Frontiers in Research: Open Science

Most science funding, whether government or philanthropic, goes into research projects, or into the buildings, equipment, and services that power them. But the layer underneath all of that — how research is produced, shared, evaluated, celebrated, and reused — is rarely funded directly, and it is the layer that sets the pace.

“Science moves… at the speed of infrastructure… the layer that sits underneath all of this… how research is produced, how it’s shared, how it’s evaluated, how it’s celebrated, and how it’s reused across the system. And… that infrastructure is severely outdated. It’s chronically underfunded.”

— Sandeep Patel, panel transcript

His through-line for the whole session: with AI and shifting research norms, this is a generational opening for philanthropic capital to reshape that underlying layer, and improve the pace and quality of science broadly.

Traditional publishing was built for the print era: a journal filters work before anyone can read it, because page counts were limited. On the web that ordering no longer makes sense. From submission, it takes seven to eight months on average to get through peer review and acceptance, and sometimes two or three years. Preprints simply reverse the order: post the work immediately with a datestamp, then let evaluation and filtering happen afterward.

“Somebody who’s an expert… [says] I don’t want to wait eight months. I can read it tomorrow, make a judgment for myself, and then start the next set of research.”

— Richard Sever, panel transcript

COVID was the stress test. Treatments like dexamethasone were disseminated rapidly enough to change clinical practice immediately; Yale’s Akiko Iwasaki said she couldn’t imagine COVID research happening without preprints.

“One physician came up to me and said, it’s no exaggeration to say that there are people alive today who wouldn’t have been because of preprints.”

— Richard Sever, relaying a physician’s comment, panel transcript

And because preprint servers do no editorial or peer-review evaluation, the cost structure collapses: they are free to post and free to read. bioRxiv and medRxiv run on a few million dollars a year, against a multi-billion-dollar publishing industry.

Biomedical research is hyper-competitive (very few senior positions, a narrow career bottleneck), so people are wary of putting work out before it’s “safe.” A preprint flips that calculus: a public datestamp gives you priority, on your own timeline.

“You put it online, you as the researcher [are] in… total control of the dissemination timing. So it can go online with a datestamp.”

— Richard Sever, panel transcript

The visible effect was at conferences. Before preprints, a great talk might draw the response, “that was interesting data when they did it two years ago; it’s just come out in Cell.” People only felt safe discussing work once a journal had stamped it. After bioRxiv, researchers began posting the day or week before a talk and presenting genuinely current work. Sever pointed to a recent Cold Spring Harbor biology-in-AI meeting where Jennifer Doudna’s keynote alone had three bioRxiv QR codes, and every subsequent talk cited one too, the kind of speed that makes waiting eight months look absurd in a field moving as fast as AI-for-biology.

In the journal process, an editor screens incoming papers (rejecting many at the gate, and each rejection restarts the clock), then sends the survivors to roughly three experts with a 14-day window. Those reviewers are already overstretched, so they routinely miss deadlines and revision cycles stack up, which is how you reach an eight-month average. A good review is thorough, but whether reviews are consistent from one journal to the next is the open question.

Sever’s deeper point is that real verification isn’t a one-time gate at all:

“The veracity of the work, the importance of the work, is something that may better emerge over time… from scrutiny of more people rather than just the three that you could beg to do it within… 14 days.”

— Richard Sever, panel transcript

The cautionary case is the “arsenic life” paper in Science: it passed full peer review at the most prestigious venue, was widely disbelieved almost immediately, and was retracted 15 years later. Sever’s suspicion is that as a preprint, the flaws would have been spotted within days. Conversely, a flawed myocarditis preprint on medRxiv was caught and corrected within 24–48 hours, a flaw conventional review would almost certainly have missed, because it hinged on a wrong population denominator for one region of Canada that only a local would catch.

A lab’s real asset is its accumulated know-how, but most of it never gets written down in a durable, findable form.

“A lot of the knowledge is just fragmented and messy across notebooks, Google Docs, slide decks, Slack channels, annotations, lots of paper notebooks and loose-leaf papers. And the most durable artifacts are the… narrative manuscripts.”

