Oct 3, 2026 · 53 min
Unrestricted AI tests the limits of safety and enterprise adoption
Inside The Startup Building Uncensored AI (Abliteration AI) | EP 2345
The episode examines who should control safeguards for powerful AI as businesses, regulators, and model developers struggle to define responsible use.
- 1Unrestricted models can serve cybersecurity, biology, and defense users when customers—not vendors—set application-specific safeguards.
- 2Enterprise AI adoption remains slower than the technology narrative suggests because permissions, security reviews, and administrative barriers still dominate.
- 3AI may reduce production costs without eliminating premium expertise, while solo founders remain constrained by coordination and execution demands.
Don't miss
Devin Thomas explains Obliteration’s layered safety approach: separating an unrestricted knowledge model from a dedicated safety model.
The brief
Jason Calacanis interviews Devin Thomas, founder of Obliteration, about building unrestricted AI models for cybersecurity, biology, synthetic data, and defense from an early home GPU cluster.
Obliteration removes baked-in refusals while letting customers define safeguards, raising a central tension: sensitive users may need flexibility, but dangerous applications still demand screening and accountability.
The discussion turns to model interpretability, frontier-model incentives, and layered safety, including a separate safety model alongside an unrestricted knowledge model.
GPU shortages and cloud support illustrate the practical bottlenecks behind AI growth, while enterprise adoption lags because permission, security, and administrative systems move slowly.
The broader debate asks what AI changes beyond models: whether professional services shift from hours to outcomes, microdramas build durable media businesses, and solo founders gain real leverage.
What was said on this episode
22 statements · 10 positive · 7 negative · 5 neutral
Frontier labs underserve professionals and companies needing sensitive AI use cases
“there's a whole bunch of professionals and companies out there that are essentially underserved or unserved by the frontier labs”
Listen at 2:03
Obliteration lets customers define their own AI safety guardrails
“our unique I guess innovation or idea on the, on the space is to allow customers to set their own guardrails”
Listen at 4:21
Major AI labs prohibit testing whether a company is vulnerable to live CVEs
“The big labs won't even allow you to test that”
Listen at 5:27
Companies should red-team AI agents before deploying them
“you probably want to red team those agents”
Listen at 6:20
Users are generally responsible for how they use AI tools
“the user of the tool today is kind of responsible”
Listen at 7:52
Chinese open models trail frontier models by approximately three months
“I would say today they're maybe around 3 months behind”
Listen at 9:46
Model capabilities may become similar enough that users cannot distinguish providers
“there just a point where like the capabilities everyone has is like roughly similar enough that the end user really can't tell the difference”
Listen at 10:01
Obliteration’s safety-policy layer adds little request latency
“it doesn't add a lot of latency to your request”
Listen at 12:59
Internet agents will execute commands according to their configuration
“if you unleash some agent on the internet, right, it's gonna go do whatever you kind of command it to do”
Listen at 14:53
Exploit benchmarks reward models for extending exploits further
“The reward function for the model is how far can you take this exploit”
Listen at 16:16
Obliteration routes user-defined policies through a separate safety layer
“users can then, when they set those policies, it just goes to the, the, the, um, the moderation layer or the safety policy layer”
Listen at 21:58
Obliteration plans to build a trusted brand for unrestricted AI
“our plan, myself and my co-founder, is really that to build a brand people can trust in this space”
Listen at 24:02
Obliteration focuses on business-to-business unrestricted AI applications
“we try to stay focused on kind of like the B2B use cases”
Listen at 25:16
GPU shortages are real and make reliable large-scale supply difficult
“the GPU shortage was fake, but now I've experienced it and it is, it is, it is, it's actually a real thing”
Listen at 27:06
Some professional services will continue charging by the hour
“you'll still have people who charge by the hour”
Listen at 33:57
Billable-hour pricing will not disappear despite AI
“I'm not buying that the hourly, uh, goes away”
Listen at 35:19
Microdramas could test intellectual property before expansion into other media
“I think it's an interesting medium for testing IP that then could go to other mediums”
Listen at 39:35
AI enables microdrama production at scale by reducing traditional production requirements
“there would be no practical way to do this if you had to make sets, if you had to make costumes, if you had to cast”
Listen at 41:33
AI video creation will eventually achieve a major breakthrough like Toy Story
“We're going to have a Toy Story moment”
Listen at 42:50
Solo-founder viability is constrained by workload, not individual efficiency
“the issue isn't the efficiency of one individual”
Listen at 48:25
Investor preference for multi-founder teams will not change merely because of AI tools
“It is not going to change just because of the AI tools”
Listen at 50:24
Solo founders with strong growth and product-market fit can raise funding
“if you have product-market fit and your consumer product's growing 5-10% a week, week over week, or 10-20% month over month, and you're a solo founder, okay, now you've just eliminated the product-market fit risk, so you will be able to raise money”
Listen at 50:36
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Featuring
OpenAIListen to the full episode and explore every guest, topic, and moment on PodLume.