Driven by Compression Progress

🧠 Free Recall ❓ Quiz 🔍 Cues 🔮 Simulation 🥷 ShadowBox 📐 Encode Cards

Main Ideas from Schmidhuber's "Driven by Compression Progress"
Schmidhuber's central claim is that a single simple principle—rewarding compression progress—explains a striking range of cognitive and aesthetic phenomena. Here are the key ideas:

The Core Principle
An intelligent agent should store its full sensory history, continually try to compress it better (find regularities/patterns), and reward itself whenever its compressor improves. The "curiosity reward" is proportional to the number of bits saved by a new, better compression of past data. A reinforcement learner then chooses actions that maximize this intrinsic reward, steering the agent toward experiences that yield further compression breakthroughs.

Redefining Beauty and Interestingness
Schmidhuber draws a sharp distinction between two concepts that are usually conflated:
Beauty of some data, relative to an observer, is proportional to how compressible it is given that observer's current knowledge. A face close to one's internal prototype, or a proof with a short description, feels beautiful because few bits are needed to encode it.
Interestingness is the first derivative of beauty over time—the steepness of the learning curve. Something stays interesting only while the observer is still learning to compress it better. Once fully assimilated, it becomes boring even if still "beautiful."

Rejecting the Shannon/Boltzmann Notion of Surprise
Traditional information theory says white noise is maximally "surprising." Schmidhuber rejects this: pure noise is boring because it can't be compressed further, just as a blank wall is boring because it's already fully predictable. True novelty lies between these extremes—in data with regularities the observer hasn't yet discovered.

Art, Science, Music, and Jokes as By-Products
Once you accept the principle, many things fall out naturally:
Scientists design experiments to find laws that compress observations (Newton's gravity compresses falling-object data; relativity compressed more).
Artists combine familiar elements in ways that reveal hidden regularities, making the combination more compressible than its parts.
Music lives in the sweet spot between too-familiar (already compressed) and too-random (incompressible). Schoenberg is harder for many listeners because its regularities are less accessible.
Jokes combine known concepts so the punchline reveals an unnoticed shared regularity—the same mechanism as learning to juggle and grinning at your own progress.
Consciousness is just the by-product of building an internal symbol for the agent itself, since the agent is involved in all its data.

The boundary between creator and audience blurs—both are running the same compression-progress loop.

No "Ideal Ratio" of Order to Complexity
Earlier information-theoretic aesthetics (Birkhoff, Bense, Moles) posited some objective ratio of expected to unexpected content. Schmidhuber argues there is no such universal ratio—what matters is the change in compressibility for a specific observer with specific prior knowledge and compression abilities. This is why taste is subjective and time-dependent.

Implementation History
The paper also surveys Schmidhuber's own work since 1990 implementing approximations of this idea:
1990: reward = prediction error (problem: rewards noise)
1991: reward = predictor improvement (fixes the noise problem)
1995: reward = KL-divergence between prior and posterior (later rediscovered as "Bayesian surprise")
1997: two modules play a zero-sum betting game over algorithmic experiments

He closes with appendix material on theoretically optimal but uncomputable versions (AIXI, Gödel machines) and the open problem of building practical curious agents within fixed computational budgets.

The unifying thread: a lot of what we call intelligence, taste, and creativity might reduce to one resource-bounded agent trying to make its model of the world shorter.

Examples

Web Design Aesthetics
The history of web design reads almost like a textbook illustration of "interestingness is the first derivative of beauty."

Skeuomorphism (roughly 2007-2013). When Apple shipped iOS with leather-stitched calendars, felt poker tables, and glossy buttons casting realistic shadows, this was genuinely interesting to users because their compressors hadn't yet built good models for "digital interface." The visual cues borrowed regularities from physical objects users already had compressed schemas for—a button looks pressable because it has a highlight on top and a shadow underneath, exactly like a real button. The compression progress came from mapping unfamiliar digital affordances onto familiar physical ones. It was beautiful in Schmidhuber's sense: short description (it's just a button) for an otherwise novel domain.

Then it became boring. Once everyone had years of experience with touchscreens, the leather stitching wasn't doing compression work anymore. The physical metaphor was redundant—you already knew it was tappable. The "regularity" had been fully assimilated. Worse, the visual noise of textures and gradients was now adding bits without adding meaning. So flat design (iOS 7, Material Design, Metro) felt like a revelation: it stripped away the now-uncompressible decoration and revealed a simpler underlying grammar. The learning curve shot up again.

