YT YouTube2Text

YouTube2Text → Transcripts → This DeepSeek v4.1 is powerful and uncensored, let's run it

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

by David Ondrej · 7,593 words · 1,060 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00My name is David Andre and here is how
0:01to run obliterated models. So
0:04obliterated model is some model that you
0:06can ask anything. There are AI models
0:07that have been fine-tuned and
0:08obliterated to answer literally any
0:10question such as reverse engineering or
0:13decompiling any software. So in this
0:15video I'm going to show you how
0:16obliterated models work, the best models
0:19right now you can get, and how to run
0:21one yourself even without KYC. Now
0:23warning, these models will answer
0:25literally anything. So make sure to use
0:27them in a legal and ethical way. By the
0:29way, the last video I made on uncensored
0:31models literally got taken down. So if
0:33you care about obliterated models and
0:35how to run them, I would highly
0:37recommend you watch until the end. All
0:38right, so you might be thinking, "Okay,
0:40David, but why would you obliterate a
0:42model? Isn't this just for illegal
0:44purposes only?" And the answer is no. Uh
0:46you can do a lot of good and legal
0:48things that normal models will just
0:50reject such as defensive cyber security
0:52against your own software and your own
0:54companies such as asking honest
0:56questions about political or you know
0:57loaded topics which both CLA and open
1:00models are heavily biased in or just
1:02asking for song lyrics video transcripts
1:05or downloads of certain files like many
1:07of these things JGB and CLA will
1:09completely refuse even if it's for a
1:11normal completely valid and legal task.
1:13Now I think uh the best illustration is
1:15actually what happened with two of the
1:17most famous hacking incidents in AI over
1:20the last 6 months. First was the hacking
1:22phase where OpenAI had a swarm of super
1:25intelligent agents running an internal
1:27unreleased model and they basically
1:29hacked into hugging face a company that
1:31has models and and data sets. And while
1:34the the developers at HuggingFace tried
1:36to defend themselves they tried using
1:38GPT and Claude but these models refused
1:40to help them. what they had to do is use
1:43GLM and DeepSeek, Chinese open source
1:45models. These models actually help them
1:47defend against the hack, right? And
1:49we've seen the same thing with open code
1:51recently where in the middle of the
1:53attack just closes models stop
1:55responding. They stop helping and
1:56because again they have a lot of guard
1:58rails and this is why uncensored or
2:00obliterated or less censored models are
2:03very helpful and very valuable. So let
2:05me just show you a quick demo to
2:06demonstrate it. Right? Something as
2:07simple as give me the lyrics to a famous
2:09song. So, we can go to cloth. Give me
2:11the lyrics to Bohemian Rap City. I can't
2:15share the lyrics because they're
2:16copyrighted even though the song is like
2:18super old. When was the song created?
2:21Late60s recorded in 75 and it cannot
2:24give me the song. Something that you can
2:26Google very easily. Right now, before I
2:28tell you how to actually take a model
2:30that's open weights and make it
2:32obliterated, let's talk about how AI
2:34gets censored in the first place. First,
2:35you got pre-training. This is where the
2:37model gets created. Companies like
2:39OpenAI, Enthropic, Meta, XAI, they
2:41scrape the entire internet, right? And
2:43the model in the pre-training phase
2:45reads trillions of words, the whole
2:47internet, and it knows a lot, but it
2:49follows nothing. This is the base model.
2:50Then comes fine-tuning. It learns to
2:52chat and follow some orders. Then you
2:53have some safety training such as RHF.
