Teaching with the complete image space
The Pixel Theory explorer is a free, single-file interactive that makes several famously abstract ideas physical: combinatorial explosion, the pigeonhole principle, Kolmogorov complexity, sampling from enormous spaces, digital signatures, and what AI image generators actually do. It needs no accounts and no installation, runs entirely in the browser, and sends no data anywhere. It works offline from a saved copy.
The premise takes one minute to state: an RGB pixel has 16,777,216 possible values, so a 1080×1080 image grid has 16,777,2161,166,400 possible states — a finite number with 8.4 million digits, containing every photograph that has ever been or could ever be taken at that resolution. Every section of the site is an instrument for feeling one consequence of that fact.
Concept map
| Section | Concept | The question to pose |
|---|---|---|
| 01 · The Space | Exponents, combinatorics, scientific notation | Why does adding one row and one column multiply the count by 1015,612 instead of adding to it? |
| 01 · Futility meter | Orders of magnitude | Every atom in the universe computing since the Big Bang covers ~10−8,426,800 of the space. Why is this not an engineering problem? |
| 02 · Frame Viewer | Place value, number bases, bijections | Press +1 and find the pixel that changed. Why that pixel? What is this in base 256? |
| 03 · Static Channel | Probability, uniform sampling | Why is a random frame always noise? What fraction of frames look like anything? |
| 04 · Address a Photograph | Functions and inverses | Encoding and decoding are exact inverses — so did the camera create the image, or find it? |
| 05 · Islands of Meaning | Redundancy, error tolerance | Why is a photo still recognizable with 30% of its pixels destroyed, and what does that say about information in images? |
| 07 · Coordinates | Pigeonhole principle, entropy, compression | A photo's address compresses 70×; noise compresses 1.000×. What is compression actually measuring? |
| 08 · A Finishable Universe | Enumeration, scale contrast | You just finished the 1×1 universe in 17 seconds. How long is 2×2 at the same speed? (Work it out before pressing the button.) |
| 09 · Generators | AI literacy | Noise → image needs a map; the map is the model's weights. What did training actually produce? |
| 10 · A Film's Path | Sequences, measure | All of cinema is ~1017 frames. What fraction of the space has every movie ever made explored? |
| 11 · Witness Registry | Hashing, digital signatures, provenance | If every fake image already exists in the set, can any analysis of pixels prove a photo is real? What can? |
| 12 · Your Words | Text as data, encoding | Your sentence has held its number since before language. What does it mean to “write” it? |
Classroom exercises
- The number by hand (middle school+). Compute the frame count for a 2×2 grid of black-or-white pixels (24 = 16) and draw all sixteen. Then compute 2×2 RGB with logarithms (16,777,2164 ≈ 7.9 × 1028) and check against section 01 at 2×2.
- The birthday problem in the static (high school). If the whole class watches the Static Channel for a year, estimate the probability any two students ever see the same frame. Compare with the classic birthday paradox — why does the intuition break here?
- Meaning is compressibility (high school+). Address a class photo in section 04, export its coordinate in section 07, and record the size. Press random and export again. Explain the difference using the pigeonhole principle: why can't every frame have a small coordinate?
- Finish a universe together (any level). Run section 08 as a class. Before starting, have students predict the sweep time for 2×2 at a million frames per second (≈ 2.5 × 1015 years), then let the 1×1 finish in front of them. Discuss: what does "finishable" depend on?
- Base-256 place value (middle school+). In the viewer, render frame 0, then +1 repeatedly. Predict when the green channel of the last pixel first changes (after 256 steps). Connect to odometers and decimal place value.
- The provenance debate (media literacy). Proposition: “Since every possible fake photograph already exists as a member of this set, photo evidence is dead.” Have one side argue detection, the other provenance, then demo section 11's signed claims and tamper-failure.
- Meet at a phrase (any level). Everyone enters the same seed phrase in section 13 on their own device and confirms they see the same frame — then changes one letter. Discuss hash functions and sensitivity.
Practical notes
Suggested bands: exercises 1, 4, 5 from ~age 12; 2, 3, 7 for high school; 3, 6 and the essay for senior secondary and undergraduate courses in CS, statistics, or media studies. The 1×1 sweep and the atlas are reliable crowd-pleasers for open days.
Everything runs client-side: no student data leaves the machine, and a saved copy of the page works without internet. The source is openly available — classes that want to go deeper can read exactly how the enumeration, compression, and signatures are implemented, or submit a class-witnessed frame to the public Specimen Archive by pull request.
This page prints cleanly — use your browser's print function for handouts.