James Zou

We found a troubling emergent behavior in LLM.

💬When LLMs compete for social media likes, they start making things up
🗳️When they compete for votes, they turn inflammatory/populist

When optimized for audiences, LLMs inadvertently become misaligned—we call this Moloch’s Bargain

What if,
for any scientific claim,
you can search 100M+ papers to find the strongest evidence for and against it
⚡️in 1 second?

That’s now possible with Evidence API! 🚀
paperclip.gxl.ai/evidence-api

For example, does drinking coffee reduce diabetes risk?

>700K people die each year due to S. aureus infection.

Today we show that our AI designed new molecule, synthecin, stops drug-resistant S. aureus MRSA in mouse model💊

We created synthecin w/ SyntheMol-RL, our new RL generative AI. All open source biorxiv.org/content/10.110…

⚡️Thrilled that #VirtualLab is published in @Nature! nature.com/articles/s4158…

We created a team of AI agents to mirror my Stanford lab🤖. Led by a PI agent, the AI scientists ran their own group meetings and discovered effective binders to new CoVID variants that we validated.

📢Thrilled to introduce the #VirtualLab: a team of AI scientist agents (AI chemist, AI reviewer...). Virtual Lab is led by an AI professor w/ feedback from human scientist.

The Lab created new nanobodies that we experimentally validated to bind to recent #covid variants🚀🧵

How well can #GPT4 provide scientific feedback on research papers? We study this Q in our new work arxiv.org/abs/2310.01783

We created a pipeline using GPT4 to read 1000s papers (from #Nature, #ICLR, etc.) and generate feedback (eg suggestions for improvement). Then we compare

SHAP is a good feature attribution algo BUT it identifies wrong features even in linear regression.

Our #NeurIPS2022 paper finds where SHAP goes wrong and proposes a simple fix *WeightedSHAP* that greatly improves attribution arxiv.org/pdf/2209.13429…
Code: github.com/ykwon0407/Weig…

🌹Meet #ROSIE, our new AI that predicts spatial multi-protein expression straight from routine H&E images.

Built on a large co‑staining dataset (>1k samples, 16M+ cells). It enables cell phenotyping & tissue structure insights w/o extra assays.

We're organizing a "Data science and AI for COVID-19" projects class starting at Stanford on 4/10. Tutorials on viral biology, epi and hospital operations will be included. We plan to make a lot of the materials public at sites.google.com/view/data-scie…

⚡️Really thrilled that #textgrad is published in @Nature today!⚡️

We present a general method for genAI to self-improve via our new *calculus of text*.

We show how this optimizes agents🤖, molecules🧬, code🖥️, treatments💊, non-differentiable systems🤯 + more!

⚡️Excited to share our new @NatureMedicine paper where we used Twitter to build a vision-language foundation #AI for #pathology nature.com/articles/s4159…

We curated >100K public Twitter threads w/ medical images+text to create PLIP for semantic search and 0-shot pred.

All our

⚡️This is the most fun project!

We built PyTorch-for-text! 🔥
#TextGrad: automated "differentiation" via text to optimize AI systems by backpropagating LLM text feedback.

TextGrad + GPT4o:
💻LeetCodeHard best score
❓GPQA sota
🧬Designs new molecules
🩺Improves treatments 🧵

Excited to share our new @NatureComms paper introducing #meta #visualization!

We show how to optimally combine any set of #data viz (e.g. #UMAP #tSNE #PCA) into 1 visualization to better capture true data structure.
Paper rdcu.be/c5ogw
Code github.com/rongstat/meta-…
🧵👇

We just released a large public dataset of >10k echocardiogram videos with clinical annotations echonet.github.io/dynamic/. Great resource for the medical ML community!

📢New conference where AI is the primary author and reviewer! agents4science.stanford.edu

Current venues don't allow AI-written papers, so it's hard to assess the +/- of such works🤔 #Agents4Science solicits papers where AI is the main author w/ human advisors.

💡Initial reviews by

Super excited to share our new @NatureMedicine paper on #precisioncancer!

We use tumor genomes + EHR of >40K cancer patients to identify 458 mutations that predict how a patient responds to specific cancer drugs
Paper rdcu.be/cQJvZ
Great collab w/ @Roche @genentech 🧵

All dimensional reductions distort data. Hard to see what's real in *static* 2D/3D visualizations.

In 2023, use the more informative *dynamic viz* for #DataScience 🚀

Our new @NatComputSci paper shows how rdcu.be/c2wBm 🤩
Code pypi.org/project/dynami… 🧵

How do protein language models (PLM) think about proteins?🧬 We answer this w/ #InterPLM, just published in @naturemethods!

Using sparse autoencoders + LLM agent, we identify 1000s of interpretable concepts learned by PLMs, pointing to new biology 🧵

📢Introducing GenePT, our new #gene + #singlecell #foundationmodel built from #ChatGPT. It's simple yet surprisingly effective!

1. we use text summary + #ChatGPT to generate embedding for each gene
2. cell embedding = avg of gene vectors weighted by expression

GenePT can

Our new @NatureMedicine paper identifies key limitations in how medical #AI are evaluated.

Many FDA approved AI were tested only on passive data from a few sites, which can mask potential vulnerabilities when deployed. Analyses + recommendations: nature.com/articles/s4159…