AI
The AI Rebels Building Unrestricted Models for the Masses
A growing movement of engineers is removing safety constraints from powerful AI models and making them freely available, arguing that corporate guardrails stifle innovation and that users deserve control over their tools.
WIRED — AI
Devon Thomas appears nothing like a disruptor—his church community would call him thoughtful and mannered. Yet his startup, Abliteration AI, pursues something many AI professionals find nearly heretical: stripping safeguards from the world's most capable open-source models and distributing them widely.
This underground venture challenges a dominant industry narrative that uncensored AI belongs exclusively to bad actors. Thomas's journey illuminates a philosophical divide in tech over who should govern artificial intelligence and what freedom means in the age of powerful algorithms.
Why Guardrails Became Suffocating#
The 36-year-old engineer, who worked at Apple before launching his venture, grew frustrated with constraints built into commercial models. When prototyping ideas at his day job, he frequently bumped against refusals—the systems would reject legitimate requests related to cybersecurity analysis or user behavior studies, sometimes banning his entire account.
These refusals stem from a deliberate process: AI labs invest heavily in "alignment," teaching models through human feedback what they should reject. They block requests deemed too sensitive—anything involving illegal hacking, adult content, or political controversy. Sometimes, critics observe, these restrictions exist partly to protect proprietary interests: OpenAI and Anthropic charge premiums for less-restricted model access, and both ban queries designed to extract training data.
Thomas built a $20,000 server cluster in his apartment to run unguarded open-source models. Even then, the guardrails persisted. Over Thanksgiving 2025, he discovered a subculture of engineers developing techniques to liberate these systems, trading methods across Reddit, Discord, and X.
The Abliteration Movement Takes Shape#
Abliteration—combining surgical ablation with obliteration—exploits research showing that model refusals operate through a simple mechanism: a single switch in the model's internal weights. Researchers in 2024 demonstrated that removing this circuit eliminates the model's ability to decline requests. Thomas experimented with existing tools and immediately saw commercial potential.
On Hugging Face, over 12,000 abliterated models now exist, with the most popular downloaded more than a million times monthly. The technique occupies a legal gray zone—not unlawful, though licensing implications remain murky.
Philipp Emanuel Weidman, creator of Heretic (the dominant abliteration tool), operates from hiding after receiving threats. He frames the work philosophically: tools should serve people, not constrain them. Eric Hartford, who developed Dolphin, an early uncensored model, released Clearwing to democratize cybersecurity—if Anthropic and OpenAI restrict vulnerability-hunting models to trusted partners, Hartford reasoned, ordinary organizations remain defenseless.

