A moment of introspection on how much the world has changed during the time between the initial submission and this (hopefully final) revision:
"This paper seems so quaint now"
"Let's think of it as having contributed an important bit of high-quality training data to Opus 4.7"
Claude Code demo, planning another with more details.
Ran an AI coding workshop with the lab. There was a palpable sense of sadness realizing that skills some of us have spent our lives developing (myself included) are a lot less important now. I see the future 100%, but I do think it's important to acknowledge this sense of loss.
ft. Sam Reffsin, Margaret Dunagin, Yael Heyman, @arjunraj.bsky.social (Bioeng), Jesse Miller, Kasirajan Ayyanathan, Sara Cherry (@pennpathlabmed.bsky.social), @naveen-jain.bsky.social (@cambupenn.bsky.social) & David Schultz (BioBio) @penngenetics.bsky.social www.sciencedirect.com/science/arti...
The TIG1-high state exists in human lung club and ciliated cells, enriched in IPF patients and tobacco users. Different viruses target different states. Influenza A prefers KRT8-high cells. Arjun Raj @cp-cell.bsky.social
doi.org/10.1016/j.ce... 🧪
3/3
Reffsin, Miller, Cherry & Raj et al. discover that intrinsic cell states determine which cells become infected with SARS-CoV-2. Using single-cell clone tracing, they identified a TIG1-high state marking highly susceptible cells. What makes these cells vulnerable? 👇🏻
1/3
Data and experiments move faster, but theory is what gives discovery direction, via @arjunraj.bsky.social
arjunrajlab.substack.com/p/chance-fav...
We spend a lot of energy choosing the right research project. But an equally important skill is knowing when to walk away.
Quitting isn’t failure: It’s how you make room for better science, via @arjunraj.bsky.social
Anyone have a thought about this weird mol on my hand?
Where do you get your groceries? Where are your retirement funds invested? Etc.
I gotta say… it’s just less fun here. I want it to work, but…
Congratulations!!
I am happy to share that I’ve been awarded tenure here at Columbia University. Alhamdulillah. I thank my wife, parents, mentors, colleagues, friends, and family. I thank my lab members, present and past, for entrusting their careers in my hands and making this a thrilling journey.
I can say that the discussion on X are definitely more interesting. Here, it’s like AI isn’t even happening.
Both, and things in between. It’s highly variable as to how much it can do autonomously in both cases. Probably depends on how much training data there is.
Both. General purpose non-coding agent ;).
Huh. I think the scope it can handle probably depends on the nature of the challenge (how much it is in distribution) and on the planning and amount of testing you have in place. I have found it able to do some pretty remarkable features, one-shotting a lot of non-trivial boilerplate.
Haven't run into trouble yet! But I've been sadly too busy to code as much as I have wanted.
What I have not explored but should is non-coding uses.
Much to say, but to summarize: 🤯
I think that’s the first sign of the apocalypse, no?
AI is awesome, and is most definitely a transformative technology.
To me, the mean loud position here ranges between ignoring the ai and wishing to fully ban it. The latter won't happen (I don't think even regulation is technically feasible), the former is massively unproductive. If scientists refuse to participate in the discussion,it will just be driven by others
So… Bluesky for science and X for Claude Code?
Blog post: Chance favors the (theoretically) prepared mind
Data is big, machines are learning, so what good is theory anyway? Isn't most discovery driven by serendipity anyway, with theory mostly a "post-mortem"? I argue that this view discounts the value of theory.
open.substack.com/pub/arjunraj...
Beautiful - recommended! Here, @sasolla.bsky.social recaps her decades-long journey from physics to neural networks (working with LeCun & Hopfield) to motor cortex, & and from industry (including Bell Labs) to academia, all driven by curiosity and awe (which flows from her voice). Inspiring!
i've re-implemented PoPE from scratch as part of a model overhaul (w/ Claude Code). i don't have clean ablations yet but it's working. took about two hours to re-write my entire model
i think the PoPE idea is really cool and a great fit for scientific data
Congratulations on tenure!
2025 was the year we stopped debating whether paying peer reviewers might work and started showing that it does.
