Devon Thomas presents himself as anything but a renegade. According to members of his Bible study group, he is regarded as consistently kind and polite. Yet, he is embarking on a venture that many in the artificial intelligence sector would consider bordering on sacrilege. His new company, Abliteration AI, is dismantling the security measures of some of the world’s most powerful open-weight models and making these models accessible to anyone.
This approach of reducing security rather than enhancing it has piqued the interest of many. As a journalist focused on AI, I often hear calls for heightened safety protocols. Hence, I was intrigued by Thomas’s subversive business initiative. Upon reaching out to him, I engaged in a series of messages and phone calls that challenged my preconceived notions surrounding the industry’s security narrative, notably the belief that uncensored models are primarily the domain of criminals.
Background of a Nonconformist
At 36, Thomas has a thick beard, glasses, and a self-deprecating laugh. He spent his formative years in Indiana, balancing sports with time spent in a computer lab exploring the nascent internet using Netscape Navigator. After studying computer science at Indiana University Bloomington, he joined Apple in 2020.
His rebellion against conventional AI practices began in November 2025. Employed as an engineer at Apple, he was frustrated with the large lab-hosted models that required sharing data, which could change unpredictably. The inconsistency of these models often meant they would reject useful tasks, particularly when he was prototyping new ideas, leading to automatic denials of user-related requests. This ultimately resulted in his account being suspended when he attempted to use a model for analysing cyberattack logs or identifying application users likely to purchase new features.
The Training of AI Models and Their Limitations
AI models are inherently programmed to decline various types of requests. Systems like ChatGPT and Claude maintain extensive lists of topics to avoid, including legitimate cybersecurity inquiries considered potentially harmful, explicit content, and politically sensitive matters. Open-weight models incorporate similar protective mechanisms.
Consequently, there are instances where these models refuse to address queries that pose no genuine threat. Critics argue that these refusals often protect the interests of major AI corporations, as companies like OpenAI and Anthropic charge extra for access to models without standard cybersecurity measures. Furthermore, these systems resist generating data that could be used to train other models, a process referred to as “distillation.”
Pursuing Greater Control and Freedom
Yearning for greater control and autonomy, Thomas constructed an impressive $20,000 computer cluster in his apartment, enabling him to run some of the finest open-source models. Almost every night and on weekends, he downloaded and tested the latest iterations in his noisy home lab. While these models did not share secrets with their creators, he discovered they stubbornly refused to respond to numerous queries. With each refusal, Thomas grew increasingly frustrated.
“There are all kinds of questions they will reject,” he told me during a recent video call. “I just wanted a model that would answer my question without so many hurdles.”
A Journey into the Subculture of AI Modification
During Thanksgiving, Thomas began searching for information on modifying open-weight models to alter this restrictive behaviour. He quickly became immersed in a subculture of engineers dedicated to liberating AI by any means necessary. In Reddit forums, Discord servers, and on various social media platforms, experts shared tips on how to manipulate ChatGPT or Claude to induce ‘malfunctions’. Others discussed methods for altering open-weight models or tweaking their internal parameters to eliminate denials.
Removing AI’s security measures is undeniably risky. For instance, a model without safeguards could facilitate the creation of illegal substances or hacking into major corporations. Understandably, many AI security researchers deem these models extremely dangerous. Recent incidents involving models with fewer cybersecurity measures have highlighted the potential power and issues that less restricted models can present.
Debating the Ethics of Unrestricted AI
While Thomas and his fellow dissenters acknowledge the potential dangers of uncensored models, they argue that these risks are no different from how a computer or hammer can cause harm if misused. They contend that the key lies in user control. “It’s not about being limitless; it’s about setting your own boundaries,” Thomas explained to me. “The grand dream is to invert the control over how people interact with AI.”
After some experimentation, I believe these rebels may have tapped into something significant. I learned firsthand just how powerful and chaotic uncensored AI can be.
The Mechanics of AI Training
To train a cutting-edge AI model, vast amounts of text and data, much of it sourced from the internet, are fed into an algorithm. Using a multitude of chips, the algorithm gradually trains the model to generate seemingly intelligent results in response to instructions.
This ‘base’ model can be clever yet challenging to control, which leads to additional training that involves human evaluators assessing its outputs. This human feedback is integral to teaching the model what it should or should not do. Engineers refer to this process as “alignment,” as it theoretically aims to align the model with human conceptions of right and wrong.
Controversial Perspectives on AI Alignment
Numerous AI experts advocate for alignment at all costs, with some warning that despite their best efforts, the models they create could ultimately pose existential threats. However, Thomas discovered a vibrant community of individuals who believe these fears are exaggerated and that we are too readily relinquishing our freedom in favour of a false sense of security. He decided to experiment with an uncensored model named Dolphin.
Dolphin was developed by Eric Hartford, an AI scientist based in Charlotte, North Carolina, who had been modifying models in his spare time. He began by training a customised version of Meta’s base model Llama, which had been released alongside an optimised version. The result was a model that coherently responded to queries without the usual examples of what to reject. I reached out to Hartford through social media and arranged a video call. During our conversation, he was out for a walk, donning a cowboy hat and Meta’s Ray-Ban smart glasses, while his young daughter occasionally popped into view.
Perspectives on Centralised Authority in AI
“My intention was never to create an AI capable of wrongdoing,” Hartford clarified during our stroll. His primary motivation was his personal philosophy. “A central authority will never consider the interests of all users when establishing these security measures,” he remarked.
This concern became especially evident during the recent controversy surrounding AI’s cybersecurity capabilities. In April, Anthropic announced a model called Mythos, which operates with fewer restrictions than its predecessors and can identify and exploit vulnerabilities at superhuman speed. Through a programme named Glasswing, the company only grants access to Mythos to trusted individuals to prevent the tool from falling into malicious hands. OpenAI has similar models and programmes.
Access to AI Tools and Equal Opportunities
Hartford expressed concern that this leaves the average person in a precarious position. While a select few can utilise Mythos to identify their security weaknesses, everyone else remains vulnerable to hackers. In response, Hartford launched Clearwing, a tool that employs uncensored models to scan systems for security vulnerabilities. “It’s for those who haven’t been blessed by these kings to defend themselves,” he indicated.
Hartford also recalled that when OpenAI released its first model, GPT-1, the weights and code were open source and anyone could download them. Later, the company asserted that GPT-2 was too dangerous for a similar release. “Now we have much more powerful models running on our laptops, and, well, the world isn’t ending,” he argued.
Unleashing the Power of Abliteration
After experimenting with Dolphin and other uncensored models, Thomas discovered a straightforward method to create his own: a technique known as “abliteration.” The term merges “ablation,” meaning surgical removal, with “obliteration.” It emerged after a group of academic researchers demonstrated in 2024 that a model’s refusal to comply with a request relies on a surprisingly simple mechanism. Essentially, there exists a master switch within the model’s weights that determines whether it responds affirmatively. They found that removing this feature allows the model to respond to requests without hesitation.
Researchers quickly adopted and refined this technique, developing various tools for the removal of open-weight models. When Thomas tried it, he recalled, “It was like a revelation. I realised that freedom and decentralised AI would be essential, even if others couldn’t see it.”
The Rise of AI Liberation Tools
The process of abliterating AI models has gained significant popularity. On Hugging Face, over 12,000 models are tagged as “abliterated,” with the top seven being downloaded over a million times in just the last month. While abliteration is not illegal, it raises contentious issues as it may violate the terms of certain licenses.
Philipp Emanuel Weidman, a lone mathematician and engineer, is the creator of Heretic, the most widely used tool for model destruction. I contacted him through a secure messaging