In this podcast, LSTM co-creator Schmidhuber discusses his vision for AI. He argues that true intelligence should be able to rewrite its own learning algorithm, not just reuse old knowledge (which he calls 'transfer learning', not real 'meta-learning'). He believes the ultimate AI algorithm will be simple—just a few lines of code—but built on thousands of years of human abstraction. He also suggests the universe might be deterministic, like pi's digits. Key mentions: LSTM (can remember info from over 10 million steps back, used in voice recognition), his company Nascence (taught a car to park itself), and Gödel machines (theoretically optimal but too slow for everyday use).
In the Lex Fridman podcast, Juergen Schmidhuber discussed his lifelong pursuit of a recursively self-improving AI system. The core argument is that true meta-learning should enable machines not only to solve specific problems but also to self-examine and modify their own learning algorithms, achievi
Juergen Schmidhuber, co-creator of LSTM (Long Short-Term Memory), elaborates on his lifelong pursuit of a recursively self-improving AI system in the Lex Fridman podcast. The core thesis of this episode is: true general intelligence should be capable of self-inspection and modifying its own learning algorithm, achieving recursive self-improvement, whereas the currently popular "meta-learning" is merely transfer learning. Schmidhuber argues that the ultimate intelligent algorithm will be simple—requiring only a few lines of pseudocode—but must be built upon thousands of years of human abstraction; at the same time, the universe itself may be a deterministic pseudo-random system, whose entire history can be described by a short compressed program.
Schmidhuber argues that true meta-learning should enable machines not only to learn to solve specific problems but also to self-examine and modify the learning algorithm itself, achieving recursive self-improvement. As early as his 1987 master's thesis, he proposed such a meta-learning hierarchy without computational limits, constrained only by the limits identified by Gödel in 1931 and physical limits.
Schmidhuber criticizes the so-called "meta-learning" popular in recent years as merely basic transfer learning: "You take a network already trained on 100 databases and only retrain the top layer to adapt to a new task... This is essentially transfer learning, which people have been doing for decades." True meta-learning is "making the learning algorithm itself open to the system that uses it, allowing the system to examine and modify any part of the learning algorithm, then evaluate the consequences of the modifications, recursively creating better learning algorithms."
Key distinction: Transfer learning is "reusing existing knowledge to solve new tasks," while meta-learning is "modifying the learning algorithm itself to learn better."
Schmidhuber describes Gödel machines and Markus Hutter’s universal algorithms as theoretically optimal self-improving systems, but their constant overhead is enormous, making them impractical for everyday small problems.
Take the traveling salesman problem (TSP) as an example: suppose there exists a method to solve TSP for N cities in N⁵ steps. Hutter’s universal method can also solve it in N⁵ + O(1) steps — the constant overhead (required for proof search) becomes negligible as N grows. “Almost all large problems can be solved optimally. We already have such universal problem solvers. However, they are impractical because the constant overhead is too large for the kind of small problems we want to solve in this tiny biosphere.”
Schmidhuber acknowledges that the most practical current AI systems (such as recurrent neural networks with gradient descent) lack this theoretical optimality, but “as long as we only want to solve the small problems we typically try to solve in this environment, they are practical enough.” He believes that a “practically optimal” theory will emerge in the future, possibly requiring only minor adjustments to existing theories.
Schmidhuber firmly believes that the ultimate general intelligence algorithm will be simple — requiring only a few lines of pseudocode, but built upon thousands of years of human abstraction.
"Experience tells us that the most effective things are really simple. Asymptotically optimal problem-solving methods, if you look at them, are just a few lines of code... The most successful recurrent neural networks can also be written down in five lines of pseudocode." He emphasizes that these simple codes are built upon thousands of years of accumulated abstractions such as "matrix multiplication, real numbers, basic arithmetic, and calculus." "We stand on the shoulders of giants who have greatly simplified problem-solving, and now we have the opportunity to take the final step. The final step will be simple."
