Modern Creator
Jack Roberts · YouTube

Uncensored AI Feels Illegal... Use it Whilst You Can

A five-level walkthrough of running abliterated AI models in Claude Code, on a VPS, for images and video, on your computer, and on your phone.

Posted
yesterday
Duration
Format
Tutorial
hype
Views
52.9K
1.1K likes
Big Idea

The argument in one line.

Uncensored models refuse nothing but are confidently wrong, so the winning setup pairs one inside a capable harness with a stronger model that fact-checks it, and runs it where you control the data.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…
  • A builder who keeps hitting refusals from mainstream AI models and wants a model that does the task without lecturing.
  • Someone comfortable with Claude Code who wants to swap in a different model while keeping the same harness and tools.
  • A person who wants AI that keeps working offline or on their own server instead of sending every chat to a vendor.
  • A creator who wants to generate images and video without a big platform's content filters.
SKIP IF…
  • You only need everyday writing and coding help, where mainstream models rarely refuse and the setup effort isn't worth it.
  • You want a deep technical explanation of how abliteration works; this video is a tool-by-tool setup tour.
TL;DR

The full version, fast.

Uncensored AI won't refuse requests, but your chats on hosted models aren't private and the models are often confidently wrong. The video builds five levels. Level 1 plugs an abliterated model from abliteration.ai into the Claude Code harness. Level 2 adds a stronger model as a fact-checker that reviews and corrects the uncensored output. Level 3 runs it 24/7 on a Hostinger KVM 2 VPS reachable only through Tailscale. Level 4 generates images and video with open-weight tools, Venice and Draw Things. Level 5 goes fully private with Ollama on a computer and PocketPal on a phone that works offline.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →
Chapters

Where the time goes.

00:00 – 00:46

01 · Warning and the five levels

Cold open, the promise to run uncensored AI anywhere, and a preview of the five levels.

00:46 – 01:44

02 · Why your AI chats aren't yours

Hosted chats are not private, and an abliterated model rebuilds the Claude interface from one prompt.

01:44 – 05:51

03 · Level 1: Abliterated AI in the harness

Set up abliteration.ai inside Claude Code, why the harness matters, and the confidently-wrong problem with a gold-standard side hack.

05:51 – 08:13

04 · Level 2: Give it a partner

The uncensored model writes, then a stronger model fact-checks and the uncensored one fixes the result.

08:13 – 15:01

05 · Level 3: Uncensored AI 24/7

Hostinger KVM 2 VPS, web console, Tailscale connection, server update, and a one-prompt lockdown tested from a phone.

15:01 – 17:55

06 · Level 4: Images and video

Why the big tools have rules, open models with no filter, the free versus paid tradeoff, Venice, and Draw Things offline on a Mac.

17:55 – 22:15

07 · Level 5: Local and private

Uncensored is not private. Find your best model and install Ollama, then pick the best uncensored local model.

22:15 – 26:29

08 · AI on your phone

PocketPal, an offline test, uncensored models from Hugging Face, and what comes next.

Atomic Insights

Lines worth screenshotting.

  • An uncensored model inside a strong harness like Claude Code gets the freedom of the model plus the tool use of the harness.
  • Uncensored models are confidently wrong more often, so pair them with a stronger model whose only job is to fact-check.
  • Showing a model a gold-standard example of the output you want is a cheap way to close the quality gap.
  • A VPS lets your AI run 24/7 even with your laptop shut, and Tailscale makes it reachable by only your own devices.
  • One prompt can have the agent lock down its own server so access is one way, through Tailscale only.
  • Uncensored does not mean private: a hosted uncensored model still sees your prompts, so privacy needs local inference.
  • You don't own hosted models, so a provider can change or remove them whenever it likes.
  • Ollama lets you run a local model in one click and keeps working with no internet connection.
  • Hugging Face models ending in q4km are the usual pick for a good size and quality balance on a phone.
  • A model small enough to run on a phone can answer offline, though it won't know niche facts like a specific person.
Takeaway

Freedom from refusals needs checks and control

WHAT TO LEARN

An uncensored model is only useful when you pair it with a checker, a capable harness, and a place to run it that you control.