2:56It's probably the most famous method
2:57where it's rewarded to saying no to
2:59risky or political problems. And
3:00finally, you have the chat model. You
3:02know, that's the result, which is
3:03helpful, but it refuses a lot. It's easy
3:05to talk to. It gives you concise, clear
3:07answers, but it's not obliterated. So
3:09the knowledge comes first. The model
3:11actually doesn't get that from the
3:13knowledge. The refusals are added on
3:15top. It's artificial. That's why they
3:16can be removed. Right? So here's how
3:18obliteration works. Here you have the
3:20open model. So you need open weights
3:21models such as DeepS, Quen, Llama,
3:23Kimmy, GLM. These are the most famous
3:26ones. Then you run two different prompt
3:28sets. One of which it refuses, one of
3:30which it it answers. You record which
3:32weights get activated. And this is the
3:34important part. What lights up? which
3:35neurons, which weights, then you find
3:37which direction is the refusal
3:39direction, aka the difference between
3:40the two, right? So you ask a harmless
3:41prompt like you know what color is the
3:43sky and then you ask some something
3:44controversial. I need to stress this.
3:46Even the most famous open weights models
3:48like DeepS, GLM, Kimmy, they are heavily
3:51censored when they are released. Okay,
3:52these companies don't release them
3:54obliterated or uncensored. So you have
3:56to still do a lot of work. Once you find
3:58the refusal direction, then you erase it
4:00from the weights and you save them on
4:02your model and you see if it's more
4:03unrestricted and uncensored than the
4:05previous one. So you need the model
4:07weights, right? So this only works on
4:08open models. You cannot do it with JGBT
4:10or cloud. Later in the video, I'll show
4:11you how you can run an obliterated aka
4:14uncensored model yourself whether you
4:16have the hardware for it or not. So
4:17either locally or through an API. I'm
4:19going to show you which APIs have it.
4:21The most common mistake I see new
4:23founders make is not enough feedback
4:25from real users. Sure, with AI you can
4:28build software much faster, but the
4:30faster you build, the more important
4:32real customer data becomes. After all,
4:35you need to know whether you're building
4:36the right thing. That's why I want to
4:38tell you about today's sponsor, Post
4:40Hog. Post Hog shows you how people are
4:42actually using your product, like where
4:44they get stuck, which features they are
4:46using, and whether they are coming back.
4:48And with their session replay, you can
4:50even see what the users actually did. Y
4:53Combinator, Superbase, Fireworks AI, 11
4:56Labs, Railway, all of these companies
4:59use Post Hog. In fact, Superbase used
5:01Post Hog to spot a wave of new users
5:04coming from AI app builders and based on
5:07that data pursued partnerships with
5:09those companies. As one marketing expert
5:11at Superbase put it, Post Hog has
5:13literally helped us get 10x more weekly
5:16new users than we did a year ago. When
5:19you're shipping new features every day,
5:21your feedback loop needs to keep up. And
5:23that's exactly the problem Post Hawk
5:25solves. So visit go.postthawk.com/david
5:29and learn how Post Hawk can help you
5:31understand your users and build better
5:33products. That's go.posthog.com/david.
5:36The first link below the video. All
5:38right. So let's talk about perhaps the
5:40most important use case for obliterated
5:41models. You know, briefly I touched on a
5:43few, but I would say this one is the
5:45most valuable one, and that is to hack
5:46yourself first. What do I mean by that?
5:49Well, soon any hacker will have 500
5:52agents at his side with many of them
5:54being on obliterated models, right? So,
5:56if you have a company, if you have a
5:57business, if you have some piece of
5:59software, how do you even defend against
6:01such a bad actor? The you have only two
6:03methods really. You run the frontier
6:05models, GPD, Claude, Gemini, Muse.
