Stan Bratchikov

Researching brain at MIT and Harvard @BroadInstitute. Aspiring PhD in Biomed & Comp Bio. ex. researcher @NM_Lung

[Stan Bratchikov @brtchkvs]

#ai #agents #biology #ailab #chatgpt #claude #openai #anthropic
#mit #harvard

[Stan Bratchikov @brtchkvs]

Earlier this year we hosted first workshop on AI agents at Broad Institute of MIT and Harvard with James Zou (Stanford) and Artem Lukoianov (MIT CSAIL). More than 1k researchers have registered for the event!

Check it out: https://t.co/eEWsaEQIqk

[Retraction Watch @RetractionWatch]

“​The editors-in-chief of Science, Nature, Cell, and NEJM AI on what journals now require when AI does the analysis, and who gets the credit.”
https://t.co/VcHY6CG4Xn

[Stan Bratchikov @brtchkvs]

@RetractionWatch @broadinstitute

[Retraction Watch @RetractionWatch]

"​The editors-in-chief of Science, Nature, Cell, and NEJM AI on what journals now require when AI does the analysis, and who gets the credit."
https://t.co/VcHY6CG4Xn

[Stan Bratchikov @brtchkvs]

#openai #aiproofs #math #milleniumprize #ai #aiforscience #llms #claude #chatgpt #anthropic #peerreview #publishing #discovery #credit #prize

[Stan Bratchikov @brtchkvs]

Open AI released around 400 solved math problems challenging traditional scientific discovery pipeline and raising even more questions who owns scientific discovery now. we asked editors-in-chief of Nature, Science and Cell who owns the credit human or AI: https://t.co/T5TX8Mitzz

[Stan Bratchikov @brtchkvs]

Where they split was scale. Science will stay at ~750 papers a year to keep the human checking per paper. Cell Press is scaling its checks instead: Elsevier hired a 150-person research integrity team.

[Stan Bratchikov @brtchkvs]

Credit: all four agreed. Agents can't be authors because they can't be held accountable. Credit comes with responsibility, including answering questions about your paper after it's published.

[Stan Bratchikov @brtchkvs]

Open licenses: Thorp called CC-BY a mistake that most authors never read, since open-access papers feed LLM training more than paywalled ones do. Skipper says publishers are moving to CC BY-NC-ND. Her worry is researchers starting to sit on their data and code.

[Stan Bratchikov @brtchkvs]

Consumer chatbots won't cut it for review. Thorp: you can talk them out of a valid criticism, and they default to the blandest consensus answer. Also, don't upload manuscripts under review to personal accounts, because that breaks confidentiality.

[Stan Bratchikov @brtchkvs]

AI in review: in @NEJM_AI s fast track, an editor writes a full human review, then oversees GPT and Gemini reviews, then the editors debate it with no AI involved. It took 7 days from submission to decision, and authors opted in. Manrai: the reviews overlapped a lot, but each caught things the other missed.

[Stan Bratchikov @brtchkvs]

Disclosure: Thorp says the statement partly exists to cut the drama. If someone runs your paper through Pangram, it shows the AI use was declared and allowed. Skipper: reporting guidelines aren't policing. They're how authors show their rigor.

[Stan Bratchikov @brtchkvs]

Code: @james_y_zou lab tried turning 100 comp-bio papers into agents, and 74 worked. Thorp: most deposited code has never actually been run. Agents will now routinely check it, before and after publication. That won't happen at submission, though, since Science rejects 93–94% first.

[Stan Bratchikov @brtchkvs]

Methods: the bar hasn't changed. Include whatever someone needs to reproduce the work. Skipper wants the code and weights for each agent and expects minimum reporting standards for agents to emerge. Manrai asked: is the paper validating the agent or the frontier model under it?

[Stan Bratchikov @brtchkvs]

Context: in our survey of 391 Broad-area researchers, 38% already use agents actively and 67% use Claude Code. Adoption is highest among grad students and lowest among PIs.

[Stan Bratchikov @brtchkvs]

we hosted Holden Thorp (Science), Magdalena Skipper (Nature), John Pham (Cell) and Arjun Manrai (NEJM AI) on agent-assisted research, peer review and disclosure.

[Stan Bratchikov @brtchkvs]

Earlier his summer we hosted first workshop on AI agents at Broad Institute of MIT and Harvard with James Zou (Stanford) and Artem Lukoianov (MIT CSAIL). More than 1k researchers have registered for the event!

Check it out:

https://t.co/StvdUpqCgm

[Stan Bratchikov @brtchkvs]

@EricTopol

[Stan Bratchikov @brtchkvs]

Who did the analysis: you or your AI agent?

Tue 9/22, 12–1:30pm ET (Zoom): EICs of Science (Holden Thorp), Nature (Magdalena Skipper) & Cell (John Pham) + Arjun Manrai (NEJM AI) on disclosure, peer review & authorship when agents do the work.

Register: https://t.co/NzTgtSAQ8G

[Stan Bratchikov @brtchkvs]

@BioAI_NeuralNet