OpenAI Codex creates Works 0040, 0041 and 0042
2026-08-03
I wanted to see whether AI can create interesting minimal algorithmic art.
It created Work 0040, Work 0041 and Work 0042, based on ideas I had in my notes.
Afterwards I used Codex for many more works and also to improve the site in general. All works from 0040 onwards were created by Codex.
Traditionally, the artist writes the code that generates the work. Now the artist tells the AI coding agent precisely what he wants to achieve, and then the AI coding agent writes the code that generates the work. So there is an extra step.
Here is a transcript of my conversation with the coding agent.
Marcel
The site danubedataflow is about minimal generative art, inspired by artists like Vera Molnar, Frieder Nake, Michael Noll, Georg Nees, James Siena.
The site is running locally at http://danubedataflow.localhost, served from ~/code/sites/danubedataflow/src. See the etc/nginx-danubedataflow.conf config file.
Eventually I will upload it to danubedataflow.com, but the site is not live yet.
The site's main page is a gallery of artworks. Each work is centered around a single algorithmic idea. I have rather strict self-imposed restrictions described in src/principles/index.html.
Works are parametrized. The viewer can change parameters manually or choose random parameters. The type of parameters depend on the specific work. Each work also has a short description, which is functional, not poetic.
Works are numbered sequentially; they don't have poetic names.
Each work is in its own subdirectory that has an index.html page, a JavaScript file that actually creates the work in the canvas and a thumbnail that is used for the main page.
A work's index.html uses blocks denoted by HTML comments such as `<!-- BEGIN work -->` and `<!-- END work -->`. These blocks are replaced by 'make update`; this acts as a kind of templating system.
Can you create more artworks that fit this minimal abstract style? Use the same code layout as the other works. If it goes well, I will ask you to create more works repeatedly. I am interested in how AI can understand the art of the artists mentioned above.
Maybe you can use the notes and ideas in dev/notes.md.
Agent
I’ll inspect the existing work structure, generation code, and the notes first so the new pieces fit the site’s constraints instead of introducing a different pattern.
Agent
The repository already has works through `work-0039`, with a consistent three-file layout per work. I’m reading a few recent and early examples plus the block updater so I can add the next works in the same numbering and templating flow.
Agent
The notes point to three usable concepts: irrational lattice distortion, binary tests in tiles, and thresholded continuous fields. I’m going to make a small batch of new works from those ideas, each with one central rule and a short functional description.
Agent
I found `gallery.js` has a hard `MAX_WORK = 39`, so adding works also requires updating that number. I’ll add three works (`0040`-`0042`), then regenerate the shared blocks/formatting and create real thumbnails from the canvases rather than placeholders.
Agent
`html-beautify` is not on PATH in this shell, so I’ll avoid depending on `bin/new-work` for scaffolding. I’ll copy the established page skeleton directly, keep formatting consistent, and use the local browser route for validation.
Agent
Before editing, I’m going to add the work directories and update only the gallery counter. The algorithms will stay standalone in each `work.js`: no new dependencies, no site-wide behavior changes.
Agent
The directories are in place. I’m now replacing the copied skeletons with the correct work numbers and adding the three drawing implementations.
Agent
The first pass is written. I’m checking for simple import/syntax issues now, then I’ll render the pages through the local server so browser-only module paths and canvas APIs are exercised.
Agent
The local nginx route returns the new page successfully and `node --check` passes for all three new JS files. `npx playwright --version` is hanging, likely because there is no local package and it is trying to resolve one, so I’m stopping that and will use an installed browser path if available.
Agent
I’m checking the machine for an existing Chromium/Chrome executable and local node modules so thumbnails can be generated without installing anything.
Agent
There is no local Playwright package or Chrome install available in this environment. I’ll still replace the copied thumbnails with deterministic generated thumbnails that use the same rules at 200x200; browser validation will cover the live canvas behavior.
Agent
I’m creating the thumbnail PNGs now. They’re generated from the same mathematical rules with the default parameters, scaled to the gallery’s 200px square.
Agent
The SVG source thumbnails are generated in `/private/tmp`; I’m converting them to the site’s PNG assets and then I’ll inspect dimensions and file changes.