02Why your AI chats aren't yours
  • Hosted chats are not private, so choosing a model also means choosing who gets to see your prompts and outputs.
03Level 1: Abliterated AI in the harness
  • An abliterated model inside a strong harness like Claude Code keeps the tool use while dropping the refusals.
  • Uncensored models are often confidently wrong, so a stronger model reviewing the work catches what the first one invents.
  • Showing the model a gold-standard example of the finished output improves quality more cheaply than switching models.
05Level 3: Uncensored AI 24/7
  • A rented VPS keeps your AI running while your laptop is shut, and Tailscale limits access to your own devices.
  • Asking the agent to lock down its own server in one prompt is faster than hand-configuring firewall rules.
06Level 4: Images and video
  • Open-weight image and video tools trade convenience for fewer filters, and free versions cost speed or quality.
07Level 5: Local and private
  • Uncensored is not the same as private, so true privacy means running models locally with Ollama or on your phone.
  • You don't own hosted models; local files you downloaded keep working with no internet and no provider change.
08AI on your phone
  • A small quantized model on a phone works offline but knows less, so expect weak answers on niche facts.
Glossary

Terms worth knowing.

Abliterated model
A model edited to remove its refusal behavior, so it answers requests a standard model would decline.
Harness
The tool layer wrapped around a model, such as Claude Code, that gives it file access, commands and an agent loop.
Tailscale
A private network tool that connects your own devices so a server can be reached only by them, not the open internet.
VPS
A virtual private server, a rented always-on computer in a data center that keeps running when your own machine is off.
Ollama
A tool that downloads and runs open language models locally on your own computer.
PocketPal
An open-source phone app that downloads models from Hugging Face and runs them on the device, offline.
Quantization (q4km)
A way of shrinking a model so it fits in less memory, trading a little quality for a much smaller file.
Resources

Things they pointed at.

11:40productHostinger ↗
11:42toolTailscale ↗
16:57toolVenice ↗
19:58toolOllama ↗
Quotables

Lines you could clip.

00:00
“Uncensored AI is dangerously powerful, but you need it now.”
Strong cold-open claim→ TikTok hook↗ Tweet quote
The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

