pipette
ESEspañol

Why Pipette exists

Thousands of new research papers appear every day, and AI is about to multiply that number. Pipette reads all of them each morning and hands you the few worth knowing about, in the words of the people who did the work.

The problem

Science now arrives faster than anyone can read it. arXiv alone posts around two thousand new papers on a weekday, and bioRxiv, medRxiv and the big journals add hundreds more. With AI tools helping people write, that flood is only getting bigger.

The tools we have were built for a different job. Search engines work when you already know what to look for. Social feeds reward whatever is loudest. AI summaries are quick, but they put words in scientists' mouths, and a confident summary of a paper is not the paper.

Pipette is for the other moment: opening one page in the morning to see what science found yesterday, across every field, without being sold anything.

What we promise

Every word is the authors'.
Pipette never summarizes or rewrites a paper. The machine picks the sentence that states the main result and labels the paper. What you read was written by the scientists.
No ads, no tracking, no accounts.
Nobody can pay to be in the edition. There are no analytics scripts. Your saved papers and interests stay in your browser.
Honest about what a paper is.
Preprints are marked as not yet peer-reviewed. Abstracts that claim more than they show are flagged for careful reading and never make the daily edition.
Every field gets a seat.
No field may take more than three places in the daily edition, so a busy day in AI cannot push out ecology or mathematics.
We show our work.
The exact questions, the answers with their probabilities and the ranking formula are all public, on every paper and on this page.
Open data.
Everything Pipette produces is published as JSON that anyone can reuse.
Respect for authors and publishers.
When a journal's licence does not allow republishing an abstract, Pipette quotes only the two sentences it selected and links to the publisher.

How it works

  1. Collect. Every night Pipette downloads the new papers of the previous day from arXiv, bioRxiv, medRxiv and 58 leading journals through OpenAlex. Revised versions of older papers are skipped.
  2. Split. Each abstract is cut into sentences, with the mathematics kept intact.
  3. Ask. Jev, a decision model built by TypeSafe, answers ten questions about every paper. Jev does not generate text: it can only choose among options we define and say how likely each one is. That is why it cannot invent anything.
  4. Rank. Pipette combines the answers with a fixed formula, then builds the edition with diversity rules.
  5. Publish. The edition, the full list and one page per paper go live, along with the open data.

Reading yesterday's 191 papers cost US$0.03 in model usage.

The questions Jev answers for every paper

These are sent to the model exactly as written. (jev-1.13.0)

Which topic best describes what this paper is about?
What is the main contribution of this paper?
  • The paper introduces a new method, model, algorithm, technique, device or design and shows how well it works.
  • The paper reports a new observation, measurement or experimental result about the natural, living or social world.
  • The paper proves theorems or develops a theoretical or mathematical framework.
  • The paper's main contribution is a new dataset, benchmark, software tool, database or catalogue for others to use.
  • The paper reviews, surveys or synthesizes existing research, including meta-analyses.
  • The paper argues a position, proposes an agenda or discusses ideas without presenting new results.
What is the main kind of evidence behind the paper's claims?
  • Mathematical proof or formal derivation.
  • Computer simulations, numerical calculations or theoretical modeling.
  • Experiments that run software or AI models on datasets or benchmarks.
  • Physical experiments in a laboratory, including cells, tissues, animals, materials or devices.
  • Analysis of observational or real-world data: telescope observations, field measurements, health records, surveys or cohorts.
  • A randomized controlled trial or other prospective clinical trial in humans.
  • A synthesis of previously published studies: systematic review, meta-analysis or literature review.
  • No new evidence: argument, opinion or proposal only.
Which sentence in `sentences` states the paper's main result or contribution most directly?
Which sentence in `sentences` admits a limitation or weakness of this study's own results (for example a small sample, a narrow setting, or a result that still needs confirmation)? A sentence describing what was unknown before the study is not a limitation. Answer `none` if no sentence admits a limitation.
Who would find this paper interesting?
  • Only specialists in this exact subfield would care about it
  • Researchers across this field would find it interesting
  • Scientists from other fields would find it interesting
  • A curious member of the public would find it fascinating
How big a step forward does the paper claim to be, judging only from its title and abstract?
  • A small incremental improvement or a minor variation of known work
  • A solid, useful contribution to its field
  • A substantial advance that changes how the field approaches a problem
  • Potentially field-changing: a breakthrough or a first-of-its-kind result
How hard is the abstract to understand?
  • A general reader can understand the abstract
  • Understanding the abstract needs some university-level background in the field
  • Only specialists can understand the abstract
Does the abstract claim more than the evidence it describes can support, for example sweeping or promotional claims based on limited experiments?
Does the abstract describe a result that could directly affect people's health, the environment, or technology that people use?

The ranking rule

rank = appeal + 0.9·advance + 0.4·practical − 2.4·max(0, hype − 0.5) − 0.25·level, papers flagged for bold claims (hype ≥ 0.6) are left out of the edition, then at most 3 papers per field and 2 per topic in the daily edition, with the best paper of every field considered first.

What Pipette cannot do

Jev reads only the title and the abstract, not the full paper. Its labels describe what a paper claims, not whether the claim is true.

Peer review, replication and time are what establish results. Pipette is a way to notice research, not a verdict on it.

Jev is most accurate in English. It will sometimes get a topic wrong or miss a paper that deserved the edition. The percentages on each paper show how sure it was.

Sources

arXiv does not announce new papers on weekends, so Saturday and Sunday editions are smaller.

Thank you to arXiv for use of its open access interoperability.

Open data

All files are JSON. Reuse them freely under the terms of the original sources: arXiv metadata is CC0, bioRxiv and medRxiv abstracts carry the licence their authors chose, and journal metadata comes from OpenAlex (CC0).

Privacy

Pipette has no accounts, no ads and no analytics. The only cookie is lang, set when you choose a language, so the site remembers it.

Saved papers and your For you settings are stored in your browser's local storage and never leave your device, except the one sentence you type in For you, which is sent to Jev to rank the day's papers and is not stored.

The site is hosted on Vercel, which keeps standard server logs for security.

Who makes Pipette

Pipette is made and paid for by Ferced, a small software studio from Buenos Aires. We build software for companies. We built Pipette because we wanted it to exist, and we keep it free because science belongs to everyone. If you find a bug or have an idea, write to hola@ferced.com.

ferced.com · Back to today