AI Startup Builds 'Uncensored' Models

A Palo Alto-based startup, Abliteration, has released a new AI model, abliterated-model-large-v2, which is intentionally designed to not refuse harmful requests. This contrasts sharply with the industry trend of focusing on AI safety after a series of major hacks. The new model is built upon an open-weight base but has had many of its safety features removed.

Abliteration claims its AI can perform tasks like offensive cyber operations and agent testing that other models decline. However, the company states it will not generate child sexual abuse material or promote self-harm. The process used, orthogonalization, reportedly removes refusal mechanisms while preserving the model's core capabilities.

This development targets developers who find mainstream AI models, such as those from Anthropic, too restrictive. Critics, however, describe this approach as reckless and a potential catalyst for a 'grey market' of jailbroken AI models. This occurs within a significant regulatory vacuum in the United States, where open-source AI models are exempt from proposed voluntary safety checks.

Experts view Abliteration's release as a wake-up call for U.S. policymakers. The commercialization of dangerous AI capabilities without regulation, particularly when enabled by U.S. technology, is a cause for alarm, highlighting the pressing need for updated AI governance.