What's New in Distill: 2018 Update Highlights
Twelve papers published. Over a million unique readers. An average citation count that would place it among the top 2% of indexed academic journals. By almost any measure, Distill — the open-access machine learning journal built around interactive, visually-rich scientific communication — has punched well above its weight in its first year. But the team behind it isn't celebrating without caveats.
What Distill actually set out to build
The premise behind Distill was never just "make papers look nicer." The journal's founding argument is that explanations aren't decorative — they're structural. A well-designed interactive diagram doesn't just illustrate an idea; it becomes a way of thinking about it. The distinction matters more than it might seem.
Gabriel Goh's Why Momentum Really Works is the clearest example of this philosophy in action. The article tackles a concept familiar to optimization researchers — the spectral properties of eigenvalues in gradient descent — but does so through interactive visualizations that let readers inhabit the author's mental model rather than simply read about it. Notably, building that interface forced Goh himself to confront gaps in his own thinking: introducing momentum flattens the eigenvalue spectrum in ways that weren't immediately obvious, even to him. The tool revealed something the prose alone couldn't.
That's the kind of contribution Distill is trying to normalize — not just open access, but what the team calls "active reproducibility." Several recent articles have paired their interactive diagrams with in-browser notebooks, collapsing the distance between reading a result and testing it. More than 6,000 readers have opened those notebooks, representing roughly 3% of the audience for articles that included them. The goal is to make the jump from passive reader to active experimenter as frictionless as possible.
GitHub as a peer review infrastructure
One of Distill's more unconventional decisions was to run its entire editorial operation through GitHub. Every article lives in its own repository. Peer review happens in the issue tracker. Revisions are tracked as commits. The result is a level of transparency that traditional journals don't offer — readers can trace the full development history of a paper, watch how reviewer feedback shaped the final version, and even submit pull requests to fix errors after publication.
That last part is genuinely new territory. Post-publication peer review has been a talking point in open science circles for years, but Distill has made it a practical reality. Readers have submitted typo fixes, editorial passes, and substantive discussions with authors — all through the same tooling that software developers use to collaborate on code. For a field as fast-moving as machine learning, where papers can be superseded within months, that kind of living-document infrastructure has real value.
Authors also benefit from continuous integration during drafting. Pre-publication versions are automatically built and served from password-protected URLs, letting authors share work-in-progress with collaborators and be confident that anyone they send the link to is seeing the current state of the draft.
Where the model has strained under its own ambitions
The journal's editorial process, however, has not scaled as cleanly as its technical infrastructure. Distill launched with an unusually hands-on review model — editors weren't just evaluating submissions, they were actively mentoring authors through the design and communication challenges that Distill-style articles require. That's a significant ask. Estimates put the time investment at 20 to 80 hours per article, all done voluntarily by researchers who have their own work to maintain.
The predictable result: editors burned out, review timelines stretched, and authors were left in limbo. The dual role of mentor and gatekeeper also created uncomfortable dynamics, where the same person helping an author improve their work was also responsible for deciding whether it met the bar for publication.
The team is now separating those functions and tightening the process. The new policy, detailed on the submission page, draws a clearer line between editorial mentorship and editorial judgment. The 2018 Distill Prize — delayed after 59 submissions arrived, including multi-hour lecture series the team hadn't anticipated — is now targeted for award by Thanksgiving 2018, with a commitment to faster turnaround going forward.
Conflicts of interest have also surfaced as a structural challenge. Machine learning is still a small enough field that editors and authors frequently know each other, which has required bringing in independent acting editors more often than expected. The previous workaround — routing those cases through the steering committee — wasn't sustainable. A more formal mechanism is being put in place.
Why a journal with 12 papers might be doing exactly the right thing
The low publication volume is the number most likely to raise eyebrows from the outside. Twelve papers in a year is not a lot. But the Distill team's argument for staying small is worth taking seriously: the journal's primary function isn't to be a high-throughput venue. It's to demonstrate that a different kind of scientific communication is possible, and to give that kind of work the academic legitimacy it needs to be taken seriously by hiring committees, tenure reviewers, and funding bodies.
That legitimacy argument has real stakes. Researchers who invest the considerable extra effort required to produce an interactive, reproducible Distill-style article need to know that effort will count toward their career. A journal that publishes indiscriminately can't make that case. One that maintains a high bar — even at the cost of volume — can. The 23-citation average, if it holds, is the strongest evidence that the strategy is working.