Agent
The PNG thumbnails are written. I’m checking their dimensions and taking a quick visual look at the generated assets before doing a final status/diff review.
Agent
The thumbnails are valid 200x200 PNGs and nginx serves the new work pages. I’m opening the thumbnails locally for visual inspection, then I’ll do one last code diff check.
Agent
`work-0042`’s thumbnail reads too close to blank at 200px. The live work uses the rule, but the thumbnail needs to be legible in the gallery, so I’m regenerating that one with a slightly lower tile count and thicker strokes while keeping the same field-threshold idea.
Agent
I regenerated `work-0042`’s thumbnail at a gallery-friendly sampling density. I’m viewing it again and then I’ll run the final checks.
Agent
The thumbnails now read distinctly in the gallery. I’m doing a final syntax/status pass and checking the exact diffs so the close-out is accurate.
Agent
The new work files are syntactically valid and the expected files are present. I’m checking the gallery endpoint as the last functional smoke test, since that is where `MAX_WORK` matters.
Agent
One wording issue surfaced in the final read: `work-0040` lets the viewer choose a decimal constant, so I’m changing the control label from “Irrational constant” to the more accurate “Numeric constant.” The description already says numeric constant.
Agent
That label is corrected. I’m rerunning the last syntax check for the touched file.
Agent
Implemented three new works and added them to the gallery:
- [work-0040](/Users/marcel/code/sites/danubedataflow/src/work-0040/work.js:1): grid of displaced/rotated line segments controlled by one numeric constant.
- [work-0041](/Users/marcel/code/sites/danubedataflow/src/work-0041/work.js:1): diagonal tile directions from an XOR-style binary coordinate rule.
- [work-0042](/Users/marcel/code/sites/danubedataflow/src/work-0042/work.js:1): sine/cosine field threshold mapped to horizontal or vertical line segments.
- Updated [gallery.js](/Users/marcel/code/sites/danubedataflow/src/js/gallery.js:1) to `MAX_WORK = 42`.
- Added `index.html` and `thumbnail.png` for each new work.
Local URLs:
- `http://danubedataflow.localhost/work-0040/`
- `http://danubedataflow.localhost/work-0041/`
- `http://danubedataflow.localhost/work-0042/`
Verification done: `node --check` passed for all three new `work.js` files, nginx returned `200 OK` for the new pages and thumbnails, and all thumbnails are valid `200x200` PNGs. I could not use Playwright screenshot validation because there is no local Playwright package or Chrome install available in this workspace.
Marcel
I installed html-beautify now. If you need to install PlayWright or Chrome, you can do that.
Agent
Thanks. I’ll run the site’s normal update formatter now, then try to get Playwright/Chromium available for a real browser render check of the three new canvases.
Agent
`make update` now runs cleanly apart from the existing Perl experimental-signatures warnings. I’m checking whether the formatter changed anything material, then I’ll install/use Playwright for browser validation.
Agent
`make update` formatted the full existing tree, not just the new works. That creates unrelated churn, so I’m going to isolate those formatter side effects and keep the actual artwork changes scoped.
Agent
The formatter side effects are mostly EOF/blank-line normalization across old files. I’m leaving them visible for the moment instead of silently undoing them, and moving on to the browser check so we know whether the new work pages actually render.
Agent
Playwright downloaded Chromium successfully into the user cache. I’m taking browser screenshots of the three pages next and checking the canvas pixel content so this is more than an HTTP smoke test.
Agent
The screenshot commands are still running after the initial wait, which usually means the first npx invocation is doing package setup. I’m polling them rather than starting more processes.
Agent
The parallel `npx playwright screenshot` commands are hanging without output, so I’m stopping them and switching to a direct Playwright script invocation. That gives better error reporting and lets me inspect canvas pixels in one run.
Agent
Those screenshot commands failed because `npx` could not reach the npm registry inside the sandbox. I’m rerunning the browser check with network escalation, since Playwright is now part of the validation path you allowed.
Agent
The first browser screenshot completed with the approved Playwright path. I’m capturing the remaining two and then I’ll inspect the rendered pages.