Flat design then exhausted itself the same way. Once everyone could parse flat rectangles instantly, designers started adding back subtle shadows, depth, glassmorphism, neumorphism, bento grids, brutalism, "anti-design" with deliberately clashing elements. Each wave is a search for what's currently incompressible-but-learnable for an audience that has already absorbed the previous wave's regularities. Brutalism is interesting precisely because visitors have deeply internalized polished design—it violates an expectation strong enough to be worth violating.

The framework also explains why good designers chase trends but great ones don't: chasing trends means producing what's already compressed by your audience (boring), while genuine novelty means finding a regularity they haven't seen yet. And it explains the generational split—what feels fresh to a 50-year-old designer (who internalized 2005 web norms) feels stale to a 22-year-old whose baseline compressor was trained on TikTok.

Rap Music's Evolution
Rap is an even cleaner case because its formal vocabulary has expanded so visibly.

Early rap (Sugarhill Gang, early Run-DMC, ~1979-1985) had end-rhymes on the 4-beat, simple AABB schemes, mostly monosyllabic rhymes, breakbeat loops. To listeners in 1980 this was intensely interesting because the regularity—rhythmic speech locked to a sampled break—was itself the discovery. The compression progress was huge: you suddenly had a model for "spoken word as percussion." Nobody complained the rhymes were simple because the form was the novelty.

But the simple form got compressed fast. By the mid-80s, AABB end-rhymes on the 4 were so predictable that they generated almost no compression progress. Rakim is the canonical inflection point: he moved rhymes off the bar line, introduced internal rhymes, multisyllabic rhymes, and enjambment across bars. Suddenly there was a new regularity to discover—rhymes could land anywhere, not just at the end. The learning curve re-steepened.

Each subsequent era is a search for the next still-learnable regularity:
Nas, Wu-Tang (early 90s): dense internal assonance, slant rhymes, narrative compression where a single line carries cinematic detail.
Eminem, MF DOOM (late 90s–2000s): multisyllabic chains so long they become their own structural game ("orange/door hinge"-style payoffs DOOM stretched into entire verses).
Outkast, Kendrick: voice as instrument, flow switches mid-verse, meter changes—introducing regularities at the level of flow patterns rather than just rhyme.
Migos, trap (mid-2010s): triplet flows. This was the most striking recent inflection because it was almost a rhythmic compression discovery—you could subdivide the bar in a new way that was instantly recognizable yet hadn't been the default before.
Playboi Carti, "rage" rap, mumble rap: push toward texture and timbre over semantic content. Critics often call this "regression to simplicity," but in Schmidhuber's framework it's a shift in which dimension carries the regularity. Lyrical compression flattens; sonic/timbral compression opens up.

This last point matters: what looks like simplification can be sophistication in a different dimension. A listener whose compressor is tuned to lyrical density hears Carti as boring (no compression progress on the axis they're measuring). A listener whose compressor is tuned to vocal texture, ad-lib placement, and beat selection hears the same track as full of new regularities to learn. Same data, different observers, different derivatives.

Practical Takeaways

Know your audience's compressor, not just their preferences
Most creator advice focuses on what audiences like—their current taste. The framework says that's exactly the wrong target, because what they like now is what they've already compressed, which is sliding toward boring. The more useful question is: what have they internalized so deeply they don't notice it anymore? That invisible baseline is your raw material. Brutalist web design only works as a move against fluent polished design—it requires the audience to have absorbed the thing you're violating. If your audience hasn't compressed the convention, breaking it just reads as incompetence.
Practical version: spend time figuring out what your audience treats as default, not what they treat as good. The defaults are where the leverage is.

Aim for "surprising in retrospect, obvious in hindsight"
Schmidhuber's framework predicts that the most rewarding discoveries are ones where the audience thinks "of course" after seeing them—because that "of course" is the feeling of their compressor suddenly fitting the data with a shorter program. If the response is "huh, weird," you've produced noise (incompressible). If the response is "yeah, obviously," you've produced cliché (already compressed). The sweet spot is "I didn't see that coming and now I can't unsee it." Rakim moving the rhyme off the bar. Apple stripping skeuomorphism. The triplet flow.
This is a useful self-test. After you make something, ask: does this feel inevitable in retrospect? If yes, you've probably found a real regularity. If it just feels random, you may have mistaken novelty for progress.

