For bioinformatics coding, use Codex with gpt-5.6-sol. Beats Claude Code with Opus 4.8 by a long shot. Fable 5 won't even look at anything remotely related to biology.
Steven Ge
@stevenge.bsky.social
Professor, founder of Orditus. AI, genomics, bioinformatics
My 5 strategies for saving tokens—especially with Claude Fable 5: - Use Claude.ai for bounded tasks. Use Claude Code only when it needs to inspect files, edit a repo, run commands, or test its work. - Control the context. Start a new chat when switching tasks. Edit a prompt instead of piling on.
Fable 5 extended again to 7/19. Thank you, Anthropic, for motivating me! I needed deadlines in the summer! 😀
“Explain briefly in plain language, using bullet points.” I’ve been saying this to LLMs a lot recently. These models have become incredibly powerful. But their answers are often long, technical, and overly detailed. Sometimes I simply don’t have the time, patience, or knowledge to read.
Don't worry about not keeping up with AI. Nobody is. Identify and solve a problem! Even last year's AI is underutilized! For scientists, discover something new. Or create a tool that helps others.
After 18 months of R&D with a team of specialist academics, I'm VERY excited to say that @carbonbrief.org has just launched 'Project Cosmos'... ...the world's largest database of climate research Our hope is this becomes a major new resource interactive.carbonbrief.org/cosmos/index...
Project Cosmos – the world’s largest database of climate change research
interactive.carbonbrief.org
On June 5th Israel attacked the home of Lebanese sea turtle conservationist Mona Khalil. Mona was hospitalized with burns and injuries to over 70% of her body. Yesterday, June 19th, Mona died from her injuries. This wasn’t an accident.
Today's "From the Files" 06-21-26: Mountain Shack. Silverton, CO. 2025 #davidgardnerphotography #liightight #landscape #vertical #grandturkpeak #fall #mountain #structure #shack #fallcolor #aspen
1/ Many genomics mistakes are hard to catch in the code. They’re in the data you never looked at. Genome browsers can change that. Here’s how.
Anyone can create data dashboards with AI. Following this tutorial, you can use ChatGPT, React, and VS Code to turn a data file into a simple, interactive dashboard. You do not need to know how to code or understand HTML or JavaScript to get started.
vibe
gexijin.github.io
More biologists are coding their own analyses. Next: version control. Following this 30-min tutorial, you can learn how to use ChatGPT and GitHub Desktop to build a simple data dashboard. GitHub helps you track changes, collaborate, and share your work. Tutorial: gexijin.github.io/vibe/Git_Bra...
vibe
gexijin.github.io
I've been following & researching "precision" medicine / genomics for decades... So much hype. So much underwhelming or narrow applications. This seems genuinely exciting! 👇 Groundbreaking genomic test could spare millions of breast cancer patients chemotherapy www.theguardian.com/society/2026...
Groundbreaking genomic test could spare millions of breast cancer patients chemotherapy
Trial suggests patients with a low test score could be treated with hormone therapy alone with near-identical outcomes
theguardian.com
Ratio Percentile Deviation (RPD): A nonparametric, compositionally robust method for measuring the divergence of a microbial sample from a reference dataset https://www.biorxiv.org/content/10.64898/2026.05.27.728224v1
I love and respect this scientist. Carl Sagan answers a question about creation vs evolution.
Claude Code shipped /ultrareview and almost nobody is talking about what's actually new about it. It's not "AI reviews your code." We had that.
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. The result? Cognitive debt that compounds over time.
OmicClaw: executable and reproducible natural-language multi-omics analysis over the unified OmicVerse ecosystem www.biorxiv.org/content/10....
2 years ago I went to bed feeling great, woke up next morning with a painful ear and hearing loss (which became permanent). Because of a shingles infection in my auditory nerve. Get the vaccine if you're eligible www.wired.com/story/shingl...
You Should Be More Freaked Out by Shingles
The viral infection leaves millions with chronic pain, increased stroke risk, and lifelong nerve damage—yet vaccination rates remain dangerously low.
wired.com
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.
I’m giving a seminar on 20th April about refactoring crsu.shinyapps.io/MetaInsight/ to use {shinyscholar}, producing crsu.shinyapps.io/MetaInsight_... Details and sign up: www.eventbrite.co.uk/e/crsu-webin... A shorter version is also coming to R/Medicine in May #rstats #metaanalysis #openscience
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.
OmicClaw: executable and reproducible natural-language multi-omics analysis over the unified OmicVerse ecosystem. www.biorxiv.org/content/10.6... 🧬💻🧪 github.com/Starlitnight...
The new GPT-5.4 Mini is slightly less intelligent but faster and costs only 30% of GPT-5.4 and 15% of Opus 4.6. It would be great for many applications. This also makes running local models hard to justify solely on economic grounds. openai.com/index/introd...
Introducing GPT-5.4 mini and nano
GPT-5.4 mini and nano are smaller, faster versions of GPT-5.4 optimized for coding, tool use, multimodal reasoning, and high-volume API and sub-agent workloads.
openai.com
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:
I Replaced OpenClaw with Claude Code (It's Not Even Close)
YouTube video by Agentic Lab
youtu.be
--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.
I'm wondering if the reason many people are so keen on this "science through repeated prompts" idea is because that's already how they operate with their students (or, outside of academia, their team). Try this vague idea. Try again. Come back next week after trying more. Make this into a paper.
Today we're releasing AI for RStudio. It's really, really good—I'd encourage you to point it at the messiest data sources you have and see what it can do. www.simonpcouch.com/blog/2026-03...
