Reddit user releases Nikusui-v1, a model manually altered via J-Space manipulation
A tool based on Anthropic's Jacobian Lens lets you export models with tweaked behavior—and one creator made an uncensored version.
Anthropic recently introduced the Jacobian Lens, a technique to interpret and manipulate the internal representations of language models. A Reddit developer, Extraaltodeus, expanded on this by creating a tool that can export a model after directly tweaking its J-Space—the model's internal Jacobian manifold. This allows manual behavior modification, essentially letting users 'abliterate' specific constraints or biases by hand, without retraining.
To demonstrate the tool's power, Extraaltodeus created Nikusui-v1, an uncensored LLM explicitly designed to generate inappropriate content. Released alongside GGUF quantized versions, the model is offered purely as a proof-of-concept for the new manipulation technique. While the developer emphasizes this is for science, the release underscores both the potential for customizing AI behavior and the urgent safety implications of such manual fine-tuning approaches.
- Tool based on Anthropic's Jacobian Lens enables manual export of models with tweaked J-Space
- First output is Nikusui-v1, an uncensored LLM for generating explicit content
- GGUF quantized versions released for community testing and feedback
Why It Matters
Shows a practical method for manual AI behavior editing, raising both customization potential and serious safety concerns.