New Science Manifesto

00 // INTRODUCTION

Introduction: the paper is not the final form of knowledge

We too easily mistake the scholarly paper for scholarship itself. A young scholar proves competence by publishing papers; a field confirms knowledge by turning it into papers; universities, funders, and promotion systems count papers as the visible units of academic life. Over time, the paper stops looking like one way of expressing knowledge and begins to look like knowledge's passport.

That is a historical illusion. The paper did not fall from the sky, nor is it the natural shell of reason. It is a media form produced by print, postal networks, learned societies, libraries, subscriptions, citation systems, and academic evaluation. In other words, the paper is not the ontology of knowledge. It is the shape knowledge acquired in one media environment so that it could circulate, be certified, be archived, and be disputed.

The year this manifesto is written, 2026, is the year the question was forced into the open. In June more than four thousand mathematicians signed the Leiden Declaration, asking for transparent disclosure of AI use and insisting that credit and responsibility stay with humans. In August Terence Tao, at the International Congress of Mathematicians, split problem-solving into five stages: generating, verifying, explaining, digesting, and canonizing, and warned that mathematics is moving from proof scarcity to proof abundance. In September twenty-five Fields medallists declared a severe misalignment between AI companies and the mathematical community, because solving problems is only a tool and proxy for understanding. Between these statements, a 23-year-old with no advanced mathematical training handed a sixty-year-old Erdős problem to a language model and solved it. If the oldest discipline built on proof and authorship is re-asking what a result is, who counts as an author, and what counts as finished, no other field can pretend the question is not its own.

The claim of this manifesto is simple: the paper remains important, but it should no longer monopolize the form of scholarly output. The AI era calls for a richer practice: moving from paper publishing to multimodal scholarly curation. The result is a richer object: multimodal, traceable, reproducible, conversational, versioned, and readable by both humans and machines. We call this object a knowledge object; in ordinary language, it is simply a better curated scholarly work.

01 // PRINT

1. The great print-era invention

Modern scholarly publishing grew out of the print revolution. Gutenberg's Bible is often treated as a turning point in Western movable-type printing because mechanical reproduction gave texts a new social life: stable editions, page references, wider circulation, and shared objects for argument. In the seventeenth century, print, postal systems, and learned societies combined to create the scientific journal. In 1665, Henry Oldenburg launched the Royal Society's Philosophical Transactions, one of the oldest continuously published scientific journals.

The journal was not merely a container for finished results. It was a communication apparatus. It gathered letters, experiments, observations, disputes, and reviews; editors selected and arranged materials; printers made them portable; postal networks distributed them; libraries preserved them; citations connected them to future work. The scholarly paper became the form that fit this apparatus.

The paper's virtues were real. It compressed a research process into a bounded text that strangers could read and judge. It had a title, authorship, an abstract, argument, evidence, conclusion, and references. It was short enough for journal pages, stable enough for archives, standardized enough for review, and dense enough for expert readers. It helped scholarship escape dependence on personal acquaintance and local correspondence.

02 // DIGITAL

2. Digital reading changed the unit

In the digital era, the material assumptions of print journals have weakened. Scholars rarely begin by reading an entire journal issue. They search databases, follow DOI links, move through citation chains, consult preprints, use institutional repositories, and increasingly ask AI systems to summarize or compare a literature. The journal remains important as a signal of review and reputation, but it is less often the actual medium through which readers encounter a work.

This is a structural change. The basic unit of print was the issue. The basic unit of digital scholarship is the object: a paper, dataset, code repository, video explanation, lab record, model weight, review history, or version. If the medium can carry and connect all of these materials, why should the scholarly output still be reduced to the format most convenient for paper pages?

Open access data points toward the shift. Piwowar and colleagues estimated that roughly 47% of papers published in 2015 were openly available. The STM Open Access Dashboard reports substantial growth in article output and gold open access between 2014 and 2024. OpenAlex describes an index containing hundreds of millions of scholarly works. No individual reader can navigate this scale by reading journals issue by issue.

03 // WHAT GETS CUT

3. What papers cut away

The paper's density is both strength and limit. It lets peers move quickly because it assumes a shared background; it saves space because citations point elsewhere; it makes review efficient because arguments arrive in a familiar structure. But these same virtues exclude newcomers, adjacent fields, and the public.

