Reddit
A $500 RL Fine-Tune of a 9B Open Model Beat the Best Frontier Model on E-Commerce Catalog Review at 68x Lower Cost Per Listing
Fermisense published 'When Machines Take the Wheel' on 2026-07-28, reporting that a task-trained 9B open-source model reviews e-commerce product listings more accurately than the best frontier model they tested, at 68x cheaper per listing, for a total RL fine-tuning spend of about $500. The framing is scale economics: eBay carries ~2.5 billion live listings and Shopify absorbs 10M+ product updates a day, so per-listing cost dominates any accuracy edge a frontier model holds. The post drew 225 points and 66 comments on Hacker News; note the site returned 403 to direct fetching, so the specific figures here come from the indexed article summary rather than a first-hand read of the full methodology.
↳ Follow the thread