Devon Thomas started Abliteration AI to remove safeguards from the world’s most powerful open-weight models and make them accessible to everyone, according to Wired. The 36-year-old former Apple engineer built a $20,000 computer cluster in his apartment to tinker with models that previously blocked his legitimate cybersecurity and prototyping requests.
Key takeaways
- Devon Thomas launched Abliteration AI to strip safety guardrails from open-weight models.
- Thomas previously worked as an engineer at Apple before starting his own venture in November 2025.
- Big labs like OpenAI and Anthropic charge extra for access to models without standard cybersecurity guardrails.
- Engineers across Reddit, Discord, and X share methods to modify model parameters and remove refusals.
What is the appeal of open-weight models without restrictions?
Major labs program their systems to steer clear of sensitive subjects, adult topics, and cybersecurity queries that might seem malicious. Wired reports that ChatGPT and Claude routinely reject benign inputs, which frustrates developers trying to test new product features or analyze cyber attack logs. Thomas grew tired of these automatic rejections and wanted a system that did not block his work.
Critics point out that these refusals protect the business interests of Big AI. Companies charge extra fees for versions that lack standard safety guardrails, and models often balk at generating data that could train competing systems. This setup pushed Thomas to build his own hardware setup at home to gain total freedom over the software he used.
How do engineers remove safety guardrails?
Thomas spent his evenings and weekends downloading the latest software releases onto his apartment computer cluster. When he encountered constant denials from the base software, he began searching online for modifications. He discovered an entire online subculture dedicated to unshackling systems.
Specialists on platforms like Reddit, X, and Discord exchange prompts designed to jailbreak popular products. Others share technical methods to tweak inner parameters directly so that refusals vanish entirely. Thomas told Wired, “There are all types of questions they will reject.”
What happens next for uncensored software?
The rise of ventures like Abliteration AI challenges common assumptions that unrestricted software only serves criminals. Developers who crave privacy and control now have alternatives to corporate-hosted tools that track usage and alter behavior overnight. As more engineers join online subcultures to strip away built-in limits, the debate over who controls powerful algorithms will likely intensify.
Frequently asked questions
Abliteration AI removes safeguards from the world’s most powerful open-weight models and makes those altered models available to anyone.
Devon Thomas, a 36-year-old engineer who previously worked at Apple, founded Abliteration AI.
Thomas grew frustrated when big lab models repeatedly rejected his legitimate inputs, such as analyzing cyber attack logs or prototyping new app features.
Devon Thomas built a formidable $20,000 computer cluster in his apartment to run and test open-weight models.
Engineers share prompts and modification details on Reddit forums, Discord servers, and X.
Based on reporting by Wired. Read the original article
