←
Ran Blekhman9,670 posts
Ran Blekhman avatar

Ran Blekhman

@blekhman

Professor of Medicine at @UChicago. Decoding the human microbiome. BlekhmanLab.org

◷ Joined November 2012▧ 1,316 Photos and videos
Ran Blekhman avatar
Ran Blekhman@blekhman·16h

When LLMs review a paper, they draw 50% of their weaknesses from the appendix. For human reviewers it's about 8%. The general trend: human reviewers focus on the big picture, overall presentation and support for the major claims. AI reviewers nitpick.

Post attachment
Ran Blekhman avatar
Ran Blekhman@blekhman·Nov 17

Is the Bitter Lesson coming to genomic prediction? An untuned tabular foundation model matches or beats specialized methods biorxiv.org/content/10.64898…

Ran Blekhman avatar
Ran Blekhman@blekhman·Nov 15

Is it time we abandon recommendation letters in academia? Letters have ~zero signal to noise ratio, and with AI they are easy to generate. A 10-minute call is much more informative. Compare today’s 3-page "top 1%" letters with Darwin’s 1851 letter for Huxley (2 sentences):

Post attachment
Ran Blekhman avatar
Ran Blekhman@blekhman·Nov 12

Why does every AI Bio company sounds exactly like Theranos? "a CRO but really fast" (with no tech innovation, no experience in biology) sounds a lot like "a blood test but from a really small amount"

Today we're launching Petrichor (YC F26), a wet lab for frontier AI labs. We run experiments in parallel and return the data from every step.

0:45
Ran Blekhman avatar
Ran Blekhman@blekhman·Nov 11

Some concerns are valid: data sovereignty, underrepresented languages, and asking “what AI is for” are real issues. Indigenous knowledge is useful. But blurring science and values, under the stamp of @ScienceMagazine, means both science and Indigenous knowledge lose credibility

Ran Blekhman avatar
Ran Blekhman@blekhman·Nov 11

The editorial is incorrect about fundemental science concepts. It calls “Western science” linear and purely reductionist, which is obviously wrong. It says AI is trained to use it as a “control”, also wrong. No citations at all for any claim in this piece.

Ran Blekhman avatar
Ran Blekhman@blekhman·Nov 11

Public trust in science is critical, especially now. This trust relies on the idea that “science” means real claims we can test and revise. So when a @ScienceMagazine editorial describes a “unifying life force” as part of science, it's a problem

"In Lakota science, there are four parts—matter, energy, relationality, and a unifying life force" That @ScienceMagazine, arguably the world's preeminent science journal, would publish this is mind boggling.

Post attachment
Greg Jordan avatar
⟳ Ran Blekhman retweeted
Greg Jordan@gjuggler·Nov 8

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.

Post attachment

arXiv submissions. Look at the last couple of months. Insane

Post attachment
Ran Blekhman avatar
Ran Blekhman@blekhman·Nov 6

The authors didn't do anything wrong here - they clearly disclosed their AI use But, as a reviewer/editor, I just cannot bring myself to read AI-generated papers. I don't reject the paper, I just step down as a reviewer. I simply can't justify spending any of my time reading it

Peak peer review. (We submitted to ICLR) #iclr2027

Post attachment
Ran Blekhman avatar
Ran Blekhman@blekhman·Nov 5

It's time, calling for a “Rebuild American Science Act”: - Double the NIH & NSF budgets - Limits on executive interference - End forward funding - Rebuild the research training pipeline - Full transparency of success rates - Permanent H-1B fee exemption for universities