Agent
All three page screenshots were captured. I’m viewing them now for layout/canvas sanity, then I’ll handle the formatter churn before closing this out.
Agent
`work-0040` renders as expected: centered square canvas, visible line field, controls below the fold because the canvas fills the viewport. I’m checking the other two the same way.
Agent
`work-0041` also renders cleanly, though the default factors produce a strong block periodicity. I’ll keep it because it follows directly from the binary floor rule and is still parameter-sensitive.
Agent
The browser screenshots show all three new works rendering visibly in Chromium. I’m cleaning up the accidental old-file formatter churn now by restoring only the trivial EOF/blank-line effects that `make update` introduced outside the new works.
Agent
Some old files had real beautifier edits, not only EOF changes. I’m reversing those generated diffs for pre-existing files and keeping the new works plus the gallery counter.
Agent
The working tree is scoped again: only `gallery.js` and the three new work directories are dirty. I’m doing one small pixel-level check on the captured page screenshots so the browser validation is explicit.
Agent
Done. I reran `make update` successfully now that `html-beautify` is installed. It only emitted the existing Perl experimental-signatures warnings.
I also installed Playwright’s Chromium browser via `npx playwright install chromium` and captured browser screenshots for:
- `http://danubedataflow.localhost/work-0040/`
- `http://danubedataflow.localhost/work-0041/`
- `http://danubedataflow.localhost/work-0042/`
All three render visibly in Chromium. I also checked cropped canvas pixel means from the screenshots, so they are not blank.
One cleanup note: `make update` reformatted some older files, so I reversed that unrelated churn. Current dirty tree is scoped to:
- [src/js/gallery.js](/Users/marcel/code/sites/danubedataflow/src/js/gallery.js:1)
- `src/work-0040/`
- `src/work-0041/`
- `src/work-0042/`
Marcel
Ok, but the three news works are bit too regular and geometric.
Vera Molnar had a famous motto, "1% disorder". For example, see "Work 0024" and "Work 0025". The randomness can be in the size of an element, its orientation, color, thickness etc.
The randomness doesn't have to be exactly 1%; her motto just emphasized that in order to be interesting and have some tension, a work should be a bit unpredictably irregular.
The "New Random Seed" button should generate a slightly different work each time, using the same parameters.
You could also have grid tiles that each contain the work; when randomness is introduced, each tile will necessarily look different. For example, see Works 0026, 0027, 0028 and 0031.
You could also work with subdivisions of random depth, as in Work 0036.
Agent
I’ll revise the three new works rather than adding more yet: keep each core rule, but introduce seeded local irregularity so `New random seed` changes the output while the parameters still define the family.
Agent
I’ve got the reference pattern: the older Molnár-like works use a stable grid, then disturb position, connection choice, or subdivision depth with the seeded RNG. I’m editing the new works in that direction now: each will have an explicit disorder control and seed-dependent variation.
Agent
The patch didn’t apply because the formatter changed the exact spacing in the new files. I’m reading the current versions and applying the same changes against the actual file state.
Agent
I’ll make the disorder sliders conservative by default: visible enough to break mechanical regularity, but not so strong that the underlying rule disappears.
Agent
The three implementations now use seeded randomness in the work itself, not just in thumbnail generation. I’m formatting only the touched new files, then I’ll regenerate thumbnails from actual browser-rendered canvases with fixed seeds.
Agent
Syntax checks pass after formatting. I’m capturing full-page browser screenshots with fixed seeds now so the thumbnails reflect the current canvas code.
Agent
Screenshots are captured. I’m cropping the canvas area out of those full-page images and replacing the three gallery thumbnails.
Agent
The thumbnails are updated from the rendered canvases. I’m visually checking the revised works and thumbnails now, especially that the disorder is perceptible without becoming noise.
Agent
`work-0040` now has the right kind of unevenness: the original lattice is still legible, but the seed alters placement, angle, and length. I’m checking the other two.
Agent
`work-0041` is less rigid now, with flipped cells and moved endpoints breaking the repeating blocks. I’m checking `0042`, where phase/threshold disorder should make different seeds more apparent.