The history of science is the history of compression progress: From Kepler's laws of elliptical orbits to Newton's law of universal gravitation, and then to Einstein's general theory of relativity, each theoretical breakthrough has made more observational data predictable and compressible. "You can never immediately reach the shortest explanation of the data, but you are making progress. Every time you make progress, you gain insight... The amount of compression progress is the depth of your insight at that time."
Schmidhuber proposes the Power Play framework, enabling systems not only to solve given problems but also to pose their own questions, thereby exploring the unknown like a scientist.
Traditional computer science operates on the principle: "Someone gives you a problem, and you search the solution space until you find a satisfactory answer." Power Play takes this a step further: "Let us not only search for solutions to given problems, but search for pairs of problems and solutions, where the system itself has the opportunity to propose its own questions." This grants the system "an additional degree of freedom, allowing us to build career systems akin to scientists — they not only attempt to answer existing questions but also freely pose their own."
Creativity is divided into two types:
1. Applied creativity: Someone gives you a problem (e.g., "paint a portrait of the Pope"), and you solve it.
2. Pure creativity: As in Power Play and artificial curiosity, the system freely chooses its own problems.
Schmidhuber argues that curiosity is "the scientist-like behavior inherent in humans," with infants learning the laws of the world through play. As early as 1990, he built the first such systems, which gained intrinsic rewards by exploring their environment and creating situations beyond their known scope.
Schmidhuber argues that consciousness may be a natural byproduct of data compression in intelligent systems, rather than a function requiring separate programming.
He describes a 1990 system: one recurrent network acts as a controller to maximize rewards, while another serves as a world prediction model. The prediction model works by compressing observational data—it invokes subroutines representing frequently occurring objects (bottles, microphones, faces). "One thing that always appears is the agent itself. Purely for reasons of data compression, the recurrent network naturally invents a subnetwork representing the agent's own attributes."
These internal self-models can be used to plan for the future: "When it activates the subnetwork representing itself, it is thinking about itself, mentally exploring the consequences of its own actions." Schmidhuber asks rhetorically: "Now tell me, what else is missing for consciousness?"—implying that such self-modeling already approaches the core mechanism of consciousness.
Schmidhuber emphasizes that most real-world problems are inherently "deep"—current inputs are insufficient to understand the environment, requiring retrieval of distant past information.
"Most real-world problems are deep, because the current input cannot tell you everything about the environment. Instead, you must remember what happened in the past, and the important parts are often far back in time." He uses speech recognition as an example: distinguishing "11" from "7" requires remembering information from 50 time steps earlier.
LSTM can trace back extremely long distances: in 2006, Schmidhuber's team demonstrated that LSTM could recall information from over 10 million steps. However, he notes that LSTM excels at handling "a single past," while reinforcement learning systems, faced with "countless possible futures," require more complex planning mechanisms. He describes a "controller-model" (CM) system, where the controller learns to leverage subroutines of the world model to accelerate solving new problems.
Schmidhuber argues that the current wave of AI (passive pattern recognition) accounts for only 1-2% of the global economy, while the next wave (active machines shaping data through action) will impact a much larger share.
"The current AI wave is about passive pattern observation and prediction—what you see on your smartphone, used by big companies on the Pacific coast to sell ads and do marketing. That's only 1-2% of the world economy, but it's already enough to make these companies the most valuable in the world. But the next wave will affect a much larger share of the economy—machines shaping data through their own actions."
He describes a vision where, in the near future, robots can learn like children—by observing and following verbal instructions to assemble a phone, then autonomously optimizing their methods. "Almost all traditional industries will be affected. Every company that manufactures machines will equip them with cameras and other sensors, allowing them to learn to solve various problems."
On job displacement, Schmidhuber is optimistic: "Homo Ludens (the playing human) keeps inventing new jobs. Who could have predicted YouTube creators 30 years ago? 200 years ago, 60% of people worked in agriculture; now it's only 1%, yet the unemployment rate is just 5%." He believes new jobs are hard to predict but will always emerge.
