Dispatch
Inherent Labs trained a 27B 'AI Scientist' on 310 paper-replication tasks and says it beats Opus 4.8 and GPT-5.5
Inherent published Faraday on Aug 14: a 27B model trained with long-horizon RL that uses coding agents as tools, aimed at replicating research rather than answering questions. The training environment, Replica, is 310 RL tasks drawn from 100 papers, each requiring the agent to recreate figures inside fixed time and compute budgets, scored by an LLM judge with auto-generated per-task rubrics validated against human studies. The claimed edge over Claude Opus 4.8 and GPT-5.5 is largest in meta-learning, structural biology and materials science — a reminder that a small model trained on the right long-horizon environment can beat a frontier general model on that environment.
↳ Follow the thread