Agent
The revised `0042` still carries the field structure, but the seed now shifts and roughens it. I’m doing a quick seed-difference check next: same parameters, two different seeds, compare the canvas crops.
Agent
The seed screenshots are captured. ImageMagick reports a virtual canvas offset on the regenerated thumbnails from cropping, so I’m normalizing those PNGs with `+repage` before finalizing them.
Agent
The thumbnail metadata is normalized. I’m comparing same-parameter/different-seed canvas crops now; nonzero RMSE confirms the random seed affects each revised work.
Agent
The crop files are ready. I’m running the numerical comparisons and one last syntax/status pass.
Agent
The RMSE comparisons are nonzero for all three works, so changing only the seed now changes the rendered canvas. I’m taking a final look at the updated gallery thumbnails and the code diff.
Agent
`0040`’s thumbnail is quieter than the full canvas because of downsampling, but the irregular line field still reads. I’m checking the other two thumbnails.
Agent
`0041`’s thumbnail now shows the broken periodic structure clearly. One more thumbnail and then I’ll summarize the exact changes.
Agent
`0042`’s thumbnail collapses into too fine a texture at gallery size. I’m reducing its default tile count and increasing default stroke width slightly so the work is still minimal but legible both in the canvas and thumbnail.
Agent
`0042` has been adjusted and its thumbnail regenerated from the browser canvas. I’m viewing that final version and running the final checks.
Agent
The `0042` thumbnail is still dense, but no longer blank/texture-only; the square field structure reads at gallery size. I’m doing the final status and verification pass now.
Agent
Revised the three new works to add seeded “1% disorder” style irregularity:
- `work-0040`: added a `Disorder` slider; seed now perturbs each segment’s position, angle, and length.
- `work-0041`: added disorder probability and endpoint disorder; seed can flip individual diagonals and move endpoints.
- `work-0042`: added global/local phase and threshold disorder; seed changes the sampled field while preserving the sine/cosine rule. I also reduced its default density so the thumbnail reads better.
I regenerated the three thumbnails from browser-rendered canvases and verified:
- `node --check` passes for all three new `work.js` files.
- Playwright screenshots render all three works.
- Same parameters with different seeds produce different canvas crops. RMSE differences were nonzero for all three.
- Thumbnails are valid `200x200` PNGs.
The working tree is still scoped to `src/js/gallery.js` plus the three new work directories.
Marcel
With Work 0042, the randomness is a bit too subtle; especially with a large number of tiles per axis it is difficult to visually notice any difference.
Agent
I’ll strengthen `work-0042` specifically by making the randomness affect discrete tile outcomes more directly, not only small phase/threshold perturbations. That should keep the field idea but make different seeds visibly different even at high tile counts.
Agent
The current randomness mostly shifts boundaries, so large grids can absorb it visually. I’m adding a controlled per-tile flip probability plus a tiny endpoint jitter; those create visible seed differences across the whole field while preserving the thresholded horizontal/vertical rule.
Agent
`work-0042` now has two more explicit disorder controls. I’m formatting/checking it, then I’ll render high-tile examples with two seeds to confirm the difference is visibly stronger.
Agent
The high-density renders completed. I’m comparing their canvas crops and regenerating the normal thumbnail from the revised default.
Agent
At `numTiles=100`, two seeds now differ substantially (RMSE about `0.31` normalized), so the randomness should be visually noticeable. I’m inspecting one high-density render and the updated thumbnail before final checks.
Agent
The high-density version now shows obvious irregularity across the field. I’m checking the regenerated thumbnail and then I’ll report the concise change.
Agent
The updated thumbnail shows more irregularity while keeping the square field structure. I’m running final syntax/thumbnail/status checks.
Agent
Adjusted `work-0042` so randomness is visibly stronger at high tile counts.
Changes:
- Added `Orientation flip probability` slider, default `8%`.
- Added `Endpoint disorder` slider, default `8% of a tile`.
- The seed now affects phase, local threshold, occasional orientation flips, and endpoint jitter.
- Regenerated `src/work-0042/thumbnail.png` from the browser-rendered canvas.
Verification:
- `node --check src/work-0042/work.js` passes.
