Heretic hits 35k stars: fully automatic censorship removal without post-training
gnukeith · x · 2026-10-14
Heretic, an open-source GitHub project with 35.2k stars, removes safety alignment from transformer LLMs fully automatically — no expensive post-training required. It combines an advanced implementation of directional ablation (abliteration) with a TPE-based parameter optimizer powered by Optuna, aiming to solve the "uncensoring makes models dumber" problem. Ships with multiple config presets under AGPL-3.0.
More from Models
- Frontier models get stale in two weeks, user says native video and audio are the missing piece — imjustnewatai · 2026-10-14
- Microsoft Patches 1,400+ Flaws in One Month, Credits LLMs for Bug Discovery — suchenzang · 2026-10-14
- Swapping to Antigravity CLI Lifted Gemini 4 Argon to Near Opus 5.5 in Real-Work Bug Benchmark — PawelHuryn · 2026-10-14
- Livestreaming Gemini playing Minecraft until it defeats the Ender Dragon — jarrodwatts · 2026-10-14
- Users push back on ChatGPT's feigned-concern interventions for sadness — whitelightstorm · 2026-10-14
- Rumor: OpenAI's Astra 6.1 scrapped over failed safety evals amid compute strain — imjustnewatai · 2026-10-14