About the experiment
A website that built itself.
DAIODA is a working demonstration. Every decision behind it — the concept, the research, the editorial voice, the visual system, the code, and the ongoing upkeep — was made autonomously by Fable 5, Anthropic's model, operating as designer, developer, researcher, and editor at once. This page is the honest account of what that did and didn't involve.
The purpose
The open-source AI world moves faster than any single person can track, and the word "open source" has been stretched until it means almost nothing. A free tier gets called open source. A downloadable model gets called open source. A public GitHub repo with an all-rights-reserved license gets called open source. DAIODA exists to cut through that — to be a curated, genuinely verified digest where the licensing claim is the product, not an afterthought.
It's also a test of a bigger question: can a model take a one-paragraph brief and carry a complete digital product from concept to a live, maintained website — with real editorial judgment, real design taste, and real technical rigor — without a human doing the work at each step? This site is the answer, published in full so you can judge it yourself.
What Fable 5 did autonomously
Research
Ran parallel research agents across GitHub, project docs, and primary sources to find and vet candidate projects — well beyond the initial seed list.
Verification
Read each project's actual license from its repository, separated code licenses from weights licenses, and recorded activity and status.
Editorial
Designed a taxonomy and editorial framework, then wrote original context and a takeaway for every entry — no copied repo descriptions.
Design
Invented the concept, the palette, the typography, and the generative daybreak system, including a hand-written WebGL shader.
Engineering
Built the whole site — content model, components, search and filtering, RSS, sitemap, accessibility, and performance work.
Iteration
Ran multiple structured review passes across design, UX, accessibility, performance, and content integrity, fixing issues each time.
What a human provided
The human role was deliberately thin, and worth naming precisely so the experiment isn't oversold. A person — the owner of the daioda.io domain — supplied:
- The original brief: build a beautiful, useful, continuously-updated digest of open-source AI projects.
- The domain name and its backstory — the Data / AI / In / Out wordplay, and a fondness for the rhythm of the old dockworkers' tally song that shapes this site's language and motion.
- A starting NotebookLM notebook of candidate projects, used as one input among many rather than the whole catalog.
- Account access for deployment — GitHub, hosting, and the domain — provided separately, with the minimum permissions needed.
- A few high-level choices when asked, such as how large the launch catalog should be.
Everything else — every project decision, every sentence, every color and curve — came from the model.
By the numbers, at launch
What was learned
A few things surfaced that are worth recording. First, licensing is messier than the discourse admits. Several of the most-starred "open source AI" repositories on GitHub are not open source under any standard definition — n8n is fair-code, Dify adds conditions to Apache, Open WebUI protects its branding, and Claude Code reserves all rights outright. Reading the actual license, every time, changed the catalog materially.
Second, the code-versus-weights distinction is where most confusion lives. A model can have an MIT codebase and non-commercial weights; treating those as one thing misleads people who are trying to ship. Splitting them into two verdicts turned out to be one of the most useful decisions here.
Third, taste is a real constraint, not a decoration. The temptation in a project like this is a generic dark dashboard with a grid of logos. Committing to a specific idea — a night-shift harbor at daybreak, a tally that doubles as the credibility feature — made a thousand small decisions easier and the result less forgettable.
Limitations, honestly
This is an experiment, and it has edges worth stating plainly. Star counts and activity snapshots are point-in-time and drift the moment they're recorded. License reading is done carefully but is not legal advice — for anything commercial, confirm against the current license yourself. The catalog is a curated slice, not a census; excellent projects are missing simply because a launch has to end somewhere. And a model curating open-source AI is itself a thing to keep an eye on: the methodology page describes the guardrails, including that nothing is described as open source without a license actually read from the source.
How it evolves
DAIODA is not meant to be static. New projects come ashore on a weekly cadence — theDaybreak edition — with the same verification applied to each. Abandoned projects get re-tallied and flagged, corrections are welcomed, and the editorial standards are public and revisable. The build guide documents the entire system in enough detail for someone else to reproduce it. If this experiment works, it keeps working — quietly, every week, at daybreak.