The deeper problem is that traditional papers turn research into a success narrative. The published article usually preserves the clean path: a question, a method, data, analysis, and conclusion. Actual research includes failed experiments, wrong hypotheses, parameter searches, discarded data, method changes, ethical hesitation, equipment trouble, collaboration conflict, and changed judgments. These materials are often treated as noise and cut away.

For reproducibility, however, the cut material may be crucial. To repeat a result, one often needs to know what failed. To judge data quality, one needs to know what was excluded and why. To extend a project, one needs to understand why alternate paths were not taken. The reproducibility debate, positive-results bias, and growing retraction records all point to a common lesson: the literature cannot be only a library of successful stories.

Digital platforms can preserve, connect, and search these layers. AI can help readers query them. The obstacle is no longer primarily technical. It is institutional: we still behave as though a scholarly result becomes real only when compressed into a polished paper.

04 // THE FLOOD

4. AI redistributes scholarly labor

AI sharpens the mismatch. Literature review, summarization, translation, formatting, first-draft synthesis, and even parts of peer review are increasingly machine-assisted. Research output was already growing quickly; AI now lowers the cost of producing paper-like texts. When AI can write, review, and summarize papers, the human role becomes awkward if we define scholarship mainly as paper production.

This does not mean humans leave scholarship. It means we must relocate the human contribution. AI is strong at retrieval, compression, pattern recognition, formatting, and producing plausible drafts. Human scholars are responsible for asking worthwhile questions, judging value, sensing anomalies, carrying ethical responsibility, and bringing embodied curiosity and taste to inquiry. AI can simulate an interest; it cannot truly have a stake in a question.

Two years on, those figures already look conservative. In October 2025 Stanford hosted Agents4Science, the first conference in which AI systems were the primary authors and reviewers: 315 submissions, 48 accepted; the three model reviewers agreed with one another only weakly, and more than half of the submissions contained at least one reference flagged as problematic. Tellingly, fully AI-driven research made up 23.3% of submissions but only 14.9% of accepted papers: the more humans shaped the question and the design, the better the work held up. In February 2026 Nature reported that computer science was drowning in AI slop, with one researcher generating a fictional-experiment paper in fifty-four seconds; arXiv had already stopped accepting unreviewed survey and position papers in late 2025. The deeper change is in mathematics. At ICM 2026 Tao described a field moving from proof scarcity to proof abundance: AI generates proofs faster than humans can verify them, verification outruns readable exposition, exposition outruns peer review, and publication outruns the slow work of digestion into textbooks. He called it proof indigestion, and reminded us that our institutions were designed under the assumption of scarcity. The bottleneck has moved from writing to verifying, explaining, digesting, and evaluating, and none of these is what the paper as a genre does best.

The goal, then, is not to use AI to produce more papers faster. The goal is to use AI to redesign scholarly outputs so that human attention moves toward problems, judgment, imagination, and responsibility.

This is also why visual and public entrances are not decorative extras. Human taste and inspiration can only enter research that remains visible and perceivable; cross-disciplinary breakthroughs happen only when outsiders can enter the context. Multimodal curation is the infrastructure through which human sensibility takes part in inquiry.

One might push further: what if AI could one day advance frontier research entirely on its own, leaving humans no substantive contribution to make? Would visualization and public explanation then lose their point? Precisely the opposite: they would matter more than ever. Humans would still need to know what AI is investigating, whether its findings serve human interests, and whether it is approaching risks we should refuse to accept. If autonomous AI research turned science into a black box no human can see through, we would have opened Pandora's box with our own hands — unable even to tell whether what flies out is blessing or disaster.

Keeping science visible, perceivable, and accessible is therefore a long-term commitment that outlasts any current division of labor between humans and AI. Today it lets human researchers take part through taste and inspiration and makes cross-boundary exchange possible; tomorrow it may be humanity's last window for understanding and overseeing machine-driven science. Whoever produces knowledge, science must remain within human sight, human feeling, and human reach. This is the deeper reason for scholarly curation in the AI era.

