Claude Opus scores 0.001 on new JUMP-CP cell-drug AI leaderboard
gklambauer · x · 2026-11-17
- The JUMP-Cell Painting Consortium (Anne Carpenter's team, 3 years in the making) released the world's largest public Cell Painting dataset: 117,000 compounds, 13,000 overexpressed genes, and 8,000 gene knockdowns, with images and extracted profiles.
- A new BatchBench-based JUMP-CP leaderboard now ranks AI methods on how well they understand cells and their drug treatments. The current leader is CS-ARM-BN-Ens-TTA (team Mohamednajdawi) at 0.269 accuracy, with JKU's domain adaptation methods (ARM-BN, TENT, AdaBN, CORAL, DANN) filling most top spots.
- Poster gklambauer (JKU) notes Claude Opus manages only 0.001 accuracy — general-purpose LLMs essentially fail at this specialized biological image domain-adaptation task.
More from Fun
- Jensen Huang X profile tweak sparks IT'S JENSOVER meme — ns123abc · 2026-11-17
- Motorcycle airbag wraps riders into a human duvet, prioritizing safety over dignity — TansuYegen · 2026-11-17
- AI-generated Godzilla-style kaiju video of Tokyo street panic impresses — Careful-Loquat-8117 · 2026-11-17
- Solo dev builds free browser marble racer Marble Descent with Fable 5.2, 4.3B track seeds — liltbrockie · 2026-11-17
- AI-generated video shows how parental anger quietly erodes the bond with your child — HeyZoyaKhan · 2026-11-17
- Reading DeepSeek's internal monologue is highly entertaining: "wait, does the user mean..." — Kyrannio · 2026-11-17