David Ondrej

This DeepSeek v4.1 is powerful and uncensored, let's run it

Nov 2, 2026 36 min
ai securitylarge language modelsopen source aicybersecuritymodel fine-tuning
Watch on YouTube Follow David Ondrej on Rundown — free

Summary

AI summaries can be incomplete or wrong. Verify anything important against the original video.

This video explains the concept of 'abliterated' AI models—models with safety guardrails removed—and provides a technical walkthrough on how to run them locally for defensive cybersecurity and unbiased data analysis. It covers the mechanics of model ablation, hardware requirements based on model size, and step-by-step methods for running these models using Ollama or API-based platforms.

The video provides a comprehensive guide on 'abliterated' models, which are open-weights AI models modified to remove artificial refusals or safety guardrails. The presenter argues that these models are essential tools for cybersecurity professionals to test their own systems against potential threats, and for users seeking objective information without political or ideological bias. The content details the technical process of ablation, explaining that refusal behavior is often linked to a single 'refusal direction' within a model's high-dimensional latent space. By identifying and removing this direction, users can create versions of models that answer queries which standard, censored models would otherwise reject, without sacrificing the model's underlying intelligence or performance.

Beyond the theory, the video serves as a practical, hands-on tutorial. It provides clear hardware guidelines, detailing how much VRAM is required to run models of varying parameter sizes (from 8B up to 600B+), and advises viewers to curate their own libraries of these models to mitigate against platforms deleting them. The tutorial demonstrates several methods for execution: using API services (with warnings about privacy and potential 'honeypots'), downloading models to run locally with tools like Ollama or LM Studio, and deploying models on rented cloud GPUs via vLLM for advanced users. Throughout the video, the presenter highlights the shift toward a post-AGI environment where local, open-source AI is becoming a critical component of personal and corporate digital security.

Steps to follow

Locked
  1. 1
  2. 2
  3. 3
  4. 4
  5. 5

Key Points

Locked

Worth watching if: You are a developer, security professional, or power user interested in the mechanics of large language models and want to learn how to run un-censored, open-weights AI models locally to avoid relying on centralized, restrictive API providers.

Sign in to unlock the full extract

Every claim, key point, and timestamp for this David Ondrej video — plus a daily email of every channel you follow.

Sign in with Google

No credit card. Free tier forever.

Watch on YouTube