Jeremy Howard on LLM Anti-Jailbreak: Banning Prefilling Drives Users to Open Source
jeremyphoward · x · 2026-08-01
Discussing safety filter bypasses, Armin Ronacher (mitsuhiko) worried that commercial AI labs will crack down on 'prefilling' and ban model switches. Fast.ai founder Jeremy Howard replied that they used to rely heavily on prefilling before labs killed it, forcing them to use synthetic prepended dialogs instead. He warned that if commercial labs continue banning these technical workarounds, open source will become the only option for users.
More from Models
- Benchmarking Agent Harnesses: Kimi K3 Shines, Claude Code Costs 4x More — omarsar0 · 2026-08-01
- Agentic RL Gives LLMs 'Vitality': Models Become Frantic and Engaged — teortaxesTex · 2026-08-01
- Building apps directly on iPhone using ChatGPT showcased — reach_vb · 2026-08-01
- Kimi K3 Tops Open-Weight Models on ARC-AGI, Rivals Claude Opus — mhmazur · 2026-08-01
- Claude V4-Flash Reported to Have Vision Defects, Frequently Building Workarounds — teortaxesTex · 2026-08-01
- OpenAI Removing GPT-5.4 Models from ChatGPT Starting August 31 — OpenAIDevs · 2026-08-01