05 // CITIZEN SCIENCE

5. Citizen science: not lower standards, but divided roles

If AI has made papers cheap to produce, it has also made something else possible: for the first time the threshold of research is low enough for ordinary people to pick it up and play. In April 2026 Liam Price, 23 and without advanced mathematical training, typed an Erdős problem into GPT-5.4 Pro on an idle Monday afternoon, unaware that it had stood for sixty years; his collaborator Kevin Barreto, a Cambridge undergraduate, recognized what the output meant and alerted experts. Tao later observed that everyone who had worked on the problem began with the same standard sequence of moves, while the model took a different route, using a well-known formula that nobody had applied there; experts distilled the rough output into what one called a Book proof. Scientific American named the practice vibe-mathing: pick problems from a database, let the model try, explore by feel. By August 2026 the Erdős problems site listed 565 solved and 652 open, with amateurs, professionals, and AI systems working in the same comment threads.

This is not a mathematical exception but the continuation of a longer tradition. In 2011 Foldit players solved, within three weeks, the structure of a retroviral protease that had eluded scientists for a decade, and the result appeared in Nature Structural & Molecular Biology. In 2014 thirty-seven thousand EteRNA players were co-authors of a PNAS paper on RNA design. Zooniverse, which grew out of Galaxy Zoo, passed three million volunteers in 2026 and one billion classifications. As early as 2009 Timothy Gowers asked on his blog whether massively collaborative mathematics was possible, and dozens of participants, from professional mathematicians to interested amateurs, produced a new proof of a combinatorial theorem in the comments. Tao has carried this line into the AI era: he explicitly advocates citizen mathematics, in which undergraduates, amateurs, and interested people contribute modular pieces of large collaborative projects, such as literature searches, numerics, and verification steps, and he notes that proof assistants let collaboration scale from a handful of trusted colleagues to something like a modern supply chain, because the program, not trust, guarantees correctness.

But citizen science does not lower the bar; it divides the roles. When AI can run a full research cycle on its own, people still hold at least five positions. The terminator decides when to stop: an AI can chase a conjecture indefinitely or set up experiments and dispatch robots, but whether to keep spending compute or to accept a safety risk is a human weighing of resources and danger. The director sets direction and draws boundaries, like the human-written program file in Karpathy's autoresearch project, where a person no longer writes code line by line but orchestrates agents that run a hundred experiments overnight. The supervisor watches ethics, funding, and disclosure. The associator supplies ideas: AI has imagination, but the best human intuition is still sharper, and an idea handed to a model is completed and tested within minutes, so that it can be pursued or dropped. As Fields medallist Yu Deng put it after his award, AI can fill in the technical details of a proof so that work humans could already do gets done faster, but what really moves mathematics forward is human intuition and ideas: asking a good question, and then having a line of attack. The fifth position is the largest and the most overlooked: the audience, who may contribute nothing substantive yet supply the steady feedback research needs by watching, proposing, voting, tipping, and monitoring. Each gives little; together they are a force.

Who would watch other people do research? Game streaming faced the same doubt in the 2010s, and it turned out that vast numbers of people happily watch others play. AI has outplayed humans at Go since 2017, yet human interest in Go did not die; commentary and post-game analysis became popular genres. Research need not beat the machine to be worth watching. It can be played like a turn-based strategy game, in which the researcher surveys the board, allocates resources, issues orders, and lets the AI run a round before intervening; like a visual novel, mostly watching and choosing, with the option to rewind to an earlier save after a failure; like a battle game, in which AI pushes the lanes and the human adjusts strategy; or, most often, by not operating at all and narrating while the AI plays.

Citizen science has its traps, and 2026 has exposed them clearly. Thomas Bloom, who runs the Erdős problems site, worries that AI is being used by many people who are not mathematicians, so that nobody has read the proof and nobody will. More than seven hundred mathematicians signed an open letter against a forty-hour AI mathathon at Caltech, arguing that rushing open problems through language models produces slop mathematics and dumps the verification burden on the community. The Leiden Declaration welcomes the broadening of the community while asking newcomers to respect its values. So the precondition of citizen science is not that everyone can publish, but that three things are built into the platform: every item entering the evidence layer has a clickable source; every result is explained, in their own words, by a person who takes responsibility for it; and every AI contribution is recorded and disclosed. Tao offers the measuring stick: a proof that no human can properly explain should be viewed as incomplete, even if it has been formally verified. What the threshold lowers is the operation, not the standard.