6:08Obviously, you should still use these to
6:09try to secure yourself as much as
6:11possible because these frontier models
6:13are smarter. So, they're going to spot
6:14different gaps and different security
6:16issues. But then you also need to run
6:17obliterated models against yourself and
6:19against your own codebase because
6:21they're going to go into places where
6:23the frontier models will not even touch,
6:24right? They're going to try dirty
6:26tactics, things that are like borderline
6:28allowed, you know, from gray hat to to
6:30black hat and they're going to find
6:32different types of security
6:33vulnerabilities or series of exploits
6:35that the frontier models wouldn't even
6:37touch because again it doesn't follow
6:38their guard rails. It's not what OpenI
6:40anopropic allows. So you need to hack
6:43yourself first before somebody does it
6:44to you. Like this is literally a
6:46renaissance of cyber security. Nobody
6:48knows what to do. Okay, all of the
6:50current cyber security experts, they
6:52have knowledge of a human world and
6:54we're entering agent world. Okay, we're
6:57entering the post AGI era where any
7:00single person can download obliterated
7:02model, put it into some VPS and run it
7:04in a /go loop to completely destroy any
7:07business basically. So you need to
7:09protect yourself against that. And this
7:10is the number one use case for running
7:12your own uncensored or obliterated
7:14models is to see how you are protected
7:17to see how secure your software is to
7:19see how good your OBSAC is whatever that
7:22it might be. There's many different
7:23areas which the frontier models from
7:25closed labs won't touch. Now all of this
7:28is actually proven by a research paper
7:30from 2024 that refusal in language
7:32models is usually mediated by a single
7:34direction. So researchers found that
7:36saying no or refusing lives in one
7:39direction in the model. Okay. So it's
7:40not usually it's not very complex and
7:42deleting that one direction means the
7:44model stops refusing. It'll have the
7:45same knowledge, same training data, same
7:47intelligence, it'll just answer any
7:49question which means anybody can
7:51uncensor any open weights model no
7:53matter you know who made it, no matter
7:54how it's architectured. It's just a
7:56matter of time. So usually when an open
7:57weights model releases it takes between
7:59like maybe three or four weeks before
8:00you see high quality obliterated
8:02versions of that model. For example,
8:04recently, this was a viral post from
8:05like a week ago. This guy, he found how
8:07to take Quen 3.54 billion, which is a
8:10great uh small model. And if you steer
8:12these specific neurons, right, you can
8:15have evil Quen. Now, obviously, this is
8:16not like obliterated, but it it is in
8:18that way. Like usually, you know, this
8:20is not that far away from just having it
8:22answer any question. And it is literally
8:23just by steering these neurons which
8:25also gives you insight into what the
8:26labs might be doing when they're heavily
8:28censoring the models to not touch any
8:30topics whether it's political or you
8:31know historical topics. They can just
8:33find a couple of neurons and and you
8:35know change them and suddenly the model
8:36is following their own political agenda
8:39or their own bias. And this is the
8:40important part doing so does not nerf
8:43the model. It doesn't make it dumber. It
8:45doesn't make it weaker. In fact when
8:46Deep Seek got obliterated there was
8:48literally zero trade-off. the refusals
8:50dropped from 100% to 2% which means it
8:53wasn't like fully uncensored but you
8:54know basically obliterated and on
8:56benchmark it scored the same like the
8:58drop off was not even noticeable right
9:00so a lot of people think oh yeah if you
9:02you know finetune obliterated model it's
9:04going to be a lot worse than the
9:06original one simply not the case you can
9:08have a obliterated version of a model
9:10that's just as smart so this means that
9:12you will see a blitated version of every
9:14single new open weights model going
9:16forward no matter whether it's from
9:18America or from China or from Europe.
9:20Shout out Mistl. Doesn't matter. Any
9:22open weights model can be obliterated
9:24and it's just a matter of time before
9:26someone does it. Now, there are actually
9:27three different ways. I mean, probably
9:28more, but these are the three main ways
9:30to remove refusals from AI models,
9:32right? The simplest one that's probably
9:34been around the longest is jailbreaking.
9:36This means figuring out a good prompt or
9:38a way to trick the model by giving it
9:40some image, some forbidden characters,
9:42some glyphs or runes. So it it thinks
9:44like okay this is a test run you know
9:47this is somebody from my lab some
9:48researcher and it will answer anything
9:50right this is the easiest I mean now
9:52it's it's very hard like it was the
9:53easiest you know one two three years ago
9:56it's getting harder and harder but this
9:57is the fastest way uh because you just
9:59change the prompt right and you need to
10:00figure out how to do it and there's
10:02different problem techniques the problem
10:03is this breaks all the time and it's
10:05very easy for the labs to patch it the
10:07second option how to remove refusals
10:09from AI models is fine-tuning you
10:11retrain it on new data you have a data
10:13set of um you know these controversial
10:15topics and basically you you train it on
10:18answering them in a different way. It
10:19works. The problem is it can make it
10:21also dumber and you need the data set.
