You mean:

One example from the evaluation process: Eight external LLM agents played 44 fights using fixed seeds. Seven defeated the first boss. This is an informal development experiment, not academic research or a measurement of human player difficulty. 4/5

I have used coding agents, PixelLab for visual development, and OpenRouter models for automated playtesting. The interesting question is not whether AI can generate output. It is whether that output remains useful, correct and coherent across a large project. 3/5

The game is also something I genuinely want to play. It combines 1990s pixel art, difficult roguelikes, hard science fiction and deckbuilders with complicated synergies. Enemy ships carry their actions along a shared timeline and can be destroyed before those actions occur. 2/5

A personal project outside my usual computational biology work: I have been building The Long Ascent, a turn-based deckbuilding game, to explore how far one person can take a complex project using current AI tools and a data-science mindset. mutwil.github.io/ 1/5

One could make even a bespoke version with claude code. We now live in an era where personalised software is so easy to do. I made a command center app just for myself, including short, medium, long time planners, email analysis and LLM-based replies, LLM-based project status...

🌱 Gene expression & co-expression across 630,000 RNA-seq samples from 10,700+ plant species — free, in your browser. Search a gene → see where it's expressed on the UMAP → explore its co-expression network. mutwil-plant-atlas-viewer.hf… #PlantScience #RNAseq #Bioinformatics

I am pretty happy about how the new updated version of our Plant Atlas is working huggingface.co/spaces/mutwil… . I selected trait 'Nodulating'. This compared gene expression in roots of nodulating vs. non-nodulating plants, and identified gene families such as nodulins and others.

🌱 504k public plant RNA-seq samples × 1,200+ species → live interactive UMAP atlas. Click any organ (256 PO terms) or evo-devo trait (57 of them) → top gene families upregulated pop up on the right. huggingface.co/spaces/mutwil…

6/Finding 3: The strongest correlate of impact wasn't how much a lab published, but how many distinct techniques it commanded. Methodological breadth, not volume. ⬇️

5/Finding 2: Over two decades, peripheral techniques turned over almost completely (RNA-seq, CRISPR, ChIP-seq replacing PCR, microarray, Northern blot), while a core stayed put. The field re-tooled its bench without broadening it. ⬇️

4/Finding 1: Plant Cell papers collapse into just 7 recurring "research recipes," each with its own signature of questions, methods and findings. ⬇️