Jürgen Schmidhuber, TU Munich & IDSIA — 2008/2009 paper (arXiv:0812.4360): one principle — the drive for compression progress — explains subjective beauty, novelty, surprise, interestingness, attention, curiosity, creativity, art, science, music, and jokes.
The 1991 fix to the 1990 curiosity model: rewarding raw prediction error sends the agent to stare at noise forever, because noise always gives high error.
one should not focus on the errors of the predictor, but on its improvements.
Why it matters
The 1991 fix to the 1990 curiosity model: rewarding raw prediction error sends the agent to stare at noise forever, because noise always gives high error. Reward the derivative — the improvement — and the agent hunts learnable regularity instead.
Use when: Use when building any learning system, or judging your own attention: measure progress, not confusion.
The attention rule in one line: the curious agent flees both the solved and the hopeless, and hunts the learnable middle — things still incompressible but expected to become compressible through more learning.
It will get bored by things that already are subjectively compressible. It will also get bored by things that are currently incompressible but will apparently remain so
Why it matters
The attention rule in one line: the curious agent flees both the solved and the hopeless, and hunts the learnable middle — things still incompressible but expected to become compressible through more learning.
Use when: Use when deciding where to point effort: drop the mastered and the random, chase the compressible-unknown.
A punchline is a compression event: two familiar things snap together and the combined description shrinks.
comedians also tend to combine well-known concepts in a novel way such that the observer’s subjective description of the result is shorter than the sum of the lengths of the descriptions of the parts, due to some previously unnoticed regularity shared by the parts.
Why it matters
A punchline is a compression event: two familiar things snap together and the combined description shrinks. The laugh is the intrinsic reward signal for the breakthrough.
Use when: Use when writing or judging a joke: the laugh tracks the size of the snap, not the cleverness of the parts.
The paper's practical listening algorithm: skip the overplayed and the incomprehensible; hunt the edge where your learning is fastest.
He should try a song that is unfamiliar enough to contain somewhat unexpected harmonies or melodies or beats etc., but familiar enough to allow for quickly recognizing the presence of a new learnable regularity or compressibility in the sound stream.
Why it matters
The paper's practical listening algorithm: skip the overplayed and the incomprehensible; hunt the edge where your learning is fastest.
Use when: Use when picking what to read, watch, or learn next: aim at the steepest part of your own learning curve.
The paper's master equation.
This drive maximizes interestingness, the first derivative of subjective beauty or compressibility, that is, the steepness of the learning curve.
Why it matters
The paper's master equation. Interestingness is not a property of data — it is the rate of change of your compression of it. That reframes boredom, taste, and novelty as facts about the observer's trajectory, not about the object.
Use when: Use when judging whether something is boring: ask whether the data went flat, or whether your learning curve did.
The thesis in one sentence: interest is an event inside the observer at the moment of compression progress — not a quality sitting inside the data.
I argue that data becomes temporarily interesting by itself to some self-improving, but computationally limited, subjective observer once he learns to predict or compress the data in a better way, thus making it subjectively simpler and more beautiful.
Why it matters
The thesis in one sentence: interest is an event inside the observer at the moment of compression progress — not a quality sitting inside the data.
Use when: Use when the same thing bores you now that electrified you a year ago: you compressed it, and the derivative went to zero.
Kills the Shannon definition of surprise — mere unpredictability — and replaces it with learnability.
Curiosity is the desire to create or discover more non-random, nonarbitrary, regular data that is novel and surprising not in the traditional sense of Boltzmann and Shannon but in the sense that it allows for compression progress because its regularity was not yet known.
Why it matters
Kills the Shannon definition of surprise — mere unpredictability — and replaces it with learnability. Surprise that teaches is the only surprise that counts.
Use when: Use when designing anything meant to hold attention: the target is learnable regularity, not noise.
Beauty is a level; interestingness is a slope.
What’s beautiful is not necessarily interesting. A beautiful thing is interesting only as long as it is new, that is, as long as the algorithmic regularity that makes it simple has not yet been fully assimilated by the adaptive observer who is still learning to compress the data better.
Why it matters
Beauty is a level; interestingness is a slope. A fugue stays beautiful after you have fully compressed it — it just stops being interesting.
Use when: Use when separating 'this is good' from 'this still has something to teach me.'
Discovery is not a feeling — it is a measured event: a sudden large drop in the number of bits needed to encode everything you knew.
An unusually large compression breakthrough deserves the name discovery.
Why it matters
Discovery is not a feeling — it is a measured event: a sudden large drop in the number of bits needed to encode everything you knew. Newton's law is the paper's example: a short piece of code that compresses every falling apple ever seen.
Use when: Use when asking whether a finding is real: did it shrink the description of the world, or just add to it?
Collapses the two-cultures wall: the physicist's experiment and the painter's canvas are the same move — generate data governed by previously unknown regularities.
scientists are very much like artists. They actively select experiments in search for simple but new laws compressing the resulting observation history.
Why it matters
Collapses the two-cultures wall: the physicist's experiment and the painter's canvas are the same move — generate data governed by previously unknown regularities.
Use when: Use when anyone insists science and art run on different fuels.
Curiosity is not a luxury for agents with spare compute — it is the optimal prior when rewards are rare and time is finite.
the curiosity drive is such a general and generally useful concept for limited-resource RL in rare-reward environments that it should be prewired, as opposed to be learnt from scratch
Why it matters
Curiosity is not a luxury for agents with spare compute — it is the optimal prior when rewards are rare and time is finite. Build it in; do not wait for it to emerge.
Use when: Use when deciding what to build into a system, a team, or a practice rather than hoping it appears on its own.
Demystifies consciousness into an engineering move: the self is the most-used symbol in the agent's data history, so the compressor invents it.
There is one thing that is involved in all actions and sensory inputs of the agent, namely, the agent itself. To efficiently encode the entire data history, it will profit from creating some sort of internal symbol or code (e. g., a neural activity pattern) representing the agent itself. Whenever this representation is actively used, say, by activating the corresponding neurons through new incoming sensory inputs or otherwise, the agent could be called self-aware or conscious.
Why it matters
Demystifies consciousness into an engineering move: the self is the most-used symbol in the agent's data history, so the compressor invents it. No mystic value attached.
Use when: Use when anyone treats self-awareness as magic rather than as the cheapest code for the most recurring pattern in the data.
Why observers prefer faces similar to their own: the compressor builds its prototype from the most frequent input, which is you.
What they see every day in the mirror will influence their subjective prototype face, for simple reasons of coding efficiency.
Why it matters
Why observers prefer faces similar to their own: the compressor builds its prototype from the most frequent input, which is you. Taste as caching.
Use when: Use when someone treats taste as objective: it is the observer's compression history talking.
The twin-boredom theorem: white noise and a dark room are equally boring, for the same reason — both offer zero compression progress.
Neither the arbitrary nor the fully predictable is truly novel or surprising—only data with still unknown algorithmic regularities are
Why it matters
The twin-boredom theorem: white noise and a dark room are equally boring, for the same reason — both offer zero compression progress.
Use when: Use when someone mistakes randomness for novelty, or familiarity for depth.
The paper's creed: everything learnable rides on stored history.
the data is holy as it is the only basis of all that can be known about the world.
Why it matters
The paper's creed: everything learnable rides on stored history. Throwing away raw experience is throwing away discoveries you have not made yet.
Use when: Use when tempted to delete logs, notes, or records: storage is cheap, unrepeatable history is not.