Don't confuse "new to you" with "new to the field"
This is where a lot of creators get stuck. Your personal compression progress (learning a technique that's new to you) feels identical from the inside to collective compression progress (introducing a technique that's new to the field). Both are rewarding. But only the second produces work that's interesting to other people. The implication: deeply study your domain's history, not to copy it but to know what's already been compressed. Most "innovations" by underread creators are reinventions of regularities the field absorbed decades ago.

Work at the edge of your compressor, then translate
The flip side: you can only discover regularities your own compressor is sophisticated enough to find. This is why genuine frontiers are usually opened by people who are deeply fluent in the existing form—Rakim was already a virtuoso of conventional rap when he broke its rules; Jony Ive understood skeuomorphism intimately before flattening it. Mastery isn't optional; it's the prerequisite for knowing what to break.
So the practical sequence is: get fluent enough that the conventions feel boring to you, then trust that boredom as a signal. Your own loss of interest in the standard moves is often the earliest indication that a frontier is nearby.

This is gold. May 23, 2026 @ 11:20 AM

Shift dimensions when one gets crowded
When everyone is competing on the same axis, that axis is getting fully compressed by audiences and returns are diminishing. The Carti example matters here: when lyrical compression hit saturation, the frontier moved to timbre and texture. The creators who found that frontier first didn't out-compete the lyrical rappers on lyricism; they changed which dimension carried the regularity.

For any creator, the question worth asking periodically: which dimension of my craft is currently doing the compression work? And: which dimensions am I neglecting that audiences haven't learned to read yet? Film directors moving to longer takes when fast cutting saturated. Web designers moving to motion and micro-interactions when static layout exhausted. Novelists moving to structural play when prose-level virtuosity became table stakes.

Build for the second listen, not the first
Compression progress is a derivative—it requires time and repeated exposure. Work that's optimized purely for first-impression novelty often has no depth on return visits; the regularity was thin, and once seen, it's gone. Work that rewards repeat engagement has layered regularities: the audience compresses the surface quickly, then discovers a deeper pattern, then a deeper one. This is what people mean when they say great work "holds up."
Practically: when you finish something, ask whether there's anything left to discover on the third or tenth encounter. If everything's on the surface, you've built for novelty, not progress.

Segment by compressor sophistication
Different audiences have different compressors, and the framework predicts they'll have genuinely incompatible aesthetic responses—not because some have "better taste" but because their priors differ. A trained musician hears regularities in a jazz solo that a casual listener literally cannot perceive yet. This means:
You usually can't make work that's optimally interesting to both novices and experts. The novice needs the basic regularities exposed; the expert finds those boring.
Picking your audience is partly picking which compressor you're writing for. "Accessible" vs. "challenging" work isn't a quality difference, it's a target-compressor difference.
The most ambitious move is making work that rewards multiple compressor levels simultaneously—surface regularities for novices, deeper structural ones for experts. Pixar movies do this; so do the best pop songs.

Accept that your frontier will close
Anything you discover becomes part of the audience's compressor within a few years. The triplet flow was a frontier in 2013 and a cliché by 2018. This isn't failure—it's the mechanism working. The framework implies that sustained creative careers require continuing to move, because the location of the frontier moves whether you do or not. Creators who find one frontier and stay there get progressively less interesting through no fault of their own work; the audience compressed around them.

The harder lesson: the techniques that made you interesting are the same ones you'll need to abandon to stay interesting. Most creators can't do this, which is why most creative careers have a peak rather than a plateau.

The meta-point
What the framework really offers creators isn't a recipe—it's a diagnostic. When something you made lands, you can ask which regularity did the audience just compress? When it doesn't land, you can ask was this noise (incompressible) or cliché (already compressed)? That diagnostic is more useful than most aesthetic theories because it gives you a place to look when the work isn't working, instead of just "try harder" or "trust your instincts."

Unstarred

Dunk (exclude from study)

Driven by Compression Progress

Review Questions

Todos

Sensemaker