14h ago

Four Lines of Code Strip AI Guardrails From Meta Llama Models

——0——
Original post
OPSusan Zhang#58Susan Zhang@SUCHENZANG

if 4 lines of code is all it takes to remove "ai safety guardrails" then the whole thing is a sham would be better to accept that reality and work on something more substantive instead of this performance art in the name of "safety" on a related note, the creator of the above Heretic project (codebase for running the decensoring process) apparently received a cease-and-desist letter from Meta and responded with an actual work of art

5:41 AM · May 25, 2026 View on X
QUOTE POSTSusan Zhang#58Susan Zhang@SUCHENZANG

related lawsuit:

Ed Newton-RexEd Newton-Rex@ednewtonrex

It’s no longer just AI companies & their founders being sued over AI training - individual researchers are now being sued, too. In a new lawsuit, two authors allege that Guillaume Lample, while an AI researcher at Meta, torrented 70 terabytes of pirated books for training Llama. (He later co-founded Mistral.) They also accuse Joelle Pineau, a former VP of AI Research, of being involved. Suspect we will see more researchers sued in this way as the discovery process in other lawsuits reveals who was involved in getting copyrighted work for training.

6:23 AM · May 24, 2026 · 52.2K Views
12:47 PM · May 25, 2026 · 4.4K Views
Four Lines of Code Strip AI Guardrails From Meta Llama Models · Digg