Steven Ge

Excited to launch DataMap – a portable, browser-based tool for generating heatmaps and PCA & tSNE plots. No installs. No servers. Your data never leaves your device. ✔️ Interactive heatmaps, PCA, and t-SNE ✔️ Transform & normalize data on the fly ✔️ Upload CSV, TSV, Excel + annotations ✔️ Generate R code for reproducibility Ideal for visualizing high-dimensional data matrices, such as RNA-Seq data. It's a Shiny app deployed via shinylive and WebAssembly. Also available as an R package. Try it now: github.com/gexijin/datamap Source code: github.com/gexijin/datamap Preprint: arxiv.org/abs/2504.08875

• AI Agents: drives itself. Plans, acts, verifies, and repeats, but makes lots of decisions that can be wrong. Each layer fixes what the previous one fakes.

“Can AI do X?” depends on which layer you mean: • LLM: text in, text out. Fakes reasoning. Bad at math & logic. • LLM with reasoning: adds deliberation. Plans better, but can still invent facts. • LLM with tools: adds hands. Can search, calculate, but doesn't persist. 1/2

Don't surrender and outsource thinking. Cognitive debt compounds over time.

Confession: I trusted Claude Code too much and moved too fast in my bioinformatics research. AI's execution outpaced my ability to understand. I was prompting & accepting without reading the full responses.

I came back to the United States, hearing about a very sad news: Craig Venter passed away today. He is a pioneer, successfully sequenced the first Human genome, and tried to create Synthetic Cells. We lost a giant in Science. RIP. @JCVenterInst

My wife collapsed in our hotel room in New York today “Call an ambulance, I can’t breathe” she was screaming My heart dropped If she ends up in an American hospital we are financially ruined I went on the Lufthansa app to book a flight back to Frankfurt, but unfortunately pilots are on strike today “Please I’m begging you” she was lying on the floor I sighed and called 911 She is now in surgery as apparently her appendix “almost burst” I am extremely scared This is going to cost us at least $100,000 She could have received much better care, for free, in Germany I will never visit this barbaric country ever again

Our paper in @Nature today 🥳 We tracked 6,438 mice from puberty to death and mapped the genetics of *when* you die, not just whether a gene associates with lifespan. nature.com/articles/s4158… 59 loci. Two decades of data. Thread 👇 #Longevity #Aging #Genetics #Healthspan

SubcellulaRVis: a web-based tool to simplify and visualise subcellular compartment enrichment shiny.its.manchester.ac.uk/subcellularvis/ Simply copy/paste your gene/protein list @ChiaraFrancavi2 #NucAcidsRes 2022 academic.oup.com/nar/article/50…

@candybilzzard Claude Code + Skill. Its just way simpler

@StevenXGe 所以现在来看到底是像langgraph那样开发workflow，还是应该claude code + skill 这些生态。哪一种会是未来，精度bar是怎么样的

LangChain has led the way. But now the Claude Agent SDK adds a powerful new option. It gives developers the same capabilities used in the Claude Code app — including support for Claude Skills, tools, and MCP servers. You get file access, tool use, and strong context management out of the box. For example, you can create a Finance Agent that runs analyses and generates reports using your own customized Skills and MCP servers. youtu.be/i6N8oQQ0tUE?si…

@suragnair For me, Claude Code consistently gets the job done with my lazy prompts. It just gets me!

@StevenXGe We saw this on our benchmark as well. Codex is much faster than Claude Code. But performance was similar for both! x.com/suragnair/stat…

Codex CLI is fast. Claude Code is thoughtful. Codex wins on speed and cost. Claude wins on judgment. What stood out to me is that Claude Code better understands the project as a whole: the files, the context, and what I actually want done. So instead of choosing one, combine them: Claude Code for direction, Codex for execution. That feels like the most practical AI coding workflow right now.

