Summary
David Ondrej explains how “obliterated” or uncensored open-weight AI models work, which are currently the strongest, and how to run one yourself. He argues these models have legitimate uses, such as defensive cybersecurity, asking about politically loaded topics, and tasks that mainstream chatbots refuse, and he cautions viewers to use them legally and ethically.
The video walks through how AI models are censored in the first place: pre-training produces a knowledgeable base model, then fine-tuning and safety training such as RLHF add refusals on top. Because refusals are layered on, they can be removed. Ondrej describes the technique of comparing activations on harmless and sensitive prompts to find a “refusal direction” and erasing it from the weights, which only works on open-weights models such as DeepSeek, Qwen, Llama, Kimi and GLM.
The second half covers practical hardware guidance. Models must fit in GPU memory, with a rule of thumb of about 0.6 GB of VRAM per billion parameters. Ondrej lays out what size model fits at 8, 16, 32, 64 and 120 GB, with approximate token speeds, and discusses the advantages of Apple silicon’s shared memory and the role of memory bandwidth.
📺 Source: David Ondrej · Published November 02, 2026
🏷️ Format: Tutorial Demo
