НЕЙРОСЕТИ БЕЗ ЦЕНЗУРЫ – ВЫ тоже МОЖЕТЕ ими пользоваться (я расскажу как)

Summary

The video explains how “uncensored” open-source AI models are made, how to assess whether a modified model is still capable, and how to run one locally.

Jailbreaks versus modifying a model

A jailbreak prompt tries to persuade a hosted chatbot to ignore its rules for a conversation. The presenter argues that this is not the same as removing safeguards: services can still apply separate input and output filters, log conversations, and restrict accounts.

Jailbreaks can also come with an “intelligence tax.” The video cites tests in which models prompted to bypass restrictions gave less accurate answers. It also warns that apparent compliance may be deceptive: a model can produce an answer that sounds dangerous while making it wrong or harmless.

Local uncensoring, by contrast, changes a model’s weights. The video describes older, costly retraining approaches and newer methods that identify or adjust parts of a model associated with refusals. It discusses the open-source tool Heretic, which automates this process, and mentions other weight-editing approaches.

These edits can damage general capabilities as well as reduce refusals. The presenter describes KL divergence as one measure of how much a modified model’s behavior has shifted from the original, and argues that careful, limited changes are preferable to broad, crude ones.

Choosing a model and judging its quality

The video compares open model families, highlighting Google’s Gemma, OpenAI’s GPT-OSS, and Alibaba’s Qwen. It recommends considering a model’s age and the quality of its modification—not simply choosing the largest model that will fit.

Download counts and claims such as “zero refusals” are not reliable measures of quality. The presenter recommends independent leaderboards and benchmarks, and gives examples of popular builds that reportedly performed poorly in tests.

A modified model may seem fine in ordinary chat while losing accuracy on harder tests. The video particularly emphasizes failures on questions that challenge common misconceptions, arguing that users should verify important claims themselves.

The presenter also cautions that uncensoring may work unevenly across languages: a model’s behavior on English prompts may not match its behavior in Russian.

Hardware and running models locally

The video offers approximate hardware guidance, noting that performance depends on memory and model configuration:

• 8 GB of VRAM: Lightweight models, such as smaller Gemma or Qwen variants.
• 16 GB: Some mid-sized models may fit, though with limited headroom.
• 24 GB or substantial unified memory: Larger models, including some Qwen variants. Moving part of a model into system RAM can reduce speed.
• GGUF files: The presenter suggests Q4 quantization as a general starting point and recommends leaving memory for context rather than filling the GPU completely. Mixture-of-experts models are described as a special case.

Guides and tools mentioned

• LM Studio: Presented as a beginner-friendly option with a graphical interface that estimates hardware needs and helps find model files. Its downside, according to the presenter, is that the Electron-based app uses memory.
• Ollama: Described as a popular, minimal option, but the presenter personally recommends against it and refers viewers to a separate post explaining why.
• llama.cpp / llama-server: The presenter’s preferred lightweight setup. It requires starting a server, after which the model can be used through a browser.
• vLLM and SGLang: Mentioned as alternatives that may suit larger-scale or multi-user inference better than personal use.

The presenter says a setup and configuration guide is available through the channel’s Telegram.

The video also warns viewers to download files only from verified authors. It cites a fake Hugging Face page that allegedly impersonated OpenAI and distributed a Windows data stealer.

Risks and other applications

Uncensored models can be useful in legitimate cybersecurity work, where commercial filters may block analysis of attack logs. But the same capabilities can support harmful activity, and the video says safeguards are difficult to remove selectively.

The presenter argues that platforms do not apply moderation consistently: one model build may be blocked while similar ones remain available.

A brief section covers image generation. It describes how strict filtering reportedly reduced the quality of Stable Diffusion 2, prompting community forks, and notes that payment-provider policies can influence where adult-content generators operate.

PewDiePie’s announced AI model is discussed as an example. According to the subtitles, it was based on a compact Chinese model with protections removed, but its public release had been delayed and its model files were not yet available.

Promotion

A sponsored segment advertises Skillbox’s “vibe coding for profit” course. It describes a workflow in which learners use natural-language prompts to build websites, bots, and other projects, then progress to coding. The segment mentions a subscriber discount and a bonus year of English study.

Main speaker and sources

• Main speaker: The video’s narrator, KZ_Lemon4ik. No extended interview speakers are evident in the subtitles.
• Sources and people referenced: Developers of Heretic and other weight-editing tools; researchers and benchmark projects cited in the video, including Anthropic, a Stanford open-generation observatory, and Habr; and reporting or analysis about Hugging Face and open-model communities. The subtitles’ auto-generated spellings of some names and benchmarks may be inaccurate.