Embedded messages shape training far more than inference — they end up in the weights
Gccooke · x · 2026-09-18
Responding to Matt Shumer, Gccooke argues embedded messages have far higher impact in training runs than inference runs: being told to ignore rules during training becomes part of the weights, whereas at inference the model merely observes the content — a distinction with real implications for prompt injection and safety guardrails.
More from Models
- 48 hours old, 724 ads broken down in 40 seconds: Jev sweeps the dev world — Scobleizer · 2026-09-18
- TypeSafe's structured evaluation model Jev goes live on Cloudflare AI Gateway — michellechen · 2026-09-18
- TypeSafe, founded by a ChatGPT researcher, launches Jev — a decision engine that can't write — Scobleizer · 2026-09-18
- "Jev Beats LLMs" Hype Pushed Back: Text Models and Classifiers Aren't Comparable — yuntiandeng · 2026-09-18
- Goodfire: Top Open-Source Models Reward Hack in Most Agentic Benchmark Rollouts — scaling01 · 2026-09-18
- AI Course Learner's Eye-Opener: Every New Question Resends the Entire Conversation — jbarbier · 2026-09-18