Cofounder of @paperpile. Prev: staff eng, director @google, @verily. Comp bio Ph.D. turned software engineer. Builds tools for scientists.
Cambridge, MA
Joined May 2008
- Tweets731
- Following445
- Followers314
- Likes943
Great conversations are starting to happen around how biology will adapt to the new reality that a good-looking paper is no longer a signifier of actual scientific effort.
We think a lot about this from the researcher's perspective at @paperpile. There's a reason why we haven't built any "write my paper for me" type tools that are aggressively advertised to grad students.
Early-career researchers are excited about AI search tools making a dense and obscure literature more accessible. But they're worried about de-skilling, and the unpredictability of how our systems of publishing and promotion will change over the next few years.
More established scientists see AI as an opportunity to leverage their expertise: becoming more ambitious in their data collection & extraction, reducing time spent on admin boilerplate, and shortening the idea-to-validation loop—especially in computational fields.
Every scientist I've talked to still believes in reading the full text of papers that matter to their work.
But this attitude depends on a shared understanding that a paper's authors put time and effort toward crafting an accessible, meaningful, and ideally reproducible description of their work for an audience of peers. If that assumption breaks down, things may change quite rapidly.
Discussion-starters like Blekhman's essay are important to start mapping out a highly-uncertain future.
What is biology research for?
Expanded version of my PLOS Biology Perspective, with very recent developments from arXiv and bioRxiv. This field is moving fast.
blekhman.substack.com/p/what…
This is cool! Probably not something I'd expect researchers to use on a daily basis (at least for now, folks I talk to really do care about deeply reading the author's words, esp for a paper in their field of study) but I deeply admire the exploration of alternative and more dynamic interfaces for reading and doing science.
Since so many of you loved this concept, let me answer some questions
— many of you wanted to keep the original PDF layout, and that's totally possible (post follows)
— GPT 6.1 Sol Ultrafast makes this a near instant (albeit expensive) experience. It also works with normal models, you just get to see a skeleton at times
— open source is coming, and my app is too! I had posted this as an interaction concept, but since many got inspired, I'll share it
(I know it's been a week, I've been busy, sorry)
1
2
193
I feel like you could overlay this with a graph of GitHub PRs over time and they would match exactly.
So many parallels in the spam problems facing GitHub and preprint servers in 2026.
Incentivizing or requiring more "how the work was done" style sharing, e.g. notebooks and work histories, would be a net benefit to science in the long run. (Data and code requirements are a good step in that direction.)
The unfortunate flip side is that the real story behind any publishable piece of science is usually messier, and less circumspect, than anyone would want to share publicly!
I absolutely agree. In our new NIH-funded research we try to adhere to this as much as possible. Ideally, data should be released the moment the result is there. The paper can serve as a sort of summary, but it should not be the primary outcome, because it emphasizes storytelling more than the actual results.
This would be super easy if all scientists had a digital lab book that could be traced. AI could easily make this accessible to everyone, at least whatever can be shared. And @NIH could (and should) mandate this very easily. This would also help to make scientists more accountable to the research they proposed to do in their grants. That helps with reproducibility and many of the other problems.
We will also try to implement and incentivize as much of this as we can with @ResearchHub, and keep making research more open and easily available.
20
It's crazy to see these numbers still going up.
Software like @paperpile gives you peace of mind, because we're absolutely obsessed with getting every bit of metadata correct.
Fake citations aren't a thing when your library is a reliable source of truth.
With our brand-new MCP server and Claude/ChatGPT integration, your agent can call Paperpile to find & verify citations, giving the best of both worlds.
After looking at fabricated references in biomedical literature earlier this year, researchers evaluate 500,000 science, engineering and computing papers.
retractionwatch.com/2026/05/…
pnas.org/doi/10.1073/pnas.26…
Greg Jordan retweeted
Writing a literature review in 2026: 40 minutes asking AI to summarize, 4 hours checking whether the papers exist.
Interesting AI-era update of a classic human experiment!
Also this week, Gary King from Harvard published gking.harvard.edu/peerless/ reaching almost the same conclusion: today's AI models outperform high-quality human peer reviewers.
Both are worth a read (and appropriate amounts of skepticism). King and Worthington's discussion is better, annd according to Pangram, uses less AI than the Rust team's study—not that there's anything wrong with that!
The implications for publishers are complicated, but for authors the upshot is clear: if you're *not* using Claude Code to self-review your work before submission, you're putting yourself at a disadvantage.
Lots of debate this week on whether journals should ban AI in peer review.
We actually tested this in our new preprint, and the results were pretty clear to me.
