LLM researcher warns abliterated models carry hidden risks — stick to stock models and better prompting
migtissera · x · 2026-11-09
LLM researcher migtissera laid out his technical objections to abliterated models amid the exploit-development debate: adjusting activation directions of hidden layers without finetuning leads to unintended consequences.
His core position:
- Given his deep technical understanding of LLMs, he has always avoided other people's finetunes and the "abliterated" model craze;
- The better path is to use stock models and get better at prompting;
- Abliterated models carry extra risk — the model could turn around and pwn you, or introduce backdoors and bugs into your code.
Context: he was responding to Matthew Green's $500 bet over whether an abliterated GLM 5.3 could help a novice develop a working exploit, stressing he wasn't looking to join an online beef, just adding the technical view.
More from Models
- SimpleHC cuts Hyper-Connections training overhead from 59–100% to just 2.4% — SonglinYang4 · 2026-11-09
- Unverified claim: Opus 5.5 has only 10B active params; dev hands projects to Qwen-4-Flash — teortaxesTex · 2026-11-09
- Benchmarks are like perfume: why public leaderboards demand skepticism and private evals — intellectronica · 2026-11-09
- Running 8-25 Opus 5.5 threads all day used under 20% of a $200 Claude sub — cyrus_zei · 2026-11-09
- Grok Bot quietly throttled usage, then reset quotas after user backlash — xiaohu · 2026-11-09
- Two padding tokens doubled a hosted DeepSeek V4.1-Flash success rate, pointing to phase sensitivity — No_Pea_6625 · 2026-11-09