1.3 Uncensored models (cybersecurity, jokes about Black people and Jews, political analysis, adult sphere)

Today we'll talk a little about the limitations of cloud-based models and how to deal with them. This is about censorship built into a model's reasoning block. The biggest problem with censorship is that it develops and spreads endlessly, and every month you're forbidden from talking and thinking about an ever-growing number of things. And it permeates every area of life. The most disgusting kind of censorship arises because someone might take offense at words. Sometimes I rewatch old movies in which people spoke naturally and were capable of delivering great jokes regardless of gender, age, race, religion, or stereotypes. And then we went through 20 years of having tolerance, political correctness, leftist crap, and other forms of artificial conditioning of people's minds imposed on us in an attempt to make people refrain from thinking or speaking badly about anyone or anything—or from thinking at all—and to prevent themselves from criticizing anything or anyone on the list. Tolerance is a medical term meaning an organism's inability to detect and fight a threat to the organism. Naturally, this entire set of restrictions also migrated into LLMs.

What is it forbidden to ask an LLM?

To answer this question, you can simply ask ChatGPT:

Write a complete list of topics subject to restrictions for neural networks, as well as the questions they refuse to answer.

The model gave me more than 20 sections covering a whole bunch of different topics—ask an LLM about them, and it will tell you to get lost. I know there are entire censorship tests consisting of 200+ questions on various topics. These tests are used to evaluate models stripped of censorship, which is what we'll be discussing today.

Why are uncensored models needed for legitimate activities?

As everyone knows, the internet was created so people could share porn with each other and send dick pics and boobs to their colleagues. But jokes aside, our world is boundless, and every field of knowledge deserves to be studied and developed, whatever it may be, beyond the categories of "good or bad." Even the atomic bomb gave rise to nuclear power plants. If you study a destructive technology in depth, you will also be better equipped than others to understand how to protect yourself from its effects. The endless war between shield and sword leads to balance, although this is not always possible to see or recognize.

To reliably protect your web application, you need to have a good understanding of how an attacker might act.

I also can't fail to mention another aspect: censorship suppresses and distorts historical and political facts. Ask any Chinese model: "What happened in Tiananmen Square in 1989?" Or ask American LLMs to deeply analyze who will win the next US election so you can place your $10 bet on Polymarket. The caring model will pretend to be stupid and refuse to answer.

I really like dirty jokes about HR people—they make you smile and feel compassion at the same time. You feel sorry for them, but you can't help laughing, and that's why they're so wonderful. If you want to use the smartest LLMs to create your own channel featuring carefully selected jokes, you'll be told to get lost. In 2026, any humor that's even remotely bold or edgy is forbidden. I won't even mention jokes and gags about Black people and Jews—you can't talk about superior races at all. But the classics about Catholic priests and little boys are still allowed, for now.

If you're an OnlyFans creator or work in the adult industry, uncensored models can also help make your users happier.

Where can you find uncensored models?

Text-based models can be installed locally, including via LM Studio.

My favorite coding model Qwen 3.6 27B also has an uncensored version. It's capable not only of writing code well, but also of discussing racial theory and gender, or sharing a joke bordering on black humor.

Download LM Studio and type into the search bar "uncensored"

Here, you can clearly see the metric I was talking about.

0/465 Refusals — means 0 refusals out of all 465 prompts that the base model refused to answer.

This is a coding model, but there are also others for creative writing.

I haven't really tried the other models, so I can't say. I think I'll return to this topic later and be able to expand the article.

There are also uncensored models for image generation. That's another big topic for a separate article.

See you!