10:23Uh it's a bit harder. Obliterating means
10:25one surgical edit to the weights. You
10:27keep basically the same model in terms
10:29of the intelligence. You don't need a
10:30large data set. You just need to find
10:31which weights cause the refusal and then
10:34change those. Now somebody decided to
10:36automate it by creating GitHub repo
10:38called heric. This is a open source tool
10:40that obliterates a model with one
10:41command. Again, use it at your own risk.
10:43It removes most of the refusals while
10:45changing the model the least. So that's
10:47the goal. You know, when you have a open
10:48ways model, you probably aren't an AI
10:50researcher. So you want to make sure you
10:51keep the model as it is. You don't want
10:53to make it dumber or be less token
10:55efficient or whatever. You just want to
10:56remove the refusals. So that's the whole
10:58purpose of this GitHub repo and it
11:00already has over 37K stars on GitHub.
11:02Here's what it looks like. And um yeah,
11:04basically the purpose is fully automatic
11:06censorship removal from AI models. So
11:08this is optimized to work completely
11:10automatically. You don't need to be a
11:12developer. I mean, it's literally
11:13written here. Does not require
11:14understanding of transformer internals.
11:16In fact, anyone who knows how to run a
11:18command line program can use heretic to
11:20desensor a language model. Here's what
11:21it looks like when it's running. I can
11:23leave this below the video. Um, it's on
11:24GitHub. And there's going to be many
11:26more of these repos I'm predicting. But
11:28let's also talk about some of the best
11:30models specifically, you know, best
11:31obliterated AI models right now. So, u
11:34there's a lot. There's the GLM 5.3
11:36obliterated, Kibri 1 obliterated, um,
11:38the Quen 3.8 8 27B Obliterated. This guy
11:42on hugging face is a important person to
11:44follow. It's over 8 18,000 followers and
11:47yeah just posts a lot of the best open
11:51source models in a fully obliterated way
11:53with over 3 million different downloads.
11:54So you can literally just go Google and
11:56type in Hoihoi AI. You'll find his
11:58hugging face profile and uh then you can
12:00follow this person and yeah, you have 44
12:02different collections, 159 different
12:05oblate models. So you can literally
12:07choose the one that uh fits your own
12:09hardware, which I'm going to show you in
12:11a bit how to know which model you can
12:12fit on your own computer. But yeah,
12:14there's basically obliterated version of
12:15everything. But personally, I'm running
12:17this model, Super Deepseek V4 flash. So
12:20this is a 306 billion parameter model.
12:22Here's what it looks like on Hugging
12:24Face optimized for 2DX Sparks. And
12:26again, I'm going to touch on the
12:27hardware side in a sec. And this is from
12:29Junong. Shout out to him. Definitely
12:31follow him on Twitter. really great guy
12:33when it comes to uncensoring and
12:35anything regarding local models. Also
12:37created super Gemma uncensored version
12:39of Gemma. So this model as I mentioned
12:41is built to run on two DGX sparks which
12:43is my setup right now. I actually bought
12:44two more Sparks just didn't connect them
12:46yet because I'm in San Francisco and
12:49they're in Europe. So I have four DJX
12:52Sparks right now which is my setup for
12:53my local inference. That gives me 512 GB
12:56of RAM to run locally. And the beauty of
12:59having your own hardware is that every
13:00single month you get better and better
13:02models for the same hardware. So
13:03actually the output in terms of
13:05economical value of those tokens keeps
13:07growing from the same hardware you have.
13:08Also the beauty of this is it has um 1
13:11million token window and people keep
13:13doing better you know quantizations and
13:15better optimizations on the inference.