— Joel Chan, panel transcript

Ask what a finding from three years ago was, or which protocol set up an experiment, and you have to find the right person: “and if that person leaves the lab, then you’re sort of often hosed.” Discourse graphs are a way of capturing the intermediate units that usually get lost — the driving question, the hypothesis, the individual results, and the data and protocols behind them — so they can be found, shared, and built on, not just the final paper.

In labs that adopt this, knowledge starts to compound. Undergraduates navigate the lab’s open questions and requests, pick something that interests them, and produce a documented result within one to three months, with attribution attached from the start. A null result from years ago resurfaces to set a new experiment’s parameters; a departed student’s finding still motivates active work. Project context Across 10+ labs, 6,000+ findings have been captured; the deepest dataset (the Akamatsu lab at UW) holds roughly 2,000 findings and 434 open “requests for experiments” over about 40 months, of which ~29% have been claimed and ~38% of those have yielded documented results. (These project totals are context, not figures spoken on the panel.)

Almost everything we know is locked inside PDFs: long narrative papers. To build on someone’s work, you (or, increasingly, an AI) read the whole paper, dig out the one finding you need, and copy it into your own notes. Every lab does this, over and over, for the same findings. Modular research tools let scientists share the individual pieces — a single finding, the data behind it, the method — so others can find and reuse them directly. Joel’s analogy: imagine there were no grocery stores, so every time you cooked you had to buy finished restaurant dishes and pick the ingredients back out. That’s science today; modular tools are the grocery store: the ingredients on the shelf, ready to use.

The waste is visible whenever the field tries to synthesize. Joel described the recurring frustration at a workshop on AI-assisted systematic reviews:

“What are we doing? Why are we putting all this work into pulling these things out of these PDFs and then squeezing that back into a PDF again, and then… rinse and repeat?”

— Joel Chan, panel transcript

The deeper point: those findings were never not modular. The scientist who ran the study already had the claim, the data, and the method as separate things; the paper is just where that structure gets melted down, and everyone downstream pays to reconstruct it. The fix isn’t faster extraction; it’s not throwing the structure away in the first place.

Yes, especially when the representation is paired with new kinds of dashboards or social networks built on top of it. Instead of advertising yourself with 20 published papers, you expose your open hypotheses, your requests, and a network of connected work, and tools help match you to people asking the same question, often in a different field, or holding the one protocol you need.

“I can… use tools to help me find other people who are interested in the same question, but might be in a different field, or [have] a critical… method or protocol that will help me with my request for the experiment… they start off not cold.”

— Joel Chan, panel transcript

That matters most in interdisciplinary areas, where the right collaborator is someone you’d never run into at a conference and whom keyword search over papers can’t surface.

Ronen did his AI PhD during COVID, largely on Twitter, and says he might not have finished it without the community he found there. That experience led him to a thesis about how invisible this all is to science:

“Social media is really this kind of under-theorized dark matter in science. It exerts a lot of influence, and it shapes collective attention, prestige, [and] researchers’ personal intellectual journeys.”

— Ronen Tamari, panel transcript

The striking part: these effects are an accidental side effect of a platform never designed for science. Done intentionally, with the right design and engineering, there is far more on the table. The AT Protocol (the open layer beneath Bluesky) is where that experimentation is now possible, and the question Ronen is pursuing is how to treat this activity as research infrastructure rather than ephemeral platform content.

“If it’s on LinkedIn or on Twitter, there’s really very little we can do with it… if we take this seriously, [we] start thinking about how social media should actually be… treated as research infrastructure.”

— Ronen Tamari, panel transcript

One researcher posted a question and some experimental evidence on Twitter and asked for help. Within a few weeks, 150+ researchers were collaborating on what they called a “Twitter-fueled open global drug discovery effort,” producing preclinical antiviral candidates for COVID.

“They called it a way of working none of us realized was possible.”