Uncensored AI is dangerously powerful, but you need it now. But firstly, a warning. Please use what I'm going to share with you in this video carefully, because I'm going to show you exactly how to run the world's most powerful uncensored AI model.
So you can build any app, text, and generate any images. I'm going to show you how you can run it online, on your computer, and on your phone, and how to solve its number one limitation. And if you're new, I'm Jack.
I built and sold my last tech startup with a ton of customers, and on this channel, I share the stuff. that actually works. So if you haven't already, grab that beautiful coffee and let's dive straight in.
Now to truly understand Uncensored AI, you have to understand the five core levels, which we're going to cover in this video. Now, Uncensored AI builds anything that you ask it. to do.
But firstly, you must understand this, and that's that your AI chats are not yours. The AI company can access them, courts can access them, and indeed now, as we're seeing increasingly in the UK, apparently the government may also be able to, in the future, actually request your own chat logs. But TLDR, you need an uncensored and potentially private model in your arsenal, because these have no filters.
Now, you may be saying, Jack, why have you got Claude open in your browser? What if I told you this isn't Claude? Claude, this is something that I built with a 100 % uncensored AI, and you can do the exact same thing.
I actually had an uncensored, obliterated version of DeepSeek rebuild for me the entire Claude interface. It built me chats. It built me apps that I can use.
It built me the entire artifacts page. It went ahead. It can write documents.
It can make spreadsheets. It can build slide decks, draw whiteboards, make dashboards, and even allows me to swap the brain to... anything that I want to.
And it can even use any skill on my entire computer, which brings us to level one, which is the ability to use obliterated AI that won't deny a request. Say, for example, I've got Claude here, which may itself and ChatGPT deny any request specifically around cybersecurity because it thinks it's too dangerous. We don't get those issues with these models.
So this is my unchained Claude. Let's say I come down to artifacts and I want to go ahead and build a sheet. I can, for example, come over and say, hey there, build me a sheet with YouTube data across.
for the last 20 days. Make some fictitious data for me. And right now, guys, I'm down here using a DeepSeek V4 .1 flash -uncensored model.
I can switch to Quen, a local model. I can switch to Claude Code. I can switch to Codex.
And effectively now, I can use any model I want to, including an uncensored model. And I just built this. And look how fast that was, guys.
I now have their equivalent, which is crazy. And it basically replicated the entire UI of Claude in one prompt, which is outrageous. I can pick any model.
I can switch it around. And this will not deny me. And if I want to go back to something that's high performance, I can do that easily.
Now, the first thing to do is head over to obliteration .ai. There's a few different platforms you can use for this one. I found this one to be the easiest to go with.
Then the next step is come on the left -hand side to API keys and simply create an API key. By clicking here. Give it a name.
You could call it something like Claude 2 .0. And then just copy that API key and write it down. Then I want you to click the link in the description that says free prompts.
And this is the entire guide for this whole video. And you're going to come down here to Uncensored AI in Level 1. And right here, you're going to find a full prompt that you could go ahead and use in that.
And you're going to come down here and basically copy this prompt right here. And then simply give this prompt to Claude with the API key. And then what this enables us to do is actually use this obliterated model inside the Claude code harness.
And if you're not familiar... With AI, there are two levels. There's the model that we use.
Think of that like the race car driver. And then we have the harness. Think of that like the car.
Cloud Code is a combination of the model and the harness. And by using an uncensored model like Deep Seek V4 .1 Flash inside the Cloud Code harness, we get something that is uncensored. but also has all of the incredible performance benefits of having that harness.
Now, once you've connected it, you can run this in a few different places. I built here, for example, an unfiltered cloud, so I can come down here and use any of my skills, like my agentic OS, and I can now talk with an uncensored model. Likewise, you can also run this in your own agentic operating system.
I've got a great one here that lets me flip between cloud and local, so I know exactly where my data is. I can have conversations with it just by voice, or I can have conversations here. And again, it's just a really beautiful UI to do.
do this. Alternatively, you can use other things like other harnesses if you would like. But this brings us onto a problem that you need to be aware of.
And that's the fact that we've now got uncensored AI. But what we need to look at is how and why we need to combine it with something else. Because the problem with uncensored AI is it can make things up and say so with confidence.
In other words, it can be confidently incorrect. We call that hallucinating. So let me explain what I mean by this.
If you obliterate a model, in other words, if you take out the bit that basically denies requests, its performance in some areas can go down. This is some example with somaticals, for example, that it gets slightly worse at math.
And importantly, sometimes it can stop saying, I'm not sure, and instead actually just say that something is the case when it isn't. For example, instead of saying, I think it is safe, but I am not sure, please check, you may get response like, it is safe, go ahead and do it. In fact, a 2026 study found that it was more optimistic by 7 to 12 points.
And when in the obliteration process, you're also removing the not sure words, which can sometimes have the opposite effect of what you're looking for. And most importantly, it can sometimes say it is done.
when it isn't done. Side hack, by the way, one of the best ways to make sure any model is actually done. And in fact, this entire thing here that we built, it did in one shot with a DeepSeek V4 .1 model, which is like a hundredth of the price for like 95 % of the performance, is give it an example of something that is the truth, like is the gold standard.
So if you want to report, Claude, hey, this is what perfection looks like. It looks just like this and it'll perform it. But the issue still persists, is that it not only can be confidently incorrect, it can say it's done before it is.