@StevenXGe I wrote an article for ASBMB today reflecting a similar point. I appreciate you highlighting that same sentiment and that using AI within your writing process doesn't mean you're using AI to circumvent writing. asbmb.org/asbmb-today/op…

Thrilled to share our new @NatureGenet paper! We mapped human skin at single-cell spatial resolution and found that anatomy is encoded in cell states and neighborhoods, led by PhD student @paularstrpo. Link: nature.com/articles/s4158… 1/9

Vibe Writing: how I co-write with Claude Code, one sentence at a time. It’s not “AI, write my paper.” It’s me thinking out loud while AI handles the mechanics. I direct. AI drafts. I said, “move this to methods” → Claude saw it was already covered there and cut the redundancy. I flag the problem. AI diagnoses it. “The flow isn’t working” → Claude identified the issue, gave options, and I chose one. I shape the argument. AI executes. “Discuss comorbidity first” → “Then post-mortem” → “Add references” → “Mention GTEx wasn’t designed for disease” → “Condense” AI as a sounding board. “Should I discuss covariates here?” → Claude suggested adding a sentence. I edit. AI checks. "Go through my edits" → Claude caught a typo and a tense mismatch. AI reviews for readability. “Review this paragraph for flow and readability.” The whole session took minutes, instead of hours. Vibe writing isn’t letting AI write your paper. It’s writing with a tireless partner. This post was drafted by Claude, reflecting on a live session. Polished by ChatGPT.

Fascinating story about how ordinary people exposed the government's dark secrets. cnn.com/2026/03/20/us/…

OpenClaw is impressive, but you can build a more transparent AI agent with just Claude Code and Bash. Roman's video shows how to create custom agents for real-world workflows, like getting an email summary at 9 a.m. or managing email via Telegram. What clicked for me: • 'claude -p' lets you run Claude Code as a headless command • It can be triggered by schedulers like cron or by other events • 'claude -p -r' helps preserve context across sessions • You can even replace Claude Code’s system prompt • Commands, skills, and subagents are, at their core, reusable prompts That’s the direction I’m most excited about: agentic workflows that are simple, inspectable, and truly yours. youtu.be/ODKMmKCgrvw?si… #AI #ClaudeCode #CodingAgents #Automation #BuildInPublic

📢 Preprint: we present a whole-mouse-brain in vivo Perturb-seq atlas, 7.7 million cells, 1947 disease-associated perturbations, moving toward direct readout of how human genetics rewires cell states & circuits in vivo. Grateful for the Team! @NVIDIAHealth biorxiv.org/content/10.648…

Dennis Ritchie created C in the early 1970s without Google, Stack Overflow, GitHub, or any AI ( Claude, Cursor, Codex) assistant. - No VC funding. - No viral launch. - No TED talk. - Just two engineers at Bell Labs. A terminal. And a problem to solve. He built a language that fit in kilobytes. 50 years later, it runs everything. Linux kernel. Windows. macOS. Every iPhone. Every Android. NASA’s deep space probes. The International Space Station. > Python borrowed from it. > Java borrowed from it. > JavaScript borrowed from it. If you have ever written a single line of code in any language, you did it in Dennis Ritchie’s shadow. He died in 2011. The same week as Steve Jobs. Jobs got the front pages. Ritchie got silence. This Legend deserves to be celebrated.

--dangerously-skip-permissions in Claude Code worked fine for months — until today. Claude wiped all the scripts in one of my subfolders, and I hadn’t committed several hours of work yet. Lessons: Commit often. Use YOLO mode with caution.

Good news: Claude Code can run complex data analyses on its own for hours. Bad news: it makes many decisions quietly along the way. Hard-coded parameters, filtering choices, and assumptions. Most are fine, but a single bad parameter can derail the entire workflow. We need to supervise: 1. Plan thoroughly and iterate. Have it explain the core logic, then refine the plan with feedback. 2. Inspect the code. Have it list all hard-coded values and key decisions, then review the implementation carefully. 3. Verify what the data looks like at each stage. Even the best LLMs are smart MOST of the time. For scientific research and mission-critical decisions, that's not enough.