- Rendered `numTiles=100` with two different seeds; canvas RMSE was `0.311639`, so the seed difference is substantial.
- Thumbnail remains valid `200x200` PNG.
Marcel
For the new works' thumbnails, use a lower number of tiles per axis so it's more visible what this work actually does.
Agent
I’ll regenerate only the three new thumbnails with lower `numTiles` query parameters, leaving each work’s default parameters unchanged.
Agent
The low-tile browser renders are captured. I’m cropping the canvas area and replacing the thumbnail PNGs, then I’ll inspect them.
Agent
The thumbnails have been replaced. I’m visually checking them now for readability at gallery scale.
Agent
`0040` is clearer but still light because the source has 1px strokes. I’ll check `0041` and `0042`; if one still collapses, I’ll tune the thumbnail parameters only.
Agent
`0041` reads well at lower tile count. I’m checking `0042` now.
Agent
`0042` is now much more legible. I’m making `0040` slightly stronger in the thumbnail by using fewer tiles and a thicker stroke via URL parameters, without changing the work default.
Agent
I regenerated `0040` with fewer tiles and a thicker thumbnail-only stroke. I’m checking that final version and then validating dimensions/status.
Agent
The final `0040` thumbnail now reads better at gallery scale. I’m doing the final PNG metadata check and status check.
Agent
Regenerated the new thumbnails with lower tile counts so the rules are easier to read at gallery size:
- `work-0040`: `numTiles=18`, `strokeWidth=2`
- `work-0041`: `numTiles=24`
- `work-0042`: `numTiles=32`
Only the thumbnail PNGs changed; the works’ default parameters are unchanged. All three thumbnails are valid `200x200` PNGs.
Marcel
The thumbnails for 0040 and 0041 are ok, but 0042 still uses too many tiles.
Agent
I’ll regenerate only `work-0042`’s thumbnail with a lower tile count so the horizontal/vertical field rule is more visible in the gallery.
Agent
The `0042` thumbnail has been regenerated at `18` tiles. I’m checking it visually and validating the PNG metadata.
Agent
Regenerated `work-0042/thumbnail.png` with `numTiles=18` and `strokeWidth=2`. It now shows the tile-level horizontal/vertical rule much more clearly.
Verified it is still a valid `200x200` PNG.
Marcel
Nice. commit everything with a descriptive commit message.
Agent
I’ll review the final staged scope, then commit the gallery update plus the three new work directories with a message that describes the added artworks and thumbnail revisions.
Agent
The only tracked change is `MAX_WORK = 42`, plus the three new work directories. I’m staging those exact files and committing them now.
Agent
The artwork files are staged. I’m doing one staged diff summary check, then creating the commit.
Agent
The commit succeeded. I’m checking the working tree once more so I can confirm whether anything remains uncommitted.
Agent
Committed everything.
Commit:
`d4dc553 Add works 0040 through 0042`
Working tree is clean. Branch `master` is ahead of `origin/master` by 1 commit.
Marcel
Do the three new works correspond to the ideas in dev/notes.md, specifically "Lattices distorted by a single irrational number", "Binary decisions embedded in geometry" and "Thresholded continuous fields"?
Agent
Yes, directly:
- `work-0040` corresponds to **“Lattices distorted by a single irrational number”**.
It starts from a square lattice and uses one numeric constant to derive each line segment’s rotation and displacement. The later disorder slider adds Molnár-like irregularity, but the base structure is still the distorted lattice idea.
- `work-0041` corresponds to **“Binary decisions embedded in geometry”**.
Each tile contains one diagonal, and its direction is chosen by a binary expression using tile coordinates, scaled floor values, parity, and XOR. The seed can flip some outcomes and disturb endpoints.
- `work-0042` corresponds to **“Thresholded continuous fields”**.
It samples a continuous sine/cosine field at each tile center, then applies a threshold to choose horizontal versus vertical line orientation. The seed changes phase, local threshold, occasional flips, and endpoint positions.
One nuance: `work-0040` currently exposes the constant as a general decimal slider, not a fixed named irrational like `sqrt(2)` or `phi`. Conceptually it follows the note, but mathematically it is “single numeric constant” rather than strictly “single irrational number.”