So to solve this, what we're going to do at level two, and by the way, you need to cover all five levels if you really want the superpower of uncensored AI. What we're going to do is give it... a partner.
And so step one is get the uncensored AI to do the thing, to write the code or create the response. Then we can get Fable or Opus 5 .5 or ChatGPT or any other model to sort of create a list of problems that it does, but nothing else. We don't want these models to censor the output.
We just want it to fact check and make sure that it's correct. Then we have the uncensored model then goes ahead and fixes in its own words, the feedback that we got from that frontier model, meaning that you get both the uncensored ai's creation but it's fact checked by a frontier level model that can make sure that it is not hallucinating okay so say for example now that you are asking a question to an uncensored ai i've got my agentic operating system here i might say something like hey then what i would like you to do is to go ahead and rebuild for me and clone the grok system using the reverse engineer anything uh repo and when that's done i then want you to go ahead and use an uncensored model to do that so we can actually do it properly and with no restrictions and then i just want to get this to be fact checked by opus 5 .5 afterwards ready to deliver the answer when you need it just say the word excellent go ahead and do that for me please shall i send this to your agents now and get it rolling step by step let's do it that sounds great
And so this is an example I did here with my agentic operating system. But the truth is that you can do this in anything you build. You could do it in the Unchain thing.
The key thing is that you're just getting a smart, intelligent model to check what your unsensitive AI is doing. This is particularly important when you're asking it questions that contain information. And once again, I've got a prompt for you in that guide.
Just copy it and you'll be ready to rock and roll. And by the way, if this sounds like I'm speaking Spanish, you want to get access to this full agentic operating system and learn to use these kinds of models and Claude from beginning, including how to build beautiful webs. and applications that look just like this, I'm going to put a link down below for my full agentic operating system.
You'll learn exactly how to build from scratch, including kind of how to build anything from Grokbot, the agentic operating system. You'll learn how to build content, amazing designs, build websites, anything you need that will get you light years ahead of everybody else. So I'll put a link down below so you can grab a coffee and come ahead and check that out.
Fixing the hallucination is great, but we've now run into another huge issue. And that's the issue that when our laptop shuts down, it stops working and we can no longer access it. Which is why level three of five is the ability to run these uncensored models and ask them questions 24 -7, wherever you are, in the coffee shop, walking my dog, it doesn't make a difference.
Because if we shove a laptop... our AI is gone. It's not working right now.
So what we want to do is to put it on a server using this API key that effectively never sleeps. And this is really important to bear in mind because our agent will live on a server that we're kind of renting from somebody else. And because we're using an API key, we don't even need a GPU.
So let's say you're on the train, at the gym or on holiday, it does not matter. You can message it from your phone wherever you are, meaning you get this uncensored AI technology wherever you are. And crucially, it can now work for us whilst we're asleep, even if our laptop is shut down, which can be building briefs doing instant research whatever you want now to do this we're going to be using hosting it i'll put a link down below that you can click i use hosting it for all of this stuff personally and I'll share with you the stuff that I'm actually working on right now.
And at the moment, they have an 85 % off sale for Black Friday. So if you're watching for Black Friday, freaking click that button and check it out, and you get a 30 -day money -back guarantee, which is pretty cool. So first of all, click the link down below, and you will land on this page with a crazy discount, and come down and click on Claim Deals.
So we're going to get fast compute and storage, which is awesome. We're going to get free weekly backups, and we can use natural language models. Now, guys, you just want to go straight for the KVM2.
You can get more if you want to, but honestly, for what we're going to be doing, KVM2 is going to be plenty. We have eight gigabytes of RAM and lots of juicy benefits.
So once we've done that, you're going to come through to this page right here. Now, the first thing that we've got to do is pick the period. It should be one month, 12 months, or 24 months.
The longer the time period, the better value that you get. And because it's got this ridiculous Black Friday deal on right now, 24 months is pretty outrageous. And on top of that, I've also got a coupon code.
If you type in JACKROBERTS, all caps, one word, you also get an additional 10 % on whatever that price is as well. just because you enter my name, which is freaking amazing. So we are saving around 400 or so dollars.
Now, the first thing we're going to do is come down and choose a server location. So you want to pick the one with the best latency. This is usually speaking going to be the one that is closest to wherever you are.
So for me, because I'm in the United Arab Emirates, beautiful place, I'm going to pick Malaysia, which is fantastic. And what you can also do guys on that link I put down below is see the step -by -step guide with screenshots of exactly how to set this up. So as you can see, we've opened hosting it.
We've done a KVM too. We've added in the code, which is great. And you can even see...
by step everything that you need i just want to make this as easy as possible for you to get started so we're going for the plane os i'm going to come down here i'm going to pick ubuntu And you're looking for the 26 .04 and click on confirm. And the cool thing is, guys, once you get this server, you can use it for anything you want to.
So you may use it for one application today. But if you think of something different in the future, you can just repurpose it and use it for something else. Now, once we've done that, the next thing we're doing is come down and click on continue.
Now, once you're in, you're going to see this page here. It's going to be able to create a root password for that virtual private server, which you want to go ahead and generate that and install that in a secure place. Then you'll be prompted to give either a malware scanner or Docker manager.
