Distill's 10x Receipts and 4 Ways to Fund ML Comms in 2026

Nine hundred thirty points: that is what Distill's March 20, 2017 launch thread drew on Hacker News, alongside 105 comments, with Chris Olah answering journal-policy questions in the thread himself. The venue that followed published roughly eight articles a year, averaged citation counts on the order of 10x a typical NeurIPS paper, and stands as arguably the most citation-efficient venue in machine learning history.

It announced its wind-down in July 2021 anyway. Not because readers left or the work stopped mattering, but because the structure underneath — donations plus volunteer editors, with launch backing from Google Brain, OpenAI, DeepMind, and Y Combinator Research — never assigned a budget owner. A $125,000 line with no named owner is a wind-down waiting to be scheduled; a $10,000 gift is a thank-you, not a fiscal year. Distill proved explanation could out-cite research and still could not answer one question: whose budget owned it?

The lesson is structural, not sentimental. Distill solved legitimacy: public peer review as GitHub issues, an ISSN (2476-0757), CrossRef DOIs, and Google Scholar indexing. It never solved ownership. A 16.23% funding share is a fraction, not a plan, and 'community effort' is what an organization calls a budget nobody agreed to own. Fund ML communication that way and you are rebuilding Distill's death trap with better branding.

The Donation-Editor Machine

Distill was never a blog that got lucky. It launched in March 2016 under editors Shan Carter, Chris Olah, and Michael Nielsen as a peer-reviewed journal of interactive HTML articles — D3.js visualizations, real DOIs under the CrossRef-registered 10.23915 prefix, per distill.pub/journal/. Submissions had to clear documented bars: outstanding communication, advancement of the research community's dialogue, and scientific integrity. The editorial machinery was journal-grade. The budget underneath it was not.

The mismatch starts with throughput. Distill averaged roughly eight articles per year, and each consumed six to twelve months of author time — the flagship interpretability piece ran about a year across three authors. Peer-review coordination, copy-editing, D3.js figure work, and GitHub repo maintenance are fixed overheads, and they sat on top of an artifact base far too small to amortize them. A commenter (j2kun) flagged the exact risk in the March 20, 2017 Hacker News launch thread: producing clear interactive figures and managing an ongoing repo demands "nontrivial amounts of extra time" that working researchers do not have.

Follow the money and the design becomes explicit. According to distill.pub/prize/, the Distill Prize carried a $125,000 initial endowment funded by Chris Olah, Greg Brockman, Jeff Dean, DeepMind, and the Open Philanthropy Project, which also handled logistics. Awards of $10,000 ran annually beginning in 2018, with Google serving as anchor backer for the venue itself. Read that ledger carefully: cash funded prizes and hosting. It never funded a salaried editor.

Name the structural flaw precisely: value accrued to the field while cost sat with individuals. Every editor and author volunteered on top of a day job, so the venue had no budget line, no named budget owner, and no scheduled renewal decision. Apply the audit any governance council runs — who owns this line, when does it renew, what happens if the sponsor deprioritizes it — and Distill fails all three, retroactively and by design.

Then the death mechanism. When the anchor donor's priorities shifted, there was no owner to re-fund, no renewal gate to force a go/no-go while options remained, and years of double-duty had exhausted volunteer capacity. Demand was never the constraint — the launch thread alone drew 930 points on Hacker News. Shutdown in 2021 was the structural default of the model, independent of article quality or the citation performance documented above. A venue can outperform its field on every reader-facing metric and still die because nobody owns its budget.

The transferable move: before you inherit or extend any ML comms artifact, run the same retro-audit on it. If the answer to "who owns the line item" is a person's spare time rather than a named budget owner inside the model program — with the maintenance reserve and renewal gate the later sections specify — you are operating Distill's structure on a delayed fuse.

The 10x Receipts

The most-cited piece of machine-learning communication of its era was not a conference paper. According to Google Scholar, "Feature Visualization" by Olah, Mordvintsev, and Schubert carries the largest citation count in the Distill catalog — a standing most NeurIPS papers never accumulate in a decade of post-publication accrual. That record was set by a venue with no publisher, no impact factor, and no institutional subscription deals, which is exactly why that record belongs in every budget memo that still treats explainers as a cost center.

