Research Debt
Olah and Carter name a failure mode that conventional publication metrics reward but do not see. A field can keep producing results while becoming progressively harder to enter, understand, and extend. Its debt appears as poor exposition, ideas left in their first rough form, inherited abstractions and notation that impede thought, and a volume of work whose signal is costly to recover. The paper’s durable move is to treat these not as cosmetic defects but as accumulated missing interpretive labor.
The cost is multiplied on the reader side
Explanation trades effort between author and audience. In one-to-many communication, author effort is paid once while comprehension effort is paid by every reader. A weak explanation can therefore look locally efficient and be globally wasteful: each future reader reconstructs the framing, analogy, or abstraction that the author did not supply. As the community grows, that repeated cost encourages specialization and fragmentation because no member can afford to understand all the others.
This is a sharper account of noise than “too many papers.” Poorly digested work makes triage expensive: even recovering a one-sentence contribution may require a full read. Prolific publication then increases both the number of items and the cost per decision. keshav2007 supplies a reader’s defense—allocate depth through explicit passes—but it cannot remove the debt embedded in the source.
Distillation changes the idea
The proposed opposite is research distillation. This is not copyediting or popularization after the real work. Developing the right notation, visualization, analogy, and abstraction can expose missing understanding and transform the underlying idea. The paper consequently rejects two easy organizational answers: one generalist cannot polish an entire field, and the work cannot simply be outsourced to a non-expert, because good distillation requires both deep subject knowledge and design judgment.
The argument complements luhmann1981 and karpathy2026. Those works explain how a personal or agent-managed knowledge system can make understanding accumulate. Olah and Carter explain why the public source layer must accumulate too: otherwise every private knowledge system repeatedly services the same communal debt.
The institutional proposal
The final section proposes an ecosystem in which distillation is recognized as a research role, supported by a journal, prizes, and interactive-publishing infrastructure. That proposal is historically situated in Distill’s own mission and therefore also functions as an argument for the venue publishing the paper. The diagnosis does not depend on that particular institution, but the ecosystem claim is advocacy rather than evidence that those mechanisms will correct incentives.
- Durable: the reader-side multiplier on interpretive labor; the four recurring forms of debt; and the claim that better explanation can require new conceptual work, not merely clearer prose.
- Evidence boundary: this is a conceptual essay supported by examples and an informal cost model, not a measurement of debt across fields or an evaluation of interventions. It does not operationalize explanation quality, fragmentation, or the counterfactual effort saved.
- Era-bound: the 2017 Distill ecosystem, its prize amount, infrastructure, and status as a proposed career home.
- Reflexive stake: the authors were affiliated with Google Brain and the paper was published by Distill; its call to recognize interactive research exposition also legitimizes the kind of work and venue it exemplifies.
For this library, the paper is both a warning and a test. A synthesis note repays debt only if it exposes structure that saves future readers work. A compressed restatement can add another item to the noise while preserving the same interpretive burden.