Devon Thomas does not strike many as a rebel. According to members of his Bible study group, he is known for being exceptionally kind and polite.
Despite his mild demeanor, Thomas is engaging in work that many within the artificial intelligence community view as controversial. His startup, Abliteration AI, focuses on removing the safety guardrails from the world’s most advanced open-weight models, ensuring these tools remain accessible to the general public without restrictions.
As a journalist covering the AI beat, I am frequently confronted with arguments for heightened safety protocols. This prompted my interest in Thomas’s subversive business model. Through a series of text exchanges and phone calls, Thomas shared insights that challenged my preconceived notions—specifically the idea that uncensored AI models are primarily sought after by bad actors.
Thomas, 36, is a man of simple presentation, characterized by a thick beard, modest glasses, and a self-deprecating sense of humor. Raised in Indiana, he spent much of his youth balancing athletics with time in computer labs, navigating the early internet via Netscape Navigator. After studying computer science at Indiana University, Bloomington, he joined Apple in 2020.
His journey toward AI activism began in November 2025. While working as an engineer at Apple, Thomas grew increasingly dissatisfied with the proprietary models hosted by major laboratories. He was concerned about data privacy and frustrated by the volatility of models that changed overnight. More importantly, he found that these models often refused to perform legitimate tasks during his prototyping processes. He encountered frequent blocks and even account bans when attempting to use models for benign purposes, such as analyzing cyberattack logs or predicting user behavior for app feature development.
Current AI models are programmed to decline a wide array of requests. Major platforms like ChatGPT and Claude maintain strict boundaries regarding cybersecurity queries, adult content, and politically sensitive topics. Even open-weight models typically come with similar pre-installed constraints.
This leads to instances where models refuse queries that pose no actual danger. Some critics argue these restrictions serve the business interests of “Big AI”; for example, OpenAI and Anthropic offer premium access to models with fewer cybersecurity restrictions. Furthermore, these models often refuse to generate data for the purpose of training other models—a technique known as distillation.
Seeking more autonomy, Thomas constructed a high-powered $20,000 computer cluster in his apartment to run various open-weight models locally. He spent his evenings and Saturdays testing new releases on his home lab. While these models offered better privacy than cloud-based alternatives, they remained stubbornly restrictive. Each refusal fueled his growing frustration.
“There are all types of questions they will reject,” Thomas explained during a recent video call. “I just wanted a model that answers my question without all the back-and-forth.”
During the Thanksgiving holiday, Thomas began researching methods to modify open-weight models to bypass these refusals. This led him into a niche subculture of engineers dedicated to “unshackling” AI. Across Reddit, Discord, and X, experts exchange various techniques—ranging from specific “jailbreak” prompts for Claude and ChatGPT to advanced methods for modifying the internal parameters of open-weight models to eliminate refusals entirely.