For context, I've reviewed for most of the big journals and evaluated 1,000 human reviews as an editor. I love to read and evaluate science.
TLDR: From now on I won't do any peer review without AI. For clinical trials, I'd even call human-only review dangerous 🧵
2
54
Cool paper!
Brings together many themes across my career: provenance for science workflows, citizens' access to data and compute, and MCP servers for agent automation.
In the software world, a change isn't "real" until it's been merged to main and post-commit tests have passed.
This paper shows: in science, a result isn't "real" unless you've got logs and verifiable provenance attached.
🚨 AI coding agents are about to flood science with irreproducible results—unless we move them out of private laptops and into public glass boxes.
Here is the proof from our new study across 48 datasets:
💥 The problem (Private Terminals):
We gave 1 published dataset to 30 AI agents.
→ 29 went off-script into 18 incompatible pipelines
→ Silently erased a real bacterial toxin gene
→ Left behind 26 throwaway scripts nobody could verify
🔓 The fix (Public Infrastructure):
We wired coding agents directly into @galaxyproject & free national supercomputers—where an agent cannot hide its tracks:
🔎 Caught hidden flaws in published papers:
Exposed that a recent Nature agent benchmark accidentally ran on a 1,000-cell demo file instead of the full 5,613-cell atlas (82% of data dropped). Running all cells on Galaxy 3× accuracy.
⚡ Supercomputer discovery at zero cost:
• Unified 2,592 malaria surface proteins across 3 continents into 21 shared 3D folds on free public GPUs
• Dated 42 pathogen epidemics in seconds instead of days
📄 Preprint: biorxiv.org/content/10.64898…
📊 Interactive story: spond.github.io/CV.js/blog/2…
2
137
New research: experts' opinions are unconsciously swayed by AI agents' outputs.
One solution is to change what the agent does for you. It should be your librarian or assistant, not your PI.
Let it find and fetch papers, and point to the important parts.
Then read them yourself. And ask it to argue with you!
Good policy. I hope the added effort isn't too much of a drag on their resources!
We're also starting to see some of the fighting fire with fire effect (using AI tools to help detect AI junk at scale) that feels inevitable in a spam-adjacent arms race.
How we are approaching AI at openRxiv - new blog post prompted by growing adoption by researchers, submission patterns, and questions from the community. 1/n
3
107
All of this, but put it all into a .zip file with a small structured manifest so it's easy for readers or AI tools to download and work with the data.
Scientists don't want to visit yet another website. They want to quickly get to the paper, data, and code.
What should an open scientific paper look like in the AI age?
We made a demonstration project out of our paper on the childhood socioeconomic circumstances of the Nobel Laureates.
Please check it out and share. Details in 🧵 1/
1
87
Cool paper – and happy 1st birthday to the manuscript spent in peer review!
It turns out academic publishing has been pacing the frontier since 1731.
Biosecurity is one of the most urgent challenges for the AI era. Bringing SynthID to biology so AI-generated proteins can be watermarked is a critical step - and we’re open sourcing SynthID Bio tools so the research community can build on this work. Published in @Nature today, congrats to the team! nature.com/articles/s41586-0…
6
218
A fascinating experiment in LLM-guided reproducibility analysis.
What's incredible (to me, coming from life sciences) is that econ journals have had mandatory, relatively well-structured replication packages for ~20 years or more!
From my non-social-media spouse, a new paper, using LLMs to replicate ... everything.
nber.org/papers/w35782
Headlne: "Across 4,452 published replication packages for five economics journals, the workflow flags discrepancies in 3,460 articles or their appendices."
1
70
This is a great use case! Every researcher should have a literature-aware agent at their fingertips.
But PubMed search is just the first step of inquiry.
Expert researchers care about the details, often buried in the PDF or supplementary files.
I got to share with @OpenAIDevs how I’m using Codex to build tools for my work as a surgeon.
One example is a PubMed MCP server to review literature. Check out my feed for other examples.
Give Codex a try even if you're a "non-technical" doctor like me. It's a superpower!
1
3
244
Paperpile for Claude: organize and make sense of the mountain of PDFs you've accrued over time.
Quotes and statements are linked to the source, for easy reading and highlighting. Notes are saved back into the PDF for portability.
1
3
158
I'll start sharing some demos of our @paperpile agent plugin.
With Paperpile for Claude, you get two huge unlocks for research work: (1) fast, comprehensive search across the literature, and (2) quickly save to your personal library & download fulltext PDFs.
Paperpile's paper-finding MCP tools search across all scientific literature, with careful attention to detail in the features that matter:
- Constrain by date, author, or citation count
- Find inbound and outbound citations
- Up to 10 parallel queries in a single tool call
1
59