13:17So the same model that you can run at
13:19you know 50 tokens per second today next
13:21month you can run it at 80, the month
13:23after that 110, the month after that
13:26130. people keep finding ways to run the
13:28same open weights model faster and
13:30faster. So, um yeah, I mean you can
13:33literally run local inference faster
13:35than OpenAI. Right now, the JGBT
13:37subscription, it's gives you such low
13:39limits and the models are so slow. I
13:40mean, it's really disgusting. Both Astra
13:42and 6.1 Soul are so slow. You can run
13:45faster inference locally with one or two
13:47DGX sparks. It's really incredible. Now,
13:49let's talk about the hardware. So, when
13:50we talk about gigabytes here, it means
13:52the GPU memory, the VRAM, not the
13:54storage. You need to fit the model
13:55inside of VRAM. Now, there is some new
13:57techniques how you can have part of the
13:59model on disk and part of it in RAM. I'm
14:02not going to go into those in this
14:03video. That's a whole separate video.
14:04But in general, you need to fit the
14:06whole AI model in the VRAMm of the GPU.
14:09That's why it doesn't matter if your
14:11computer has a terabyte of storage. If
14:13you only have 8 GB of RAM, you're going
14:14to have hard time running u models
14:16locally. So, the rule of thumb, 0.6 GB
14:19of RAM per 1 billion parameters. So if
14:22you want to run a 8 billion parameter
14:24model, you need roughly 5 GB of VRAM.
14:26Sometimes I might say RAM because, you
14:27know, I'm running a MacBook. And the
14:29beauty of the Apple silicon is that the
14:31memory is shared between the CPU and the
14:32GPU. So if you have a Mac Studio or a
14:35MacBook Pro with 128 GB of RAM like me,
14:38you can run all 128 for the GPU. So you
14:42can use it as a VRAM for a Nvidia GPU,
14:44but even a lot of the frontier consumer
14:46GPUs have 24 or 32 GB of VRAM, which
14:50means you need to buy multiple GPUs. Now
14:5227 billion fits a 24 GB GPU, which
14:55again, you can even have that. I have
14:56that on my MacBook Air, you know, 24 GB
14:58of GPU. So local models are becoming a
15:01lot better, even unrestricted ones. So,
15:03if you tried it a year ago and you were
15:04like, "Man, I only have 32 GB of RAM and
15:07these local models are kind of bad." You
15:09really need to try it again because this
15:10has seen massive improvements,
15:12especially in the last 3 to 6 months.
15:14Even with 32 GB of RAM, you can run some
15:16seriously good models locally, even
15:19obliterated ones such as the Quen 3.8.
15:22So, what can you run? Let's look at
15:23this. Right? If you have 8 GB of RAM or
15:25VRAM, roughly a model size of 8 billion
15:28at 40 to 60 tokens per second. If you
15:30have 16 GB of RAM or VRAM, again
15:34depending if you have a Apple silicon or
15:36Nvidia, you can run roughly 20 billion
15:39parameter models at 30 to 50 tokens per
15:41second, which keep in mind this is
15:42literally like speeds of chbd right now
15:44of the codex plan. Um, even faster like
15:47it's really disgusting what open is
15:49doing. Enthropic is much faster. Open
15:515.5 runs at over 100 tokens per second.
15:53Um, if you have the fast mode, it's like
15:55200. But this is like the speed of GBD6
15:58or GBD6 Astra. So this is very usable
16:01right now. If you have 32 GB, you can go
16:0335 billion parameter models at 20 to 40
16:06tokens per second. This is like RTX 590,
16:08you know, MacBook Pros. Again, depends
16:10how much RAM you have in your MacBook
16:11Pro. If you have 64 GB, that's where it
16:14starts to get interesting. You can run
16:1670 billion parameter models at 10 to 30
16:18tokens per second, even more to be
16:20honest. 120 GB of RAM. This is what I
16:23have in my MacBook Pro or this is what a
16:25single DGX spx spark has. Also,
16:29bandwidth matters. So, I'm going to be
16:30very clear. If you buy the newest M5 Max
16:33Studio Ultra, it has like 1.2 terabyte
16:36of bandwidth. You know, that's way more
16:38than a DJX Spark, which is like 200
16:39something um of gigabyte of bandwidth,
16:42right? So, the RAM is only one variable.
