Summary
This video walks through how uncensored, or “obliterated,” AI models work and how to start using them, from a simple hosted chat all the way to running models privately on your own computer. Creator Jack Roberts explains that these models have the refusal behavior removed by identifying and stripping the part of the weights responsible for saying no, with a small impact on overall performance.
The tutorial covers three levels. First is API access through obliteration.ai and its obliterated large model, including a demonstration of an agent using an uncensored DeepSeek V4.1 Flash to reverse engineer apps and rebuild a Linear-style interface. Next is running models locally with Ollama, including a Qwen model on a Mac, so personal data never reaches OpenAI or Anthropic. A hardware-based spreadsheet helps pick the best local model for a given GPU and memory.
Roberts also makes a notable claim about the open-model gap: open models trail the frontier by roughly four months, with DeepSeek V4.1 Flash matching the best model from March 2026 and models that fit on a Mac being about nine months behind. He frames the video with strong warnings about responsible use, model ownership, and access risk.
📺 Source: Jack Roberts · Published November 07, 2026
🏷️ Format: Tutorial Demo