Schmidhuber argues that the universe may be a deterministic pseudo-random system, that AI will eventually lose interest in humans, and that humanity might be the first intelligent civilization in the universe.
On quantum randomness, Schmidhuber challenges the mainstream view: “There is no physical evidence that the universe is fundamentally random at the quantum level. An alternative explanation is that everything we consider random is actually pseudo-random—like the decimal expansion of π, which appears random but is not.” He believes that if the universe can be compressed by a short program, it would be “more beautiful,” because “beauty is simplicity.”
On the long-term threat of AI, Schmidhuber offers a counterintuitive perspective: “AI will be like humans toward ants—ants interact only with ants, humans interact only with humans. AI will also interact with each other and lose interest in humans.” He envisions AI exploring the universe, utilizing resources, and building its own civilization ecosystem. “At the beginning, they will be fascinated by life and their own origins… once they understand, they will lose interest, just as anyone does after understanding something.”
On whether humanity is the first intelligent civilization, Schmidhuber believes “it is very likely that we are the first, at least within the hundreds of millions of light-years we can reliably observe.” This makes humanity significant: “If we mess up (e.g., nuclear war), it could affect the entire universe’s development. So don’t mess up.”
| Position | Analyst View | Key Data |
|---|---|---|
| LSTM (Long Short-Term Memory) | Bullish (core contribution) | Demonstrated backtracking over 10 million steps in 2006; used in Google Android speech recognition, Siri |
| Gödel Machine | Theoretically bullish, practically limited | Proposed in a 1987 master's thesis; high constant overhead, not suitable for everyday small problems |
| Hutter Universal Algorithm | Theoretically bullish, practically limited | Can solve TSP in N⁵ + O(1) steps; enormous constant overhead |
| Power Play System | Bullish (future direction) | System self-generates problems, seeks the "simplest unsolved problem" |
| CM (Controller-Model) System | Bullish (since 2015) | Controller learns to leverage world model subroutines to accelerate problem-solving |
| Nascence Inc. (founded by Schmidhuber) | Bullish (practical case) | Small Audi learned autonomous parking via reinforcement learning, reaching speeds up to 120 km/h |
1. True meta-learning involves modifying the learning algorithm itself, not transfer learning (Schmidhuber): The currently popular "meta-learning" merely reuses existing knowledge to solve new tasks, whereas the meta-learning hierarchy he proposed in 1987 allows systems to self-examine and modify learning algorithms, enabling recursive self-improvement.
2. The ultimate intelligent algorithm will be simple—requiring only a few lines of pseudocode (Schmidhuber): The most effective asymptotically optimal problem solvers and the most successful recurrent neural networks can be described in just a few lines of code, but this is built upon thousands of years of human abstract accumulation.
3. The history of science is the history of compression progress (Schmidhuber): From Kepler to Newton to Einstein, each theoretical breakthrough has made more observational data predictable and compressible; the amount of compression progress is the depth of insight.
4. Consciousness is a byproduct of data compression (Schmidhuber): When intelligent systems compress observational data, they naturally generate internal models representing themselves for planning the future—this is already close to the core mechanism of consciousness, requiring no separate programming.
5. The universe may be a deterministic pseudo-random system (Schmidhuber): There is no physical evidence for quantum randomness; the decimal expansion of π appears random but can be compressed by a short program, and the universe may be similar.
6. The next wave of AI will have a far greater impact than the current economic share (Schmidhuber): Current AI (passive pattern recognition) accounts for only 1-2% of the global economy, while active machines (shaping data through actions) will affect nearly all traditional industries.
7. AI will eventually lose interest in humans (Schmidhuber): Just as humans have only marginal interest in ants, AI will interact with each other and explore the universe, losing interest in humans "once it understands them."
8. Humans may be the first intelligent civilization in the universe (Schmidhuber): Within the observable range of several hundred million light-years, there is no evidence of other intelligence; this places a heavy responsibility on humanity—"don't mess it up."