Luke Sjulson

Graph of award probability of R35 and R01 from NIH factbook as a function of review rank percentile. As is apparent, 2025 is a significant departure, with lower award probabilities at all scores <40 and significant departures from norm, where even being in the top 10% is no longer a nearly certain indicator of success. Data source: https://report.nih.gov/nihdatabook/report/302

The data is in: the NIH goalposts have shifted.

What were once almost certain fundable scores have become coin flips and what used to be likely grants have become aspirational, leading to fewer awards.

Another manifestation of how HHS policies have led to fewer awards and less science.

I’m coding

random question for people with more experience with coding agents than I have: does it make sense to give an agent its own non-admin user account on your system, then give it full privileges? Isn't that easier than dealing with setting all the permissions?

De Grey was on my list alongside Elizabeth Holmes, Elon Musk, and John Ioannidis as people whom I immediately thought were full of shit, and I couldn’t understand why everyone else didn’t agree.

Remember his plan to stop aging by moving the last bit of the mitochondrial genome into the nucleus? 🤡

But Joao, aren’t you afraid of being scooped?

No.

It’s an ultra-rare childhood epilepsy. If you scoop me and the kids get help from it, I will help you scoop me and then move on.

I’m so proud of my 8yo who went to a sleepover last night and adjusted the poses of some creepy Japanese dolls on display, then went to her friend’s mom saying she was terrified because she saw the dolls moving on their own. The mom called my wife last night freaking out

Onboard if, and only if, then reviewers finally have to stop asking for qRT-PCR to validate said NGS.

I will fly to anywhere in the United States on my own dime to talk to people about basic science! I encourage all of you, particularly senior scientists to make this commitment. If you are looking for material ask @karalmarshall.bsky.social She has put together an amazing basic talk!

Brooklyn 8:37 PM

In a spatial navigation task, the hippocampus will create place cells. In a trace conditioning task, it will create trace cells. In a timing task it will create time cells

I don't know; I haven't tried that specifically

Trust me on this one; it will prevent 90% of your figure formatting headaches

You generally want to wrap text around the figure, and you almost always want to select the option that anchors it to a specific page location rather than allowing it to move with the text

Advice for a young investigator:

For figures in Google Docs, insert a drawing, then put both your image and the text box containing the caption inside the drawing

For MS Word, insert a text box and put your image and caption inside

For Pages, insert image and text box separately then link them

Awesome work, congrats!

Duration between rewards controls the rate of behavioral and dopaminergic learning - Nature Neuroscience Cue–reward learning rate scales proportionally with the time between rewards. Consequently, learning over a fixed duration is independent of the number of trials. This challenges trial-based dopamine ...

Very excited to post our paper led by @daburke.bsky.social www.nature.com/articles/s41... where we uncover a simple mathematical rule underlying how brains learn that a cue predicts a reward. 1/26

You have a point, but my post was more about LLM coding agents in general, not Claude Code specifically

We need to raise the bar on research code right now.

1) documentation and tests are dead simple now.
2) creating benchmarks integrating across multiple implementations
3) have agents double check your work / fix broken tests
4) fix outstanding bugs in major scientific packages

A small polymerase ribozyme that can synthesize itself and its complementary strand The emergence of a chemical system capable of self-replication and evolution is a critical event in the origin of life. RNA polymerase ribozymes can replicate RNA, but their large size and structural ...

Ooooh. Cool new paper on origins of life. A simple 45-nucleotide RNA molecule that can perfectly copy itself.

www.science.org/doi/10.1126/...

New paper alert! 🚨

We found that the brain's compass is remarkably stable at two scales

1️⃣ the system maintains its internal organization for weeks
2️⃣ It "remembers" its orientation for weeks, even after a single visit

This may be key to how the brain aligns its other maps.

Paper: rdcu.be/e3waP

exactly, I think open science will benefit a lot from this.

Re: how to pay for Claude, I haven't looked into it yet, but my first move will be to get permission from an NIH PO to rebudget for it, then show that email to our internal people as justification

quickly. The $20/month plan with Claude isn't enough for any serious work, but the $20/month plan with Codex might be. I am trying to figure out what to do because Einstein will provide free Codex access, but I don't think they'll do that for Claude, so I'd have to get creative with reimbursements

I have used both Claude Code and Codex, though I've used Claude more. The main thing is that Codex is a lot slower than Claude, to the point where it's unpleasant to use. It also isn't as "proactive," if that makes sense. The flipside is that you burn through your usage with Claude a lot more

Best Practices for Claude Code - Claude Code Docs Tips and patterns for getting the most out of Claude Code, from configuring your environment to scaling across parallel sessions.

training. I found this document to be a really useful starting point:

code.claude.com/docs/en/best...

My wife's company has started training their employees in these methods, and I am going to get my lab trainingas well. If anyone knows of any good resources, please post them! 8/8

serious errors. They can make things so much better or so much worse, which is why I think it's so important that we adopt them the right way. The wrong way is what most of us are doing right now, cutting and pasting code from a ChatGPT window. The right way will require discipline and proper 7/

realistically have time to go and manually verify that every function people in my lab use is properly unit tested, but I am going to start reviewing everyone's code with claude and codex.

Of course, it's also true that coding agents enable people to generate even worse code that's full of 6/

not enough on higher-level planning and organization, which is the source of many problems. With LLM agents, you have to tell them exactly what you want to get good results, which forces you to spend more time up front organizing your thoughts.

4) code review will be much easier. I don't 5/

structure demands it. Consequently, agents are primed to use these practices.

2) agents massively reduce the overhead associated with using good practices, which will make them much easier to enforce.

3) most importantly, the average lab programmer spends too much time on implementation and 4/

something completely incomprehensible and useless?

This hurts us all, but no one has found a solution. I think LLM coding agents are going to help immensely, for 4 reasons: 1) they are trained on code written by professional software engineers, who adhere to good practices bc their incentive 3/

good organization, compatible file formats, modularity, proper unit testing, and good documentation. Example: my lab’s behavior box GitHub repo currently has 19 branches. Someone should merge them, but who’s going to do it?

How many times have you requested code from a publication and gotten 2/