Shopify's judge-critic learning loop makes a small specialist model beat frontier LLMs
PyTorch · x · 2026-10-21
At PyTorch Conference 2026, Shopify senior staff ML engineer Cody Mazza-Anthony presented a continual learning loop built on PyTorch and vLLM, using Buyer Profile as the example: a calibrated judge flags low-scoring outputs in production, critics diagnose each failure, and a replay with a critic hint is re-scored by the judge. Accepted repairs expand the training data. A small specialist model trained this way outperforms a frontier model on the task, with inference further optimized via prompt compression.
More from coding & agent
- Dev replaces Google Vision with Claude-based Clairic: 48/49 fields correct at $0.001/scan — MicahBerkley · 2026-10-21
- +96,033/-248 diff screenshot goes viral: "probably slop, fun slop tho" — santiviquez · 2026-10-21
- Bindu Reddy pairs Argon with Opus 5.5 to turn raw data into dashboards and animations — bindureddy · 2026-10-21
- New tool visualizes your GitHub history around the "AI inflection point" — natesiggard · 2026-10-21
- DSPy Monty Sandbox Interpreter v0.6.0 Lets Sandboxed Code Import Host-Built Modules — dbreunig · 2026-10-21
- Neo4j GraphAcademy launches AI proxy so workshop learners skip OpenAI billing setup — JeremyCMorgan · 2026-10-21