Joe Janizek

Thrilled to share that @ethanweinberger.bsky.social is becoming Dr. Weinberger in CSE2, where he is presenting his work, including the popular contrastiveVI for single-cell data (Weinberger et al. Nature Methods)! I feel so fortunate to work with such amazing Ph.D. students at @uwcse.bsky.social! 🎉🎓

Medical education 🌶️🔥take: the threat of cognitive deskilling from genAI technologies is the #1 things medical educators need to be talking about right now.

AI deployments in health are often understudied because they require time and careful analysis.⌛️🤔

We share thoughts in @ai.nejm.org about a recent AI tool for emergency dept triage that: 1) improves wait times and fairness (!), and 2) helps nurses unevenly based on triage ability

Really cool paper!

AKA the “Concrete” distribution, which I think is a much better name lol

Basically a continuous relaxation of discrete random variables, lets you do stuff like differentiating through sampling (e.g. argmax) operations

The Gumbel-softmax distribution

I've been developing a semantic search tool that covers not just bioRxiv and medRxiv, but the entire PubMed database. This means you can search across a massive collection of biomedical research using keywords, questions, hypotheses, or even full abstracts. Try it out: mssearch.xyz

Machine Learning 101? Imagining a new syllabus for a first course on machine learning.

What should we teach our undergrads about machine learning? I wrote up some ideas for restructuring Machine Learning 101.

Reading that answer, I’m realizing this may actually qualify as “a big jeopardy person”

Not so much recently. I watched it a lot as a kid, and then my wife and I were watching it regularly for almost a year right at the start of Covid — was pretty fun, you can also turn up the difficulty by trying to like, guess what the clues and answers are going to be just from the categories

yes, but only because of a final jeopardy back in 2020 that was actually trying to clue a different poem

the different midjourney variations are so interesting. like, the whole row of guys w/ really crazy eyes, not sure where it got that from the prompt, but has a real Ilya Repin -- Ivan the Terrible / Goya -- Saturn vibe

Think how much performance we might be leaving on the table by not training classifiers on increasingly invasive biometrics. Pictured: medium-term radiologist-AI centaur-configuration possibility

If this was an AI paper, you’d brand it as an interpretability technique that discovers a latent “node detection” circuit in the neural network

pubs.rsna.org/doi/10.1148/...

There is a lot of buzz about our new paper in Nature Medicine on the effects of LLMs (GPT-4) on physician management reasoning! I had TONS of fun working on this -- but what it MEANS requires some unpacking.

A 🧵⬇️
bsky.app/profile/ucsf...

Brainwide silencing of prion protein by AAV-mediated delivery of an engineered compact epigenetic editor Prion disease is caused by misfolding of the prion protein (PrP) into pathogenic self-propagating conformations, leading to rapid-onset dementia and death. However, elimination of endogenous PrP halts...

Insanely good news if this holds up.

Brainwide silencing of prion protein by AAV-mediated delivery of an engineered compact epigenetic editor

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

What's Happening Inside the NIH and NSF

A long post about what’s happening to the science funding agencies in the US and why. As mentioned, this one just kept getting longer even as I kept stripping curse words from it.

www.science.org/content/blog...

The in-context inductive biases of vision-language models differ across modalities Inductive biases are what allow learners to make guesses in the absence of conclusive evidence. These biases have often been studied in cognitive science using concepts or categories -- e.g. by testin...

New (short) paper showing how the in-context inductive biases of vision-language models — the way that they generalize concepts learned in context — depend on the modality and phrasing! arxiv.org/abs/2502.01530 Quick summary: 1/5

i use claude as a rubber duck a lot, and i always make sure to thank it. not because i think that it can appreciate my thanks but i refuse to surround myself with objects which i experience as human but refuse to treat as human. we should not be learning to dehumanize the experience of intelligence.

📈How far are leading models from mastering realistic medical tasks? MedXpertQA, our new text & multimodal medical benchmark, reveals gaps in model abilities

📌Percentage scores on our Text subset:
o3-mini: 37.30
R1: 37.76 - frontrunner among open-source models
o1: 44.67 - still room for improvement!

Anyway, not exactly a guide on how to prompt, but I do think interacting with multiple chat models is a great way to get an understanding of “what’s common” to different models/LLMs more broadly

I could imagine that playing around querying OpenEvidence and then submitting the same queries to, say, ChatGPT with and without search enabled could be an interesting way to understand what sort of questions models tend to be reliable for, when hallucinations are more likely, etc

OpenEvidence The leading medical information platform.

For more clinically-oriented things, I’ve really enjoyed the OpenEvidence platform. Currently free, and grounds all of its generations in real guidelines/trials/etc. www.openevidence.com

Business idea: Anki decks but for Involuntary Memory

better screenshots

“An AI escaped from the lab!”

“Which one?”

“Uh, something named helpful-only”

“Dear god…”

If you can use Gemini to spin up reasonable reasoning traces at a fraction of the cost, even if real expert traces are eventually better, this lets you generate proof of concept that this is worthwhile to justify the investment

59K questions * 30 min * 50$/hour is ballpark $1.5 million — you could obviously imagine generating a smaller expert set of reasoning traces, but even the 1K they eventually distilled to would still be ballpark $50K. That’s a huge investment either way

Concretely, you can do a back of the envelope calculation on how much it would have cost to generate expert reasoning traces for their dataset — they initially did 59K questions, which for this difficulty level would take experts greater than 30 min per question (see GPQA paper)