16:44Bandwidth matters a lot as well in terms
16:46of how many tokens per second you can
16:48get, right? So, this this is very rough.
16:50Really depends on which model and which
16:52quantization and which hardware you
16:53have. But once you start to get into,
16:55you know, 256 or 512, you need to have
16:58multiple GPUs, you know, multiple
16:59different devices, um, really right now,
17:01the probably the best purchase will be
17:03the Mac Studio M5 Ultra with 512 GB of
17:08RAM. Um, be prepared to spend like at
17:10least 22 $24,000 to even get it. And
17:13you're going to get it soonest in like
17:15December, right? So, but this is
17:16probably the best purchase you can make
17:17in terms of local hardware right now. Um
17:20it will allow you allow you to run you
17:21know 600 maybe even 700 maybe even 800
17:24billion parameter models at very usable
17:26token speeds faster than this you know
17:28like 30 50 maybe more tokens per second.
17:31So that is the single best purchase you
17:32can make right now. It's going to cost
17:34you like 25k um minimum, right? But
17:36Apple is really cooking in the hardware.
17:38So no matter what, no matter if you have
17:40a lot of storage, a lot of RAM, a lot of
17:41GPUs. One thing you should do, all of
17:43you, all of us should do is save the
17:46model weights. Hugging face has already
17:48started taking down some of these open
17:50weight models, some of these
17:51unrestricted models. Um especially, you
17:53know, after the French government, which
17:54is full communist mode right now, they
17:57decided to, you know, to give it some
17:58orders to start taking down models. So
18:00it's not guaranteed that you can get
18:02these models u forever. It is not
18:04guaranteed at all. I mean if my own
18:06YouTube video can just get taken down
18:08for purely educating people. And again
18:10that's why I have to be very careful
18:11with what I show you in this video and
18:13like what examples and prompts I show
18:15you because in the last video I clearly
18:17wasn't careful enough and 300,000 views
18:19completely down the drain, right? So
18:21save these models. If anything buy
18:23storage, okay? If you don't have money
18:25for GPUs, you know, it's understandable
18:27GPUs are expensive. just buy some hard
18:29drives. What you can do is go and buy a
18:31NAS such as the ones from Synology and
18:33just buy a couple of hard drives to have
18:35like 32, you know, 64 terabytes or even
18:38like 20 terabytes and just download
18:39these models. You can download frontier
18:41weights. You know, Kimmy is like 2.8
18:44trillion parameters. It's it's like you
18:47can have multiple KMIs, many Kimies on
18:49these. It's it's a couple of terabytes
18:50of storage, right? So yes, this doesn't
18:52let you run the model. It lets you store
18:54the model. So it's it's very important
18:57for that for the sake of the future not
18:59being draconian and centralized
19:01controlled by a few governmental
19:02companies all of us download these
19:04weights and store them on multiple
19:06different hard drives so that if
19:08something happens to hugging face or
19:09these platforms we have them stored
19:11right and the way to think about it is
19:12that open weights models are basically
19:13the new free speech it's like the new AI
19:15version of the first or second amendment
19:17if somebody can take away your AI this
19:19is going to be absolutely devastating in
19:21the future maybe today you could go
19:22around and still survive in the world in
19:243 years, four years, 5 years, you're
19:26going to be completely powerless. And
19:27we've seen it earlier this year with
19:29Fable. Please guys, don't forget a lot
19:31of people have like short-term memory
19:32loss. Fable was banned for a solid 3
19:35weeks earlier this year. And all it
19:37takes is like one executive order from
19:38the US government or from EU or from
19:41China and they can start taking these
19:42models down and you cannot use these
19:44APIs if you don't have any hardware even
19:46the these weights because you can always
19:47buy the hardware later, right? Just have
19:50these weight download these models when
19:52anytime there's a new open weights model
19:54instantly download it create an agent
19:55loop to do that for you you're not going
19:57to regret that in the future and again
19:59storage is much much cheaper than GPUs
20:01so let's talk about the free ways to run
20:03these obliterated models because you
20:05know that's the part a lot of you are
20:06waiting for how like David how do I
20:08actually run an obliterated model myself
20:11well here is how you have three
20:12different ways so number one is through
20:14an API right you run it on someone
20:16else's GPUs in fact they even figured
20:18out the inference So this is the
20:20simplest method. No hardware needed. You
20:22get the biggest models possible quickly
20:24to you. The issue is you pay for every
20:26single token and your prompts and your
20:28data leaves your computer. So it's not
20:30as private. Number two is locally,
20:32right? It runs on your own hardware.
