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

Episode Summary

Executive Summary: The episode centers on Obliteration, a startup that strips safety refusals from open-source AI models and lets companies set their own guardrails for cyber, red-teaming, and other high-risk enterprise use cases. Jason Calacanis and Devin Thomas debate AI safety, liability, model commoditization, and the growing demand from startups and Fortune 500s for more permissive models. The second half features Ask Jason calls on billable hours and microdramas, plus a broader discussion of AI’s impact on professional services and media production.

Main Topics: Obliteration’s model: removing refusals, restoring user control: Devin Thomas explains that Obliteration modifies open-source models by identifying and minimizing parameters tied to refusals, then adds a separate safety/policy layer so customers can define their own constraints. Cybersecurity, red-teaming, and authorized offensive testing: A major use case discussed is allowing security teams to test vulnerabilities, red-team AI agents, and evaluate defenses against the latest frontier-like capabilities that are otherwise restricted by big labs. Liability, misuse, and responsibility: Jason presses on who is responsible if the software is used illegally; Devin argues the user retains responsibility, while the company leaves paper trails and focuses on being a model provider rather than an agentic actor. Frontier-model gap is narrowing and commoditization risk: They discuss how Chinese open models like GLM and DeepSeek are closing the capability gap with U.S. frontier models, especially in cyber, increasing pressure on premium model differentiation. AI’s effect on billable hours and professional services: A live caller asks whether AI will kill the billable hour; Jason argues it won’t disappear but will shift work upstream and increase productivity, while subscriptions and project pricing gain ground. Microdramas and AI video production pipelines: A caller asks about short-form serialized dramas; Jason and Lon discuss AI-enabled production pipelines, microdramas as a testbed for IP, and whether the format is durable or just AI slop. Founders, solo teams, and startup execution: Jason argues AI helps solo founders build products faster, but investors still prefer teams of two or three because early-stage work requires too much task switching and resilience.

Key Arguments: Open-source models can be made usable for restricted enterprise work by removing baked-in refusals and reintroducing policy controls at the customer layer. Cybersecurity teams need access to realistic, unrestricted models to red-team agents, test vulnerabilities, and evaluate defenses before attackers exploit them. The user, not the model provider, should generally bear responsibility for illegal or unethical use, especially when the company is not providing an autonomous agentic layer. The capability gap between U.S. frontier models and Chinese open models is shrinking quickly, possibly to only months, which may commoditize model performance. AI will not eliminate billable hours entirely; it will reduce time spent on routine work and shift humans toward higher-value judgment, ownership, and accountability. Microdramas are attractive because they are cheap to produce with AI pipelines and can serve as a low-risk format for testing characters and IP. AI helps solo founders get to product-market fit faster, but investors still value multi-founder teams because of execution load and risk distribution.

Data Points: Gap between frontier and open models in cyber: ~3 months behind - Devin Thomas estimates Chinese/open models are now only about three months behind absolute frontier capabilities in some areas. Earlier gap estimate: ~1 year behind - He contrasts today’s gap with an earlier period when open models were roughly a year behind. Platform users: 16 million users - Odoo is described in the ad read as having 16 million users across more than 170,000 companies. Companies using Odoo: 170,000+ companies - Used in the sponsor segment describing Odoo’s scale. International hiring support: 200+ countries - Odoo ad mentions contractor hiring across more than 200 countries. Paid newsletter price: $10/month - Jason promotes the This Week in Startups newsletter subscription price during the live Q&A. Newsletter annual target: $100/year - Jason says he challenged Lon to create 100 worthwhile tactics at $1 each. Alternate newsletter target: $250/year - Jason suggests raising the newsletter to $250/year later. Microdrama length: 30 to 60 episodes of 2 minutes each - A caller describes the microdrama format as serialized short episodes totaling one to two hours of content. Fortune 500 interest: Many companies in almost every industry - Devin says the company unexpectedly received inbound interest from Fortune 500 firms, not just startups. Default safety categories: CSAM and self-harm - Devin states these are default policies in their safety layer. Model response latency: A few milliseconds - Devin says the safety layer is small enough not to add much latency.

Pivotal Quotes: "our unique, I guess, innovation or idea on the space is to allow customers to set their own guardrails" — Devin Thomas: Explaining Obliteration’s product philosophy: remove default refusals and let customers define policy. "The user of the tool today is kind of responsible" — Devin Thomas: Discussing legal and ethical liability for misuse of the model outputs. "the billable hour goes away. I just think humans move upstream and do more important work" — Jason Calacanis: Jason’s answer to whether AI will destroy billable-hour pricing in legal and professional services.

Implications: The episode signals a market for “unfiltered but controlled” AI in enterprise security and red-teaming. It also suggests AI won’t erase legacy business models, but will reprice labor, accelerate startup execution, and increase demand for trustworthy governance layers.