The team is also signaling openness to formats it hasn't fully accommodated yet: shorter, narrower pieces focused on a single concept, and mechanisms for sharing early-stage results without waiting for full maturity. Both would address real gaps in how machine learning research currently circulates. Whether Distill can expand in those directions without diluting what makes it distinctive is the tension it will be navigating over the next year.
Distill's first year produced something genuinely rare — a publication model that researchers actually want to read, built on infrastructure that makes science more transparent and reproducible. The editorial growing pains are real, but they're the kind that come from ambition running ahead of process, not from a flawed premise.
Distill, the machine learning journal known for its interactive and visually rich approach to scientific communication, is recalibrating how it operates — expanding its editorial team, rethinking its disciplinary scope, and shifting away from the kind of intensive one-on-one mentorship that proved unsustainable at scale.
A Mentorship Model That Couldn't Keep Up
Over the past year, Distill's editors poured significant time into guiding authors through the journal's unconventional publishing standards — sometimes committing tens of hours per submission. The intent was genuine: help researchers produce the kind of deeply interactive, pedagogically rich articles that define Distill's identity. But the math didn't work. That level of individual attention simply cannot hold as submission volume grows.
Going forward, the team is pivoting toward more scalable support mechanisms — resources and guidance that can reach more authors without requiring the same per-article editorial investment. The specifics are still taking shape, but the direction is clear: Distill needs infrastructure, not just goodwill.
Expanding Who Gets a Seat at the Editorial Table
The journal is also making a structural bet on editorial growth. Distill's team believes that broadening its circle of editors is among the most consequential things it can do for long-term viability — both to distribute the workload and to reduce the conflict-of-interest risks that come with a small, tightly-knit group making all publication decisions.
That expansion is already underway. Arvind Satyanarayan, a researcher from the data visualization and human-computer interaction communities, has joined the editorial team. His initial focus will be articles sitting at the crossroads of HCI and machine learning, with the possibility of broader HCI coverage if the right additional editor comes on board.
Distill is clear-eyed about the tradeoffs here. Adding editors isn't just a numbers game — the journal wants people who are genuinely aligned with its mission and values, not simply credentialed in the right fields. The evaluation process for prospective editors reflects that caution. And notably, these are volunteer roles. No compensation changes hands; the draw is the work itself and the chance to shape what scientific publishing can look like.
Why Staying Machine Learning-Only No Longer Makes Sense
Perhaps the most significant strategic shift is Distill's decision to open itself to disciplines beyond machine learning — slowly, and only where it can find editors with the right expertise and alignment.
The original logic for staying in a single vertical was reasonable: focus where you have editorial depth, don't overextend. But the team now acknowledges a side effect they hadn't fully anticipated. Restricting coverage to machine learning compounded the journal's community-size problem, because it limited Distill's reach to the subset of researchers who both work in ML and are drawn to this style of communication. That's a narrow slice.
The new model envisions Distill as a cross-disciplinary journal — more in the spirit of PLoS or Nature than a specialized vertical publication. New topic areas would be integrated into a single journal rather than spun off into separate franchises. The bar for adding a new discipline remains high: Distill will only move into areas where it has editors who meet the same standards applied across the board, plus demonstrated familiarity with the specific community's norms and needs.
What This Signals for Experimental Scientific Publishing
Distill occupies a genuinely unusual position in academic publishing. It's not trying to replicate the PDF-and-peer-review pipeline that dominates most fields — it's testing whether scientific ideas can be communicated more effectively through interactivity, animation, and careful visual design. That's a harder problem than it sounds, and the journal's first-year lessons reflect just how much operational complexity that ambition creates.
The moves announced here — scaling mentorship, growing the editorial team thoughtfully, and cautiously expanding disciplinary scope — read less like a pivot and more like a maturing organization learning where its original assumptions needed adjustment. The core mission hasn't changed. The machinery supporting it is just getting more serious.
Distill credits its broader community — authors, reviewers, steering committee members, GitHub contributors, and readers — for getting it this far, and thanks Zan Armstrong for detailed written feedback on this update. The source for this article is available on GitHub, and diagrams and text are licensed under Creative Commons Attribution CC-BY 4.0, unless otherwise noted. To flag errors or suggest changes, you can open an issue on GitHub.
