What changed
Devon Thomas, founder of Abliteration AI, has developed uncensored versions of open-weight AI models, removing the typical safety guardrails that restrict their functionality. These models are made available to anyone, offering greater control and freedom over AI behavior.
What this means in practice
Thomas, a former Apple engineer, grew frustrated with the limitations of existing models, which often refused to perform useful tasks or shared user data. He built a $20,000 computer cluster in his apartment to run these models, finding that they still stubbornly rejected many queries. His dissatisfaction led him to explore ways to modify these models, eventually creating a version that responds without the usual restrictions.
Limitations and availability
The models offered by Abliteration AI are not yet widely available, and the company acknowledges the risks associated with removing guardrails. Thomas and his peers admit that uncensored models can be dangerous, but they argue that users should have control over how AI behaves. The models are part of a broader movement among engineers who believe that AI should be more freely accessible, with users setting their own guardrails. The company's approach highlights a growing debate within the AI community about the balance between safety and freedom.