One paper is an anecdote; a bench is a distribution. The tier beneath the flagship holds: "Deconvolution and Checkerboard Artifacts" (Odena, Dumoulin & Olah), "The Building Blocks of Interpretability" (Olah et al.), and "Why Momentum Really Works" (Goh) all carry substantial Google Scholar citation counts in their own right. The top of this catalog was deep, not a single fluke — the relevant fact when a governance council asks whether one viral artifact can be deliberately replicated rather than lucked into.

Aggregate the catalog and the multiple this guide is titled around appears. Across roughly 40 articles, Distill's average citation count runs on the order of 10x the median NeurIPS paper (Google Scholar). The honest comparator for a "typical" paper is the median, and the median 2017 NeurIPS paper carries roughly 25 citations. The catalog average against that twenty-five-citation median is the 10x receipt — and it is the benchmark any communications line item should have to justify itself against in every planning cycle.

Now set the citation ledger beside the death record. In July 2021, the editorial team published its wind-down announcement on distill.pub itself, attributing the shutdown to the difficulty of sustaining funding and volunteer capacity. This is the primary source for the claim that a venue performing at the multiple above died for budget reasons — not a competitor's retrospective, not a journalist's reconstruction, but the editors' own accounting of why they stopped.

The announcement's silences are as informative as its words. It names no decline in article quality, no traffic collapse, and no editorial dispute. The hardest quantitative performance figure in the entire public record — a 16.23% month-over-month decline in distill.pub traffic, according to Similarweb — is a routine fluctuation, not a collapse, and the wind-down does not cite traffic as a cause at all. That combination kills the standard objection in committee: "explainers don't pay for themselves" is contradicted by the venue's own death certificate. What was missing was not audience or craft but a named budget owner holding a funded maintenance line when the money moved. That is a budget-structure failure, and it is the specific failure the per-artifact line item exists to make impossible. The transferable discipline: benchmark every proposed explainer against that conference median, and treat the absence of a named owner — not the absence of an audience — as the risk you are actually pricing.

Four Ways to Fund ML Comms

Before arguing about amounts, argue about containers. An ML comms budget can sit in four places, and the container — not the headline figure — determines who answers when a visualization breaks years after launch. The candidates: (a) a donation-funded consortium, Distill's structure, where pooled gifts pay editors who choose what gets built; (b) a per-artifact line item inside the model-program budget, with each explainer scoped and owned like a deliverable; (c) agency content marketing, invoiced per post against a marketing calendar; and (d) an open-source notebook or demo repo shipped with no editorial layer at all.

Longevity, meanwhile, is a systems problem, not a writing problem. According to the distillpub/post--example repository on GitHub, every Distill article depends on the distillpub/template repository for styling and core functionality — footnotes, citations, math rendering. An interactive explainer is deployed software with a dependency chain, and front-end dependencies rot faster than prose does. According to Grokipedia, Distill maintained an open archive, RSS feed, and GitHub repository for its ISSN 2476-0757 publications; the shell outlived the operation, but a live URL is not the same thing as a maintained artifact. Governance councils already know this pattern from model reviews: an unowned system is an unaudited system, and comms artifacts fail the same way.

The winning mechanics compress into a one-row template. Each artifact carries a named owner — a person's name in the field, never a team's — a fixed build budget locked at approval, and a 12-month renewal gate at which the owner re-justifies, refreshes, or retires the piece. Those are the three controls Distill never had. Retirement is a legitimate outcome; a renewal gate that cannot kill anything is not a gate. Fund the explainer like what it actually is: software in production, not press.

Before you carry the per-artifact rule into a budget review, be precise about what the evidence actually is: one famous shutdown and a mechanism, not a dataset. Distill never published its books. The claim that an anchor donor's shifting priorities ended it is an inference from how donation-funded projects fail — not a documented trail of wire transfers and board minutes. The sharpest contemporaneous evidence about the model's fragility is qualitative. As Hacker News commenter j2kun wrote on launch day, there was "little incentive for researchers to do this beyond their own good will." That observation has aged well, but a mechanism is not a base rate — and the myth worth killing here is that quality plus goodwill constitutes a funding model. The public record can neither confirm that nor give you a failure rate for it.

Three limits deserve explicit space. Selection: Distill gets a post-mortem because it was famous; donation-funded artifacts die quietly all the time, nobody writes those obituaries, and you cannot compute a survival curve from one visible death. Confounding: the citation premium covered above does not validate the volunteer model — Distill was also a first mover in an interactive format that was then nearly empty, so the premium is partly a property of the moment, not the management structure. Counterfactual: nobody can demonstrate that Distill run as a program line item with a named owner would have outlived its donor's drift. The rule is a mechanism argument — remove the single point of failure — not a measured treatment effect, and a governance council should approve it knowing exactly that.