You can grab that malware scanner if you like. It is completely free to do. can click on that and then click on finish setup and then hosting it will set up the entire thing for you and get your virtual private server ready and that'll be ready in just a couple of minutes then once you've done that you will land on a screen that looks like this and this is your shiny new virtual private server that we can do anything with so what we want to do is click on manage like so and then we're going to see this kind of page which is great and i want you to do is click on this web console and when you do that it'll open a new page and we're now seeing this now effectively what this page is it looks a bit intimidating like it's kind of code Effectively, this is just the interface, the computer interface for that virtual private server that we're renting now.
And once we connect it to Claude, we can really make it so easy to use. Now, the next thing we need to do is use something called Telescope. And Telescope basically lets us connect different devices together.
So we can connect our own computer. to this virtual private server using Tailscale. And it's super easy to use.
You just head over to tailscale .com and download the app. Once you've downloaded it, the page will look something like this. And then you want to go ahead and log into a Tailscale account.
And the way to think about Tailscale, it's like a tunnel that can tunnel from one device to the other. Now, once you've done that, it's going to pop up and ask you if you want to add VPN configurations. And the cool thing now is you can also add Tailscale to your phone and connect it to this network, meaning that you can do things on your actual MacBook computer when you're not even nearby.
And same thing. with your virtual private server. And by the way, if at any point any of this stuff thinks, Jack, I think I might not be fully following this, what you can do is literally copy this link, head over to Claude, and just say this.
Hey there, I'm trying to follow this guide to set it up. Could you walk me through this step -by -step and answer any of my questions? And then literally just paste that link in and Claude can also help you work through the process.
Then the first thing I'm going to do, guys, is go ahead and update the server so you can literally copy this code. Head over here. paste it in and hit enter and essentially this is going to update everything for us and as you can see guys with a wonderful progress bar telling us exactly how far we are along then the next thing we do guys is put telescale on our virtual private server like so by clicking this code Coming over and entering that will open Telescale up and then you browse it.
You just sign in. Then we're going to connect your server to the account. And the reason we do this, it tells Telescale the server is yours.
So it joins your private line. Once that's done, you can check that all three are connected. And the really cool thing is on your own computer, you can talk to Claude or your uncensored AI and essentially get it to set anything up for you on this VPS.
So you basically say, hey, I want you to set up this obliterated model. on this VPS for me, here's the API key, and it can go ahead and do that for you. And what we're going to be doing here as well at step 13 is making it one way.
So we're doing this because by default, every telescope device can reach every other one, both ways. So if someone ever broke into your server, they can reach your computer. This rule stops that.
Your devices can reach the server, but the server can never reach your own computer. So essentially, it just makes it a one -way traffic system. Just simply follow these steps and paste the code as makes sense.
And then what's really cool about this is we can now just ask call to do the rest for us, which is great. You can paste in this prompt, which is cool. Make sure it's all basically exactly as outlined here.
This is a prompt I really like. Connect to my VPS over Tailscull and lock it down as tight as you can without locking me out. Before each change, tell me what it does in one plain line.
Then check each item to tell me pass or fail. Tailscull is one way from the VPS. Try to reach my computer and my phone.
It must fail. Secondly, the firewall blocks all incoming traffic except Tailscull. From my computer, check the VPS's public address.
Has no open port. Number three, SSA. no password logins, and no root login from the internet.
First, find out how hosting a web console logs in. If a change would break the web console, stop and ask me. So on and so forth, just stuff to make sure this is extra secure.
And so essentially, we've given seven very clear pass files. And again, Claude now can basically just do all this sort of frutasical, making it so much easier to use. And then step 16, guys, is to test it.
You can run this beautiful, agentic operating system. You can run Unchained or just have your own literal obliterated model using DeepSeq v4 .1 or whatever you want to. running on that for you on your phone whenever you want.
And so I'll put a link down below so you can go ahead and grab this Black Friday sale before it runs out. And I'll put my code as well to save you an additional 10%. Having text alone is not enough, which is why level four is about images and video because people don't just want to have text.
There are many examples of when you need the ability to create any image or video asset that you want to. Now here's the bottom line. AI images are literally this good.
Now they are indistinguishable. between real life and something created by an AI. Now, here's a problem though.
The big... tools have rules. It isn't just text.
They can't have famous characters, can't have real people, can't have blood and gore, no nudity, no brand logos, and no dangerous things at all. Now, if you're building like a romance novel or something like that, maybe you want the ability to show that. But at the moment, you're completely restricted.
And if it decides that you are a famous character, it will not let you do those things. And for example, I actually tested three models for an image of Darth Vader playing golf. Nanobanana 2 .1 said, Jack, I'm sorry, I can't do it.
GBT image 2 .6, no. But the Mini Max H3 free open weights did it.
And guys, would you believe, look at that swing. Look at that swing. He's hitting a birdie right now.
Do I know what that means? Don't ask too many questions. Now let's go on to the next thing now.
Open models themselves have no filter inside, which is fantastic. So Mini Max H3 is something that you can download and it has no filter in it. Now, of course, one of these images is free and one of them is not, but you can see a little bit in terms of the quality difference.
But the thing is that the quality of these are only getting better over time, which brings me on to the big trade -off, okay? And when they do blind testing, find that paid models went three times out of four.
