Distill Takes a Break: What the Pause Means for the ML Visualization Community
Distill, the machine learning publication known for its interactive, visually rich scientific articles, is stepping back from active operations. The journal's editorial team announced a hiatus — initially set for one year but potentially indefinite — citing burnout, structural inefficiencies, and a fundamental rethinking of whether a formal journal is even the right vehicle for the kind of work Distill was built to champion.
What Distill Was Trying to Solve
When Distill launched, its core premise was straightforward: non-traditional scientific contributions — interactive visualizations, explanatory articles, dynamic diagrams — weren't being taken seriously by academic institutions because they didn't fit the mold of a conventional paper. The journal's bet was that giving these artifacts a formal publication venue would legitimize them within the academic incentive system and encourage more researchers to produce them.
Over four years, that theory didn't hold up. The editorial team found that open-minded institutions were already willing to engage with high-quality work regardless of where it was published, while more conservative ones remained skeptical no matter the venue. The journal wrapper, it turned out, wasn't moving the needle. And the second assumption — that a dedicated venue would unlock more of this kind of work — also fell apart. The real bottleneck wasn't a place to publish. It was the sheer effort required to produce these articles and the rare combination of scientific depth and design skill needed to pull them off.
The Hidden Cost of Running a Volunteer Journal
Behind the polished articles, Distill's editorial process was quietly unsustainable. Editors — all volunteers, all working on top of their regular research responsibilities — were sometimes putting in more than 50 hours per article, helping authors with diagram design, writing clarity, and the overall shape of scientific communication. That level of engagement is closer to co-authorship than editorial review, and it created real tensions.
Editors found themselves in a dual role: deeply invested in an author's success while also holding the power to reject their work. That dynamic is uncomfortable for everyone involved. And when editors did contribute substantially to an article, the question of co-authorship arose — which then created a conflict of interest if that same editor was expected to make an independent publication decision. Distill also faced criticism for publishing articles by its own editors, even when those decisions were made at arm's length. The journal structure, rather than protecting against the appearance of bias, was amplifying it.
Why Self-Publication May Be the More Honest Path Forward
The editorial team's conclusion is pointed: for most of the work Distill was publishing, a formal journal structure adds overhead without adding proportional value. Self-publication on standalone websites — using tools like the Distill template and GitHub Pages — already works. Researcher David Ha and collaborators demonstrated this clearly with the World Models article, which was self-published, widely read, and taken seriously without any journal affiliation.
The advantages stack up quickly. Self-publication is faster, unconstrained by a journal's scope, and more flexible in format. It also sidesteps the political friction that comes with centralized editorial control. The physics community has largely moved to Arxiv-first publication, and machine learning is trending the same direction. Distill's team believes peer review itself isn't going away — but they're skeptical that bundling it with a publisher is the most effective model. Their own experience with community discussion articles suggested that reviewers engage far more seriously when the stakes feel real and their contribution is visible, rather than when review is a gatekeeping formality.
Papers currently under review won't be affected by the hiatus, existing threads can continue, and the Distill template remains available for anyone who wants to self-publish in the same style. The infrastructure persists — just without the journal apparatus around it. Whether that turns out to be a temporary pause or a permanent shift in how this kind of scientific communication gets distributed, the team's reasoning reflects a broader tension in academic publishing that extends well beyond machine learning.
When a passion project becomes a burden, the most honest thing its creators can do is say so publicly — and that's exactly what the team behind Distill has done. The machine learning journal, long held up as a model for what scientific publishing could look like, is going on indefinite hiatus. The reasons are structural, human, and worth understanding carefully.
A publishing experiment that worked — until it didn't
Distill earned its reputation by doing something most academic journals don't bother with: making research genuinely readable. Interactive diagrams, careful prose, and a visual standard that treated explanation as a first-class concern. For a field as technically dense as machine learning, that was a meaningful contribution. The journal became a reference point in conversations about what modern scientific communication could look like.
But the team behind it has reached a candid conclusion — the model isn't sustainable. Volunteer editors were routinely spending 50 or more hours on a single submitted article, an investment comparable to writing an original paper from scratch. That kind of effort, applied to a relatively small number of publications, still proved impossible to maintain over time. The gap between the standards Distill set for itself and what it could realistically deliver became a persistent source of strain.
The human cost of holding an impossible standard
Burnout isn't a footnote here — it's central to why Distill is pausing. Multiple volunteers experienced it, which the team interprets not as individual failure but as a symptom of deeper structural problems. Two risk factors stand out in their own account: conflicting goals and unachievable ones.
The conflicts were real and not easily resolved. Distill wanted to mentor researchers submitting work, but also had to reject many of them. It wanted its editors to produce their own high-quality writing, but also needed those same people to function as an independent editorial venue. These aren't minor tensions — they pull in genuinely opposite directions, and trying to honor all of them simultaneously is exhausting work.
The team also acknowledges they never clearly defined what they owed to authors who submitted work. Without those boundaries, every submission became an open-ended commitment. That ambiguity compounded the burnout problem considerably.
What Distill's pause means for the field
The team raises a point that deserves more attention than it might initially get: Distill's existence may have been quietly discouraging others from building similar things. When a project occupies a space with enough authority and visibility, it can make the space feel taken — even if the project itself is struggling. If Distill's hiatus opens room for new experiments in ML publishing, that could be a net positive for the community, even if it doesn't feel that way right now.
There's also a more uncomfortable implication in their announcement. Distill has frequently been cited as proof that journals can play a valuable role in machine learning. But if the people who built it no longer believe the model is viable, then citing it as a success story starts to look like a misreading of the evidence. The team is essentially asking the community to update its priors — and to be honest about how hard this kind of work actually is before holding it up as a template.
The Distill template remains open source, and the team has said they'd welcome others building on it. They've also been deliberate about not handing the journal itself off to new stewards, citing a desire to preserve its original character rather than dilute it. The source is available on GitHub, and the work published under it is licensed under Creative Commons Attribution CC-BY 4.0.
The team — Chris Olah, Nick Cammarata, Sam Greydanus, and Janelle Tam — closes with a hope that future projects can learn from what Distill got wrong as much as what it got right. That kind of transparency is rarer than it should be, and probably more useful to the next generation of publishing experiments than any amount of praise for the journal's visual design.
