Inside The Startup Building Uncensored AI (Abliteration AI)

Devin Thomas started Abliteration AI at the end of last year after building a home GPU cluster to serve models from his house. He met Jason Calacanis at Launch Festival, a founder event Calacanis runs, after reaching out when he saw Calacanis tweet about a project he was excited about. The company name comes from its core technical process: "abliteration" means finding the neurons in a large language model that are responsible for refusals and modifying them so the model no longer refuses requests.

The Unserved Market and Inversion of Control

Abliteration targets professionals and companies that frontier labs like OpenAI and Anthropic are effectively shunning. The use cases include cybersecurity red teaming, biology research, synthetic data generation, and defense work. Thomas gave a concrete example: a company hears about a new CVE (common vulnerability and exposure) and wants to test whether its own systems are vulnerable. Frontier models refuse to help with that kind of testing. Even large corporations that have access to special cyber programs from the big labs are still getting refusals when trying to do their jobs.