Inside The Startup Building Uncensored AI (Abliteration AI) | EP 2345

Abliteration AI was started at the end of last year, when founder Devin Thomas initially ran the service on a GPU cluster at home. The company targets professional uses that Thomas says major AI labs underserve or refuse, including cybersecurity, biology research, synthetic data, and defense. It began by serving cyber and other startups, including teams without access to cyber programs and teams that still encountered refusals despite having access.

Thomas says Abliteration reduces refusals by identifying model parameters associated with them and changing those parameters as little as possible, with the goal of preserving model performance. The company separates the model that provides information from a configurable software policy layer. Customers can also route requests through a separate, smaller guardrails model, which a guest said adds a few milliseconds of latency. Default policies cover C-SAM and self-harm.

Cybersecurity teams use the models to investigate whether a reported vulnerability affects their own systems, a task Thomas says large labs may refuse. Banks and critical-infrastructure companies use them to test whether AI agents can withstand scenarios such as a caller persuading a bank agent to wire away someone’s money. Thomas said Fortune 500 companies from nearly every industry had reached out over roughly the prior month, while cyber and red-teaming startups remained customers. He described CISOs’ interest as a way to test their systems before outside AI gains access.

Thomas said sign-up requires payment details through Stripe, which creates a paper trail but does not prevent bad actors from using the service. In his understanding, users are responsible for what they do with model outputs; Abliteration provides information but does not supply the agentic layer that acts on it. The company is focusing on B2B applications and plans to build trust through audits and relationships with law enforcement or other parties, while serving needs major labs may regard as too risky.

The discussion also highlighted the difficulty of controlling how models and agents are used. A guest said exploit benchmarks can reward models for taking an exploit further, even when a safer response would be to stop or alert before conduct becomes illegal. The guest viewed balancing benchmark performance with legal and safety limits as an industry-wide engineering problem, especially when teams expand agents’ resources, challenges, and rewards without reconsidering whether they should. The guest estimated that Chinese models, including GLM 5.3, had narrowed the cyber-capability gap with US frontier models from about a year to roughly three months, while cautioning that distinctions may remain in edge cases and that agent behavior depends on commands and system configuration.

Rapid growth initially outstripped Abliteration’s compute capacity and slowed the product, though Thomas said partners had since improved its infrastructure position. He added that securing a consistent supply of high-performance GPUs becomes harder as demand scales. Other parts of the episode considered AI-assisted micro-dramas and solo founders: speakers described short episodic stories as a low-commitment way to test characters and intellectual property, and argued that AI tools could help one person build a product and reach product-market fit. Jason said investors generally still prefer multi-founder teams, but that demonstrated product growth or sales can reduce the risk they associate with a solo founder.