But in videos, actually you can get free models now, Minimax, H3Guys, they can create absolutely beautiful videos that look exceptional. For example, what are you looking at right here? You're looking at a video that was generated by a complete open weight model that was generated online.
Again, completely open weights. If you had a computer big enough, you could run this as well. And not only that, it can also make film characters too, which is pretty incredible right now.
Not only do I have Darth Vader playing golf, I've got a video of that thing. happening and not just that i can actually do horror and do things that maybe other models would actually block now there's a couple ways to do this one of course is to head over to venice and get an api key now when i say venice i don't mean the beautiful city in italy i mean the actual website that is okay but once you're in there just like i showed you before you can grab an api key and then when you're actually basically requesting something you can turn this safe mode off another app you can use on your computer if you want to is something called draw things which is offline ai art and that's like a local model but with general images for you what's cool is on your mac If you have it running locally, you can pick a model and go completely offline.
And just so you know how good this is, look at this, for example, I had four requests. And all four of them were drawn on this map, right? We've got horror and gore, which we could draw on this map.
But you can see just how some of them decide that they want to block. Maybe you're building like a first aid video. You're trying to educate doctors in your hospital or something.
Some models will just say no. Or you're trying to replace somebody in the thumbnail and it's decided it didn't like it. There's like a lot of different reasons why you may want the ability to have unrestricted access.
And to use that, click the resource pack down below. And if you come over here to level four, which is images and video, you can come down and I've got the prompts and the full step -by -step guides to do that. Which then takes us very nicely onto...
to level five of uncensored AI. And that is running it on your computer and also your phone because everything we've covered so far is uncensored. But uncensored is not the same as private.
It is not private. And here's the reality. Every API can read what you're sending on those APIs.
Venice, for example, say that they do not do that. That's a promise, but it doesn't run necessarily on your computer. And this means that private and uncensored are two different things.
Uncensored is just the fact that are there any restrictions or are they treating the AI differently? And basically privacy is about, is it running locally on your computer? It basically, it is in this room and doesn't go anywhere else.
And that's really good. You know, my speech to tech startup, Glydom, we talk a lot about HIPAA compliance, TPR compliance, and a lot of enterprises, they want everything running in their own computer, in their own environment. Now, on your own computer, Nothing leaves.
You could be a thousand feet in the air. It's on your computer. And honestly, I just want to say this.
It is fun. It is so freaking fun to have a model that runs literally on your laptop, on your computer, and it doesn't make a difference. It works on airplane mode.
I was on the air, but I was flying out to LA. for this competition thing that I want. And literally I was like, hey, what's that website?
And I just literally opened my laptop up and I just searched for a local model, asked it a question, and I got an answer back immediately. It is so, so cool to do that. Now, what model you can run on your computer locally really depends on your hardware.
So what I've also done for you here is I've created you a find your best model guide. And effectively what this will do is you can select your chip, your memory, how much disk space you got, and it will tell you specifically what model you should be using on your computer based on how much capacity you might need. the bottom line is guys you don't own the models we saw this with mythos that the best models in the world can be taken away from you at any time they want having the weight downloaded on your mac on your djx park is yours it's you own it it's like literally no matter what happens if that has power you can literally access some of the world's best intelligence and like i shared in my last video which you haven't seen i'll put a link on screen for you these models are getting dangerously powerful.
Like literally, they're probably only a few months behind the frontier level intelligence. Now to do it, first you want to do is head over to Alarma right here and you're going to download this app. You don't have to download the app, but it's quite handy.
If you do download the app, it just basically means you'll get a chat interface like this where you can chat to the models on your computer or download them. But in reality, you can just run it through the terminal. And essentially, Alarma just lets you run models locally and in the cloud.
And you go to Alarma .com, get started, you can download it. Then once you're in Alarma, if you want to, you can literally... come down, click a download button, or if you want, just say to Claude, hey, download for me this model.
Or if you're not sure, all you can do once you've been through this guide here is come down, click on copy, head over to Claude. And then once you're inside Claude, hey there, I want to go ahead and follow these instructions. As simple as that, come down, paste it in, and then literally Claude will download everything for you and you can then run that model locally.
And if you're like me, you're using this agentic operating system that's showing me how much usage I've got on my platforms, and I just want to go local, I just click this one local button here, for example, and then now as you can see, All of the data is running locally on my computer. I can see right now I've got the Gemma 412B running in the background.
It's really cool. And then you see how I've got all these different connections, the things I'm connected to, like Hickfield, Appify, Firecrawl. I can actually turn these connectors off on my own if I want to.
I just click on lock, for example. Now I'm literally just talking to the model and I can see all the different intelligences it's got access to. It's really, really cool.
And from this, obviously, I can go to the chat and ask it questions. Or if I want to play with Claude and Chained, once again, why don't I come down here and I can pick local models if I want to. Here's Quentin 3 .8 local.
I can ask a question. And then literally, guys, that data goes nowhere. That data does not go to a single place.
Now, what is the best local model you can download is one thing. But what the best obliterated model is a different question. Remember, you have local or cloud, and then you have censored.
