Harvey post-trained an open model with expert data, nearly doubling completed legal tasks
ccerrato147 · x · 2026-11-11
Legal AI firm Harvey already validated the specialized-model route: it post-trained an open model with expert data from Mercor and others, reporting nearly twice as many completed tasks on its held-out legal benchmark as the base model, without using customer data. The author argues that once a company's AI usage is large enough, training a specialized model pays for itself as the upfront bill spreads across many uses — so cheaper inference could actually grow Mercor's customer base.
Related event: Harvey nearly doubles legal benchmark performance with expert post-training(2 posts)→
More from Companies & People
- Google DeepMind hosting student researcher on 1000+ agent cooperation, London, Jan 2027 — HaydnBelfield · 2026-11-11
- Databricks teases next-gen Genie data AI coworker at DATA + AI LIVE with OpenAI's Greg Brockman — matei_zaharia · 2026-11-11
- Investor: Mercor's data and evals now core to nearly every frontier AI lab — marcbhargava · 2026-11-11
- markletree joins Anthropic to build its production agents platform — RealGeneKim · 2026-11-11
- Automating his own bug-filing job at Apple lifted output from 50 to 300 reports a day and earned a fast promotion — jdluk87 · 2026-11-11
- Gary Marcus: Goldman data shows Anthropic's Q4 growth slowing to ~11% QoQ — GaryMarcus · 2026-11-11