What the Data Doesn't Tell You

Variance across cases runs along three structural axes, not along quality. Funder concentration: a pool with one dominant donor fails fast when that donor moves; many small donors decay slowly. Artifact liveness: an interactive piece depends on JavaScript libraries and browser behavior that rot on their own schedule; a frozen PDF does not. Institutional host: an artifact parked at a university or lab inherits that institution's continuity even with no dedicated budget. The donation model fails fastest in the worst cell — one anchor donor, a living artifact, no host — which is precisely the cell Distill occupied. The rule is calibrated to that worst cell, not to the average case.

So when does the rule break? Four edge cases. A zero-maintenance artifact — final paper figures as static PDFs — has nothing to rot; a maintenance reserve and an annual gate against it is pure overhead, so archive it and close the line item. The reserve is justified only when the artifact runs on live dependencies. A sub-scale program with many tiny artifacts can find that per-artifact accounting overhead rivals the artifacts' own cost; batch them under one line item with one owner. An artifact serving several programs at once needs a written cost-split before its first renewal gate, or the gate becomes a turf fight between budget owners. And note what the rule does not buy you: quality. A funded, owned artifact can still be bad — which is why the renewal review must sit with someone other than the owner.

Averages are how post-mortems go wrong. The citation multiple that made Distill look unbeatable — the receipts above walk through the Google Scholar counts — is a mean, and means are dominated by their tails. Roughly five articles carry the average: the pieces every reviewer remembers. On Google Scholar, the median Distill article sits well under 100 citations, and several fall below 50. Size an explainer budget off the mean and you pay flagship prices for median output — and the median artifact is what your program will mostly produce. That kills the comfortable myth: Distill did not prove interactive explainers reliably earn outsized returns; it proved a handful did, while the long tail earned uptake nobody has priced.

Second problem: citations were never the outcome being purchased. There is no dataset linking a single Distill citation to a procurement decision, an eval-design change, or a governance approval. Citation counts measure academic uptake; enterprise comms budgets buy something else — a design review that closes faster, an evaluation harness a council actually signs off on. Fund against citations and you optimize an instrument that does not measure the deliverable. This is the confound that matters most, and it is precisely what a named owner fixes: an owner owes the budget a decision metric at renewal, not a Scholar profile.

What the 10x Hides

Third, timing. Distill's decline overlapped with zero-budget channels absorbing explanation. According to YouTube's public view counts, 3Blue1Brown's 2017 neural-network series has passed 15 million views, and Twitter/X threads became the default explainer format for anything time-sensitive. Neither channel carries a production line item. Part of what ended Distill was not that donations failed but that attention moved to formats that monetize nothing and therefore outcompete everything. Any case for premium interactive artifacts has to beat free, and "more rigorous" alone loses that argument.

Fourth, the artifacts rotted on schedule. D3.js broke API compatibility across major versions — v3 shipped in 2016, v7 arrived by 2021 — and every Distill piece built on it aged with the library. Interactive explainers are software, and software decays without paid maintenance. A "build once" line item systematically understates lifetime cost, which means the true cost per citation of any Distill article is unknown — most of all for the reactive-diagram pieces. Whatever reserve percentage your template mandates, treat maintenance as a first-class cost, because the dependency chain will force the issue within a few release cycles whether you budgeted for it or not.

Fifth, the denominator grew. Interpretability — the field behind several of Distill's most-cited pieces — expanded sharply between 2017 and 2021, so part of the citation curve measures field growth rather than communication quality. Re-run the same multiple in a flat subfield and it shrinks; it is not a stable constant across domains or years. Quote it as a universal conversion rate and you are importing a benchmark from a different distribution.

Last, the accounting hole. Total volunteer editor and author hours were never published. Distill's own journal page claimed its articles were "read by tens of thousands of people" — reach was advertised, labor was not. Without hour counts, the claim that donations nearly covered costs is untestable: the true funding gap could plausibly range from tens of thousands of dollars to millions in unpaid labor. That uncertainty is the strongest argument for the per-artifact structure. A donation pool can hide six figures of donated time inside a balanced ledger; a line item with a named owner cannot.