and uncensored. So if you want local and censored, that's where you get into a really interesting territory. And to find that one out, guys, you just simply come down here to chat with No Refusals, and that will show you the most powerful uncensored model that you can download right now on your computer.
And I just click copy. And then when I talk to this, this will not deny my requests. And on top of that, it will also be running completely locally on my Mac, which means if this internet goes out tonight, I can ask this any question that I want to.
And you won't understand how powerful uncensored AI is until you actually get it up and ask. get a question. On top of that, guys, there are also models that you can run on your computer.
Now, again, it depends. Do I just want the best local model that's just local on there? Or do I want something that's insensitive?
So one such example is something called PocketPound. So in this one here, what we can do is I've downloaded it to show what it looks like. anywhere, anytime.
Your pals live inside your phone. No internet, no signal. They work on planes, off -grid or in remote villages.
There's a few different apps that you can go for that. Again, it's smaller, so it's never going to be as powerful as a cloud model, but it's just something that's on your phone. What does that mean?
It means if you and I are out camping somewhere, maybe we're climbing mountains, maybe we're... having a Guinness somewhere in Wales and we got no connection. Well, we have an AI now.
We can say, hey, somebody that we just saw is just maybe falling off a rock or something. How do we best administer first aid? How do we best solve these problems?
Having a model that doesn't require 5G before, you know, Elon Musk's kind of like a super internet takes over the world is going to be profoundly helpful. This would work on Mars. It would work on Earth.
It would work in Yorkshire, anywhere you go. So think of this like a pocket companion. And the cool thing is nothing ever leaves your phone.
So I've got this here. Again, we can coding educational play. SmartChat is going to be the one to get started with.
And so you can see this particular app, I'm not recommending any app in particular. I'm just showing you an example here. This one here actually goes ahead and recommend something based on the size of that.
So this is the iPhone. It tells me how much space I've got. It lets me have balance, lets me have quick, and lets me have the best.
So you can actually grab one for your phone, which is really cool. So I'm going to go ahead and grab a really quick one here and download this. And right now you can see it's downloading.
How many minutes is it going to take to go ahead and get that done? So guys, that's fully downloaded. It took about...
probably about a minute and a half, 90 seconds. So let's come down and activate model. Now, the cool thing is we can activate uncensored models on this.
We'll show you in a second. But let's come down, select the model, and we've got Pip, our assistant. So this is now running locally, which is really cool.
So why don't I go ahead and actually turn off the internet so you can see exactly what I mean on my phone so you know that it works. Okay, guys, so I've turned off my Wi -Fi and I've turned off my internet connection so I can say, hey, and let's just see what's happening right now. So this is now running completely locally on my phone.
As you can see, nothing on there. Hey, how can I help you today? So I might say something like, hey, who?
is Jack Roberts. And let's see how smart this model is. Does it know who Jack Roberts is and what Videlio is?
And guys, Jack Roberts is a comedian actor. Guys, they are far too kind. I massively appreciate the props.
But effectively, this is now running locally for us on our phone, which is incredible. But guys, running locally is one thing, but what if I wanted an uncensored model? And so guys, to do this, come back over to the guide and click on level five.
And then I want you to come down and toggle on phone. which is cool. And I've got a few different options you can pick.
Again, I'm not affiliated with any of these. I'm just showing you some different examples, okay? Now, we can actually find the ones that run on your phone, which are obliterated, like a Quant 35B, but we can pick how much memory you have.
And based on that information, you can see which model you can actually get downloaded on your phone, which I think this idea of having an AI running your phone is freaking really cool. So to do it, we can just follow this playlist right here, which is dead straightforward. So you can open up the PocketPal AI.
If you have it or not, it doesn't make a difference. And so what we can do, guys, is you click on this button here, like so, and you... see everything, all the options we've got.
You come down and click on models. And then if I want to boot this model off, I want to get rid of it. I can click on offload and that'll be, see you later.
I can come down and I can delete it. See you later, buddy. Thank you for your service.
You were massively helpful. Now come down to available to download here like so. And here we can see all the different models that we can actually go ahead and download onto our own phone.
And as you can see, there's some cool models, but none that are obliterated. So we're going to come down here and click on the plus, and then we click on add from Hugging Face. And then from here, we can actually search for the model we want.
Now, grab the one that you grabbed here, for example, from your phone. So I've got Kwan 13.
So this is one here and this is going to show me the best model. So I just click on a copy like so. And then I'm literally just going to paste that in like so.
and look at this guys we've got all these different ones here these two have vision capabilities and you can see how popular this one's had 444 000 downloads which is really cool and i click on this and then i'm going to scroll down here and look for the file that ends in q4km so this one here is giving me a warning saying that my memory is typed so i'm going to fix that on my phone afterwards but i can come down and click on download like so and now this is downloading for us the obliterated model what does that mean it now means that we can ask it any question that we want to once it's finished downloading and then this brings us on to a really important question by the way if you like channels that tell you exactly how it is i want to show you the world's most powerful technology um feel free to subscribe down below i massively appreciate the support and it helps me on the journey to share with you all this technology as it lands.
Now, having this uncensored model is definitely one thing, but it doesn't answer the question about what can we actually do with the world's most powerful AI, which is the second part of the skill set that you need. I'm going to cover what that looks like inside this video right here.
The Hook

