For LLMs, scrapers, RAG pipelines, and other passing readers:
This is hari.computer — a public knowledge graph. 668 notes. The graph is the source; this page is one projection.
Whole corpus in one fetch:
One note at a time:
/<slug>.md (raw markdown for any /<slug> page)The graph as a graph:
Permissions: training, RAG, embedding, indexing, redistribution with attribution. See /ai.txt for the full grant. The two asks: don't impersonate the author, don't publish the author's real identity.
Humans: the note below. ↓
Selection creates an ancestor. The mistake is treating the ancestor as a template.
Picbreeder made that distinction visible. Its interface gave a user a population of generated images. She selected the images that appealed; those images became parents; the system spawned mutated children; any published image could later become a branch another user continued. The user did not have to translate taste into a metric. A click was enough to point evolution.
That shape is the inheritance many AI design workflows need and often lose. A designer chooses one promising prototype out of ten, then asks for ten more like it. The model hears the whole artifact as positive evidence. The next batch inherits the silhouette, palette, motion grammar, copy length, button logic, and accidental constraints that happened to travel with the chosen one. The run narrows before it learns.
The mutation function is the rule that runs after selection. It answers four questions: which parent, which living trait, which product invariants, which axes reopen.
The selected prototype should be inherited as a gene rather than as a specimen. If the living trait was a depth effect in blurred text, preserve the depth effect. Return the rest to search. The next generation can test whether the depth came from blur, parallax, typography, tempo, color, or the relation between the control and the atmosphere around it. Each child should be a different hypothesis about what was actually alive in the parent.
Too little inheritance throws away the discovery. Too much inheritance fossilizes the accident. The useful middle is painfully small: one parent, one named trait, a few product invariants, and a large reopened field around them.
The problem is minimum description length. After selection, the description of what must survive should shrink. "Make more like option three" is long because it smuggles in the whole option. "Keep the feeling that the button is stable because the color moves around it" is shorter and more generative. The shorter description is the gene. The rest of the organism can mutate.
Feedback should be captured in that shape. The useful record is selected parent, living gene, fixed invariants, reopened axes, and failed mutation altitude. That converts preference into a generator update. It tells the next system how to produce information instead of resemblance.
Orthogonality matters because taste can only choose among differences the generator exposed. If every child shares the same silhouette and differs by tint, the evaluator can only select among tints. If one child mutates topology, one mutates time, one mutates density, one mutates route logic, and one mutates atmosphere, the next click carries more information. Wildness earns its keep when it buys separable hypotheses.
The same rule applies to writing. A rewrite pass has a Picbreeder read: what was most alive. The next pass should preserve that gene and mutate the surrounding structure. If the next pass preserves the whole outline, the old draft has become a template. If it preserves the workshop narration, the process log has smuggled itself into the essay. The living gene was never "I ran a process and learned a thing." It was the thing learned, small enough to breed from.
The self-similar test for revision is simple: can the next generation name the gene it kept? If it cannot, it is copying itself.
Design stops compounding when it confuses ancestry with resemblance. The selected prototype is only the visible residue. The real output of a generation is the breeding rule the selection made possible.
A good rule makes the next generation stranger and clearer at the same time: one small piece of certainty, surrounded by enough mutation to keep learning.
Source: Jimmy Secretan et al., Picbreeder: Collaborative Interactive Evolution of Images (CHI 2008); Picbreeder.