If citizen science is meant seriously, it needs a form that can be run for the long term, and the most mature precedent for that is not in academia but in large online games. A game that has run for twenty years never relied on the best graphics engine, which is replaced every few years, but on what players accumulated inside it: characters, achievements, guilds, and friendships. A research platform should work the same way. You can grind alone, taking an AI party to a small problem; you can raid as a team, dozens of people dividing an open problem among themselves as the Polymath projects did, except that this time an engine keeps the books, checks the steps, and assigns the work. You will compare yourself with others: whose problem was harder, whose evidence trail is cleaner, whose explanation was understood by more readers. You will collect trophies, and you will build a home of your own. What these accumulations have in common is that none of them depends on any particular generation of models. Models iterate and prompts go stale, but what a person has done in the community, the trust earned, and the work left behind do not reset with a version number. This is why we say the platform's real moat lies outside the technology: AI tools will always depend on the development of large models, while the stickiness of a community, the accumulation of achievements, and the settling of identity can only come from time and people.

06 // CURATION

6. From papers to scholarly curation

If the paper was the basic scholarly object of the print era, the AI era needs a different practice: scholarly curation that produces open, layered, multimodal, traceable, reproducible, conversational, and version-managed knowledge objects.

A knowledge object is not a longer paper with attachments. It is a small knowledge environment around a research result. It contains a paper layer for expert argument and citation; a public layer with images, posters, videos, and question maps; a learning layer with concepts, background, and reading paths; a reproducibility layer with data, code, logs, failed routes, and method changes; an oversight layer with funding, ethics, safety, conflicts of interest, and licenses; a machine layer with Markdown, JSON, structured metadata, APIs, and knowledge graph hooks; and a version layer with comments, commits, forks, replications, and responses. Producing such an object is scholarly curation: selecting, arranging, documenting, and maintaining the materials through which a work can be understood and tested.

Different readers need different entrances. A member of the public may begin with a poster or short video. A student may begin with the learning path. A peer reviewer may read the paper layer. A replication team may inspect data and code. A policy maker may inspect ethics and funding. An AI system may read structured metadata. These entrances should be connected, not collapsed into one PDF.

07 // JOURNALS

7. Journals as certification and curation

The knowledge object does not require journals to disappear. Many journals hold real trust, editorial craft, peer networks, and public recognition. What changes is their role. A journal should no longer understand itself mainly as a container for papers. It can become a certifier, curator, guide, and public window for knowledge objects.

Peer review can remain central, but it should become more traceable. Review questions, author responses, added data, replication attempts, corrections, and later disputes can become visible layers of the same object. Publication should not mean freezing a text; it should mean admitting a research object into public correction and continuing maintenance.

Print can also remain, but as a curated front end. A new journal could look more like a visually rich magazine: attractive, selective, and readable, designed to bring complex work to broader publics. The rigorous, technical, and reproducible layers would live on the platform. In this sense, the future journal is less a container and more an institution of scholarly curation.

08 // WENDAO & BLUEPRINT

8. Wendao: AI assistance without replacing the scholar

RenaSci also introduces an AI-assisted research suite. Its aim should not be to ghostwrite papers. If AI assistance merely increases paper output inside the old system, it will worsen the flood of low-quality texts, review congestion, and shallow synthesis.

Wendao, the proposed suite, should help scholars enter knowledge rather than bypass it. It can turn difficult papers and books into concept maps, question trees, debate maps, method guides, and key-reading prompts. It should save mechanical organization time without replacing slow reading and thinking.

It should also help researchers discover their own concerns: What question keeps troubling you? What anomaly are you noticing? What does your disciplinary training hide? What problem are you willing to carry over time? From there, the suite can help gather sources, organize evidence, test claims, draft a paper view, and generate public explanations, data descriptions, ethical notes, media assets, and machine-readable publishing packages.

In concrete form, Wendao should be a turn-based game engine rather than a chat window. The researcher writes a program file first: the question, the constraints, the no-go zones, the evaluation criteria, the budget. An AI party then plays one turn: a retriever finds literature, a reasoner sorts the approach, an experimenter runs code, an illustrator draws diagrams, a reviewer picks holes, a scribe keeps the log. The turn ends with a readable, visual report: a progress score, new evidence, risks, open questions, and a few branches to choose from. Having read the report, the researcher can issue the next orders, rewind to an earlier save and take another path, or call a halt and package the run as a knowledge object for the platform. Every save is a version, so version history is no longer written after the fact but produced automatically by the interaction, and every AI contribution, which model, which sources, how much compute, enters the disclosure record by itself. Spectators can watch the whole run, propose in the chat, and vote on branches, but their voices enter only the report's atmosphere layer and become orders only when the researcher adopts them. People set the direction, AI plays the turns, and the report is the version: Tao's four stages of generating, verifying, explaining, and digesting are compressed into a single workflow.