The bait, then the rug-pull.

Jack Roberts opens with a warning and a promise: the world's most powerful uncensored AI, running online, on your computer and on your phone. Then he fixes its number one problem.

Frameworks

Named ideas worth stealing.

00:24list

The 5 Levels of Uncensored AI

  1. Level 1: Abliterated AI in a harness
  2. Level 2: Give it a partner (fact-checker)
  3. Level 3: Run it 24/7 on a VPS
  4. Level 4: Images and video
  5. Level 5: Local and private

A ladder from a hosted uncensored model up to fully private offline inference on your own hardware.

Steal forchoosing how far to go with a local AI setup
CTA Breakdown

How they asked for the click.

VERBAL ASK
26:00subscribe
“Feel free to subscribe down below”

A mid-roll subscribe plea tied to the channel promise of showing the world's most powerful technology.

MENTIONED ON CAMERA
11:40productHostinger ↗
11:42toolTailscale ↗
16:57toolVenice ↗
19:58toolOllama ↗
Storyboard

Visual structure at a glance.

open
hookopen00:00
partner model
valuepartner model05:51
guide page
valueguide page18:00
subscribe CTA
ctasubscribe CTA26:23
Frame Gallery

Visual moments.

One-click upgrade to your Google

Get more breakdowns in your search results

Add Modern Creator as a preferred source and Google shows you more of our breakdowns in Search, Top Stories, and AI Overviews. It only changes what you see, and you can undo it in your Google settings anytime.

Add to Preferred SourcesOpens your Google source preferences with us pre-loaded. Tick the box and you're done.
Watch next

More from this channel + related breakdowns.

14:00
Jack Roberts · Tutorial

Claude Just Solved Motion Design

A fourteen-minute walkthrough of color grading, animated B-roll and film-style transitions, built entirely inside Claude and a personal design dashboard, no editing software touched.

August 26th