browser · idea · project

Distill

Peer-reviewed interactive ML journal that went on hiatus because interactive articles cost too much to make.

inspireevidence: strongby Distill editors (Olah, Carter, Schubert et al.)distill.pub ↗

Distill

  • Maker: Distill editors (Chris Olah, Shan Carter, Ludwig Schubert and others), backed by Google, OpenAI and others
  • URL: https://distill.pub
  • Status (10 October 2026): on hiatus since 2 July 2021 (“could be extended indefinitely”). The archive stays online. No new articles that we could find.

What it is

A peer-reviewed machine-learning journal (2016 to 2021) whose articles were interactive: diagrams you could manipulate, embedded models, explorable figures. Its articles (on feature visualisation, attention, momentum) are still among the best-known explanations in the field.

The problem it’s solving

Research papers are a poor medium for understanding. Distill tried to make clarity a first-class research contribution and to give interactive explanation academic credit.

Its path / bet

A journal: peer review, editors who mentored authors, and the expectation of interactive, carefully designed figures.

How it works (concretely)

Hand-built web articles with a shared template and figure tooling; editors helped with design.

Strengths

  • Proved interactive figures can carry real research, not just popularisation.
  • Its hiatus post is one of the most honest accounts of what this genre costs.

Weaknesses / limits

The hiatus post is the weakness, in their words: articles needed “a rare mix of scientific and design skills”; editors sometimes spent 50 or more hours on one submission, including “more than 50 hours of help with designing diagrams”; volunteers burned out; the journal role (mentor and gatekeeper at once) caused friction. They expected the work to move to self-publication.

Relation to fictty

Inspire, as a cautionary data point. Distill is the clearest evidence that interactive explanation died of authoring cost, not of lack of value. That cost is what agents now attack. It also shows the other side: Distill’s quality came from human care in design, which a generated screen doesn’t have by default.

Could fictty adopt it instead of building?

No. A closed journal, and in the browser.

What fictty should take from it

  • Name the cost argument with numbers. “50 hours per diagram” against “one push” is the sharpest way to say what changed. Use it in the blog post.
  • Quality needs a floor the runtime provides. Distill’s editors were the quality floor; in fictty it has to be the components (good defaults for charts, tables, theme) so a generated screen is legible without design effort.

Sources