— Ronen Tamari, on the COVID Moonshot, panel transcript

It is rarer than you might hope, and it has grown rarer still. After Elon Musk bought Twitter in 2022, researchers began leaving, API access grew expensive, the third-party tools people had built broke, and the community fragmented. The upside is that open platforms now create room to rebuild that collaborative energy deliberately.

The panel’s shared answer: no single unit wins. Papers, preprints, discourse-graph nodes, and social curation serve different purposes and audiences. Sever’s framing is a constellation:

“The way I like to picture it is… like a constellation. You’ve got a paper… [and] that constellation links out to many pieces of different data in other places. There might be a pre-registration, there may be peer reviews, maybe a big dump of data… if you’re interested in the genome sequence, then… that’s the center.”

— Richard Sever, panel transcript

The center of the constellation shifts depending on who’s reading and why. And we should drop the fiction that the millions of papers published each year are being read cover to cover:

“The brutal truth… is most papers are not read by even the people who open them to read them.”

— Richard Sever, panel transcript

Sever’s point: an AI agent can do what almost no human does — read a 40-page paper end to end, every figure — and then take a human down a tunnel to the one piece of data they need, or back up to a higher-level summary. Joel added that coding agents have shifted the economics fast, roughly since late 2025:

“Agents don’t get tired. So they can… have the formal structures in there; if [they’re] formal structures, you can check them.”

— Joel Chan, panel transcript

He pointed to an emerging spec (“Astra”) for agent-produced results that are structured rather than narrative-shaped, and to a paper provocatively titled “The Last Human-Written Paper.” Crucially, narratives and modular pieces aren’t in competition: a narrative is still a good way to say “this is a bundle of work I feel confident about,” shipped alongside the constellation of pieces that let others verify and synthesize over time. Ronen added the durable counterweight: expert curation survives AI.

“Even the AGI people on Twitter, they’re still sharing papers with each other… getting their best reads from social media. So the curation is really still crucial.”

— Ronen Tamari, panel transcript

Ezra argued that narratives are part of the human experience of science and worth preserving, which means we may need to make papers easier to read and discuss, since people increasingly lean on AI summaries instead. Sever agreed we’ll always need narratives; what changes is the ratio.

“The narrative is more than the data, and the data are more than the narrative. They’re different things, and what we may see is the ratio change.”

— Richard Sever, panel transcript

A Nature News & Views piece covering three advances is read by far more people than the underlying papers; that’s a narrative-to-data ratio we already accept. Meanwhile, many papers function less as communication than as career currency:

“They are being generated as… career currency, as demonstrations of output. They are not… being written with the idea that many, many people are reading them first to last.”

— Richard Sever, panel transcript

Sever’s first answer: we’re not correcting the record fast enough. Whole swaths of the literature are quietly wrong. For instance, thousands of papers used an antibody everyone believed targeted one protein when it actually targeted another of the same name. Faster, more open dissemination catches these sooner. His favorite illustration:

“The late Craig Venter… published a paper in PNAS… saying that his group could predict what you looked like based on your genome sequence. And that very afternoon… that data had been completely reanalyzed by Yaniv Erlich at the New York Genome Center. And he had posted a paper on bioRxiv saying the work was wrong.”

— Richard Sever, panel transcript

Joel’s answer was about synthesis and cumulative theory-building, which is throttled by how scattered the evidence is. A systematic review takes 12–18 months for a single narrow question; the UK’s Foresight obesity map took nearly a year, was never linked directly to its evidence, and was out of date almost as soon as it was used; the IPCC is openly worried that climate synthesis can’t keep pace as the literature grows exponentially.

“On average… systematic reviews these days… take about 12 to 18 months to do for… a single, narrow question.”

— Joel Chan, panel transcript

Not more papers; we’re drowning in papers. The honest target is reuse: how fast a good finding gets picked up and built on. Four dials are worth watching: rate of reuse, rate of self-correction (do wrong results get caught?), real-world decisions (does a policymaker actually act on a finding?), and how fast validated findings get shared at all. The place the new stack compresses most is the gap between “I have a result” and “someone builds on it,” today often measured in years.

