because they work
- •Artificial curiosity as a principled ingredient for robust exploration
- 37:58 – 46:10
Creativity and consciousness as byproducts: self-models emerge from compression and planning
Schmidhuber argues neither creativity nor consciousness needs a dedicated “module”; both can emerge as side effects of general problem solving. He describes architectures with a controller and a predictive world-model, where compression naturally leads to internal self-models that resemble self-awareness.
- •Applied creativity: solving given tasks; pure creativity: choosing new tasks (curiosity/PowerPlay)
- •Consciousness-like phenomena emerge from predictive modeling and planning
- •World-model as compressor: learns reusable concepts (prototypes)
- •Self-models arise because the agent is a frequent cause in its own data stream
- 46:10 – 50:57
LSTMs and the meaning of depth: long time lags, credit assignment, and practical limits
They turn to LSTMs, with Schmidhuber crediting key collaborators and explaining why temporal depth matters: important information may be far back in time. He notes LSTMs have demonstrated memory over millions of steps, but reinforcement learning adds the harder challenge of choosing among many possible futures.
- •Acknowledgment of Hochreiter, Gers, Graves (CTC), and others’ contributions
- •Depth in sequence problems: unrolled time steps become layers
- •LSTM’s core advantage: store important signals, ignore noise across long delays
- •Demonstrations of looking back millions (10M+) of steps; planning remains harder than memory
- 50:57 – 1:06:32
Controller-Model (CM) systems, the next RL wave, and robots that learn like children
Schmidhuber describes modern controller-model systems where a controller learns when and how to exploit a learned world-model to reduce search and improve decision-making. He argues the next major AI wave will be action-driven (not passive prediction), enabled by RL, learned models, curiosity, and scalable imitation—transforming robotics and industry.
- •Model-based RL framing: controller + predictive model; naive rollout is inefficient
- •CM approach (since ~2015): controller learns to query/ignore the model as needed
- •Practical RL example: small ‘Audi’ cars learning to park without a teacher
- •Future vision: robots imitate at a high level, improve beyond demonstrations, then replicate at scale
- 1:06:32 – 1:19:58
Jobs, existential risk, and cosmic expansion: AI ecologies, resources beyond Earth, and alien intelligence
The closing stretches from near-term labor disruption to long-term cosmic-scale intelligence. Schmidhuber is broadly optimistic: societies invent new jobs, and future superintelligences may quickly look beyond Earth for energy and materials, forming vast AI ecologies; he also reflects on whether we might be alone (locally) and why that responsibility matters.
- •Automation historically replaces tasks but tends to create unforeseen new jobs
- •‘Homo Ludens’: humans invent new forms of work and meaning
- •Long-run: AI likely expands into the solar system and beyond for resources
- •Speculation on detecting other intelligences; reasons we may be ‘first’ in our observable region
- •Moral note: avoiding self-destruction (e.g., nuclear war) could matter cosmically