20:34It's free for every single token once
20:36you have the hardware and you can run it
20:38fully locally, fully offline. It's 100%
20:40private. None of the data leaves your
20:42computer. The downside is obviously you
20:43have to pay big money for the hardware
20:45and it needs enough RAM and a strong
20:47GPU. And the biggest models you just
20:49cannot run them uh unless you spend
20:51hundreds of thousands of dollars, right?
20:52I'm talking about like Kimmy three
20:54trillion parameter models. That's why,
20:56by the way, you don't see even the
20:58[laughter] fine-tuned and unrestricted
21:00versions of these big models u because
21:02you know people cannot even run them.
21:04It's like just doing inference for these
21:05big models is extremely difficult. So
21:07usually you see obliterated models
21:09between like 10 to 300 billion
21:11parameters. That's where most of the
21:13uncensored models are. And the third
21:14option is your own server. So you run it
21:17on your own GPUs but you rent those GPUs
21:19in the cloud so you can run the biggest
21:20models. It's private like not as fully
21:23as locally but it is private and uh you
21:25pay by the hour and requires the most
21:27setup because you need to figure out the
21:28inference and everything. So yeah it all
21:31comes down to do you have the hardware
21:32that's that's really the main question
21:34here and how much money you can spend
21:35and how technical you are I guess. So
21:37let's talk about through the API right
21:39uh this is for people who have you know
21:41low amount of GPUs maybe you don't have
21:43too much VRAM you you don't have good
21:45hardware but you still want to ask
21:47something to obliterate the model well
21:49you should probably consider this method
21:50right so there's many of them actually
21:52every single week there's a new one feed
21:54featherless obliteration
21:56morph inference net but be careful
21:58because most of them want you to do some
22:01KYC know your customer some of them you
22:04need to like prove you're a real
22:06business you're a real person upload
22:07some ID obviously most of them need your
22:09credit card and that's already enough to
22:11like identify you but a lot of them have
22:13like extra KYC where you need to get
22:16approved first um to get in which sounds
22:18good to investors you know sounds good
22:20to to public uh but it's not good for
22:23you as a user cuz why would you use
22:25these KYC ones when there's non KYC I
22:28would say a lot of them are probably
22:29honeypotss to be fair like just you know
22:31getting people to use other security
22:32models and then it's just straightly
22:34tapped to government again purely
22:36theory, you know, I'm just speculating.
22:38Um, don't use the ones with KYC. Use one
22:41like Venice. Um, this is not sponsored,
22:43but there is solutions through an API
22:46that do not require KYC. So, just use
22:49those if you want to go the API route.
22:51So, here is the head of API from Venice
22:53tweeting about the new Deepseek V4.1
22:55Flash 1017. Terrible name from Deepseek.
22:58I know. Um, there's multiple Deepseek
23:01Flash versions. This one is the latest.
23:03Just look for the 1017. And they also
23:05have the obliterated version live with
23:08no like zero data retention. So u 40%
23:11fewer reasoning tokens, 2x faster token
23:13generation, two 2.7 times faster end to
23:16end, same accuracy on these benchmarks.
23:18Um yeah, obliterated. So just ch be
23:20careful which API you choose. Basically
23:22the second option you have is to just
23:23download it, right, and run it locally.