And acceleration has a second half that AI made urgent: trust. AI made producing plausible science nearly free, while checking whether it’s true (and whether it matters) got no cheaper. Acceleration isn’t only making findings faster; it’s being able to trust them faster. Project context The scale of the synthesis bottleneck is large enough that the Wellcome Trust and ESRC recently committed ~$100M to evidence-synthesis infrastructure, and a Lancet working group estimated $100B/year in research time wasted partly on redundant studies timely synthesis would have prevented.

Ronen: the AT Protocol science ecosystem is full of energy and badly under-resourced: 20+ projects, mostly volunteer-run, many built on weekends. He noted a striking convergence: “the properties that make social media protocols compelling are the same properties that the research communities actually want.” Grants would let people pursue this seriously, and he has curated a portfolio of them, MOSAIC (Modular Open Science, AI and Collective Intelligence), to surface these projects to funders on Catalyze.

Sever’s answer was blunt and structural:

“Infrastructure, infrastructure, infrastructure… The science publishing industry is a multi-billion-dollar industry. bioRxiv [and] medRxiv cost a few million dollars a year to run… In terms of… bang for your buck in changing the world, investing in this type of infrastructure… is really good.”

— Richard Sever, panel transcript

The point is leverage: bioRxiv and medRxiv are the substrate that lets every downstream experiment — decoupled review, trust signals, modular reuse — exist at all. He credited Chan Zuckerberg Initiative funding for enabling 35,000+ COVID papers to go up on the servers during the pandemic.

Joel pointed to two concrete units of funding. The first is convening: meeting funds and hackathons that put scientists, builders, and standards developers in the same room. He and Matt had run one, the MIRA workshop, the week before.

“You can fund meeting funds or hackathons; we just ran one last week, where you had scientists, builders, and standards developers all in the same room, prototyping. That’s the short term. Longer term, you seed social change by encouraging, as a condition of funding, that research experiment with different ways of sharing your work.”

— Joel Chan, panel transcript

That second unit is the funder pilot: making “experiment with new ways of sharing your work” a condition of the grant, so new norms get seeded rather than left to chance.

Rajesh pushed the panel past tools: the harder skill is generative, systems-level collaboration (within a lab and across labs), and that should be invested in too. The panel agreed completely. Joel pointed again to hackathons and to community norming work (PREreview, the bioRxiv community) as infrastructure in its own right. Ronen connected it back to the lineage of these tools:

“There’s some testimonials from researchers that have been using them and they report huge improvements in thinking and doing of science… it’s tools for thought they’re called, and they… help you think.”

— Ronen Tamari, panel transcript

Sandeep closed the session on the same note: a funder who cares about curing a disease should care about the infrastructure underneath the science that gets there, because you don’t cure a disease without researchers sharing, collaborating, creating knowledge, and vetting it.

The framing the team came ready with: the paper is a great story and a terrible database. Keep it as the narrative front page; stop using it as the storage format. Concretely, from scratch you’d build:

• Addresses for findings: modular, citable identifiers (think DOIs, but for the individual results inside a paper), with versioning, indexing, and provenance. Today we’ve only minted addresses for whole papers, so the pieces inside can’t be linked or tracked; a concrete, fundable gap.
• “Killer experiment” dashboards that match the experiments most worth running to the people who can run them.
• A lab-notebook layer that captures human and AI reasoning behind a result; today those agent conversations just vanish, and the trail of who and what contributed is lost.
• “Interpretable RAG”: a searchable library of individual results, with data and provenance, that AI agents can navigate reliably, instead of chopping papers into arbitrary chunks and guessing.
• A credit / funding layer that routes resources to the findings everything else depends on.

The connective tissue is a stack: Joel builds the tools scientists use to capture work (the lab bench), Ronen builds the infrastructure that stores and connects it (the plumbing), Richard works on publishing, how it reaches the world (the front door). Where they connect: a bioRxiv paper can be the human-readable front page of a machine-readable web of findings underneath, so you keep the story and get the database.