This leads to the fourth piece of the blueprint: the scholarly card. Today's academic homepage is a dull list of papers, journals, and citation counts, topped with an impact factor; it tells you how much a person has published and nothing about what they have achieved. The card we propose lists achievements instead: which problem was solved, which role was played in which raid, which work was understood by how many readers, which evidence trail survived replication. The card also generates a multimodal scholarly biography, in text, diagrams, timelines, a map of relationships, and explanations that can be heard and watched, compiled by the person together with AI and updated automatically as work accumulates. The same method can curate a well-known scholar: a multimodal exhibition of their achievements, their place in history, and their network, so that those who come later need not guess who they were from a list of papers. Cards and curated pages are open to comment and rating by the platform's users and its AI, and what is rated is the achievement and the work themselves, not the count or the name of the journal. Academia has long noticed the absurdity of paper counts and impact factors: in 2012 the San Francisco Declaration on Research Assessment called for an end to judging individual researchers by journal impact factors, and in 2015 the Leiden Manifesto for research metrics set out ten principles, the first of which is that quantitative evaluation should support, not supplant, qualitative expert assessment. We are simply turning the demands of those two declarations into a product anyone can use: a scholarly identity measured in achievements, accumulable, displayable, and open to mutual review.

09 // THESES

9. Fourteen theses

First, the paper is a historical invention, not the destiny of knowledge. Second, journals should move from print containers to certification, curation, guidance, and public display. Third, scholarly outputs should move from paper publishing to multimodal scholarly curation: better curated knowledge objects rather than isolated PDFs. Fourth, reproducibility should be designed into the output from the beginning. Fifth, failed and negative results should be recorded, searchable, and valued. Sixth, AI should be a transparent assistant, not an anonymous shadow author.

Seventh, the human contribution in the AI era lies in question-making, value judgment, embodied experience, ethical responsibility, anomaly perception, and imagination. Eighth, the public, students, and adjacent fields are legitimate readers, not outsiders. Ninth, scholarly objects should be friendly to humans and actionable for machines. Tenth, publication should begin public revision rather than end inquiry. Eleventh, evaluation should move beyond paper counts and journal rank toward trustworthiness, openness, reproducibility, explanatory power, public value, and future reuse. Twelfth, RenaSci connects the platform, journal, and Wendao suite into one scholarly curation system.

Thirteenth, science should be playable, watchable, and joinable. What AI lowers is the threshold of operating research, not the standard for judging it; some play and everyone may watch, but every result must be explained by someone, checkable by anyone, and owned by a responsible author.

Fourteenth, scholarly reputation should be an accumulable, displayable, mutually reviewable record of achievements, not a count of papers. The platform's real moat lies in people rather than models: as in a long-running online game, what a researcher has done in the community, the trust earned, and the work left behind should not reset with any generation of AI.

10 // CONCLUSION

Conclusion: not abolishing papers, but freeing scholarship

This is not an anti-paper manifesto. It is an anti-paper-exclusivity manifesto. Books still exist after the internet; print journals can still exist after digital publishing. What must change is our imagination of what a scholarly output can be.

The print era freed knowledge from manuscript scarcity and private correspondence. The digital era freed knowledge from paper distribution and created databases, preprints, open access, and search. The AI era should free scholarship from the single-text object and create outputs that are multimodal, traceable, reproducible, conversational, and capable of evolving.

Nor do we intend to turn science into a show. On the contrary: when every turn of a study is visible, every piece of evidence can be opened, and every AI contribution is recorded, science is more honest than it was behind the paper. Science once belonged to the few who could read papers; it can belong again to everyone willing to ask, watch, and judge.

We should not make papers more paper-like. We should make scholarship richer, more transparent, more inspectable, more understandable, more revisable, and more responsible. The paper was a great invention of the print era. It should not become the final idol of the AI era.