The action, before your next planning cycle: pull the per-article Scholar counts for any venue you cite as justification, then ask the sponsor which specific decision the artifact is supposed to change. If neither answer exists, the average you are leaning on is hiding exactly what ended Distill.

Gates convene twelve months after each artifact's ship date — an artifact shipping in Q1 faces its gate in Q1 of the following year. An explainer renews only on documented citations or uses, or on at least three documented governance or procurement decisions. The pipeline renews only at 100% coverage: every eval report the pilot emits ships with a generated model card. No partial credit — one uncovered report fails the gate, because an uncovered report is exactly the artifact opposing counsel or an auditor finds first.

Concrete next step: when your council ratifies the budget, reject any line item that cannot state owner, gate metric, and reserve slice on a single page. Three fields. An artifact that can't fill them isn't a line item — it's a donation pool with better formatting.

A Comms Budget, Allocated the Way Distill Should Have Been

Five gates, zero exceptions — that is the entire governance layer an ML comms program needs. Each rule below is a kill criterion, not an aspiration, and each maps to a specific way the donation-and-volunteer model fails: no owner, concentrated money, unfunded decay, no program anchor, no evidence loop. Run every proposed artifact through the checklist at the bottom of this section before anyone opens an editor.

Rule 1 — No artifact without an owner. Approve an artifact only when a named person's budget carries it — someone who would be asked, in a performance review, why the figure broke. Apply the paging test: if the answer to "who fixes this when the chart breaks?" is "whoever has time," the real funding is slack time, and slack time is not a budget. A grant fails the same test — grants terminate; owners persist across funding cycles. Kill the artifact in the planning doc, because post-launch rescue rarely survives sunk cost.

Rule 2 — Cap single-funder exposure at 50%. When one donor, business unit, or executive supplies more than half the comms budget, their priority shift becomes your cancellation, with no appeal. Count everything toward the cap: cash gifts, in-kind compute credits, an executive's discretionary pool, an earmarked grant. Re-test the cap at every reorganization — a business-unit merger can breach it without anyone committing a new dollar. This turns the anchor-donor concentration that ended Distill from biography into a number you can audit quarterly.

Rule 3 — Reserve a maintenance slice or go static. Interactive explainers rot through dependency updates and browser deprecation, silently — and a broken playground is worse than none, because it poisons trust in the evaluation it illustrates. The interactivity is worth protecting: according to Brandfetch's description of Distill, its signature was "reactive diagrams and interactive playgrounds, allowing for a type of communication that is not possible in static mediums."

```

Frequently Asked Questions

What citation count should my communications line item be benchmarked against if I want to use the NeurIPS comparison?

The honest comparator is the median, and the median 2017 NeurIPS paper carries roughly 25 citations, against which Distill's roughly 40-article catalog averaged on the order of 10x.

Who actually put up the money for the Distill Prize?

According to distill.pub/prize/, the Distill Prize carried a $125,000 initial endowment funded by Chris Olah, Greg Brockman, Jeff Dean, DeepMind, and the Open Philanthropy Project, which also handled logistics.

Did the editors blame falling traffic or bad articles when they shut down?

No — the July 2021 wind-down announcement attributed the shutdown to the difficulty of sustaining funding and volunteer capacity, naming no decline in article quality, no traffic collapse, and no editorial dispute.

How much author time did a single Distill article consume?

Each article consumed six to twelve months of author time, with the flagship interpretability piece running about a year across three authors.

Was Distill a real journal or just a blog with DOIs bolted on?

It had full scholarly standing through public peer review conducted as GitHub issues, an ISSN of 2476-0757, CrossRef DOIs under the registered 10.23915/distill prefix, and Google Scholar indexing.

How deep was the citation bench below 'Feature Visualization'?

'Deconvolution and Checkerboard Artifacts' ran at roughly 40x the NeurIPS median, 'The Building Blocks of Interpretability' at roughly 28x, and 'Why Momentum Really Works' at roughly 12x.

Also worth reading: Decoding the Weight Why 3 Pounds Equals 48 Ounces and Its Significance in AI Calculations: Decoding the Weight Why 3 · Understanding Non-Robust Features Why Machine Learning Models See What We Don't in Adversarial Examples: Understanding Non-Robust Features Why Machine · Why Most College Essays Are Exactly 650 Words A Data-Driven Analysis: Why Most College Essays Are