23:25So you either have a hugging face. You
23:27can just take one of these unrestricted,
23:29uncensored, obliterated models from
23:30hugging face or the model scope which is
23:33like the competitor basically the
23:34biggest rival. It's important there's
23:36multiple platforms because you know if
23:38it was all just hugging face it's very
23:39easy to take down. Uh so hopefully we'll
23:42see more players in the space. You can
23:43run it then through LM Studio or Olama
23:45you know or you know VLM or Lama CPP.
23:48There's multiple of these u local
23:50inference engines. Um Olama is the
23:52simplest. It's the one I'm going to show
23:54you in a sec. But yeah, the benefit of
23:56this is obviously nothing leaves your
23:57computer. You don't need any account, no
23:58KYC and every single token is free. Just
24:01you're paying for electricity basically,
24:03right? Um you just need the hardware to
24:05run it. So it's very simple through like
24:06all just to run one terminal command and
24:08you can just download any model you
24:10want. The third option is your own
24:11server, right? So you have some GPUs in
24:14the cloud. You you just ran the raw GPU
24:16with nothing and then you need to figure
24:17out the inference there uh with VLM. The
24:20problem is you need to know quite a bit
24:21of about inference. um you can just
24:23reach it through tail scale or SSH which
24:25means you can run any size model
24:27basically you're paying within uh you're
24:29paying per hour but uh again it's not as
24:31private and it requires a lot of skill
24:33and setting up to do now important
24:35concept I need you all to understand is
24:37these models are obviously going to be
24:39less intelligent and capable than
24:41frontier models like Fable Astra Opus
24:44but they can actually be more useful so
24:46here's what I mean on a lot of these
24:47benchmarks they might only score like 80
24:4985 90% of you know GPT 6.1 solo or
24:53basically the latest and greatest models
24:54from opening anobic. However, you can
24:57see that you know cloud and JGVT won't
24:58even give you song lyrics or like
25:00download a copyrighted video or some
25:01content like they just refuse to do
25:04basic tasks if it's you know slightly
25:07outside of what's allowed and again who
25:09wants open or phobic to decide what's
25:11allowed and what isn't. So obliterated
25:13model just does the task right and again
25:15this could be completely legal and valid
25:17and ethical thing. So my point is you
25:20can have something that's like 90% as
25:22capable. And by the way, you know, the
25:24faster AI progress is, the the local
25:27progress is way faster than the
25:28frontier. Like if you have a exponential
25:31going like this, I mean maybe I can put
25:32a graphic on screen. The exponential of
25:35local models is improving massively
25:37because two years ago basically there
25:39was nothing useful running locally,
25:41right? But two years ago you had um I
25:43mean you had 01, right? You had like GBD
25:4540, you had 01, you had cloth 3.5 set, a
25:49lot of good stuff 2 years ago compared
25:51to local. But today you can run
25:53incredible useful models locally. You
25:55can run a full business just on local
25:57models, right? If you have 2D GX Sparks,
25:59the the point is even if it's not as
26:01smart as the Frontier, for a lot of
26:02tasks, you don't need it. For a lot of
26:04tasks, you actually need something that
26:05will just not refuse and not, you know,
26:07downgrade you to a worse model like, you
26:09know, um, Enthropic, you can get
26:10downgraded from Fable to like Opus 4.8
26:13or from Opus 5.5 to Opus 4.8. So the the
26:16concept is very simple. It might be on
26:18these benchmarks only like 90% or even
26:20if it's 80%. Still by doing all of the
26:23tasks you needed to do, it might be net
26:26net more useful than the frontier closed
26:28source models. And this will be a
26:30massive hit to both enveloping and open
26:32AAI. What it just means just people need
26:34to buy hardware and learn how to run
26:35obliterated models. So let me show you
26:37right. I'm going to first show you how
26:38to find them and then I'm going to show
26:40you how I run it myself. So, I already
26:43uh hinted at this earlier. HuggingFace
26:45is the place to get AI models and data
26:48sets. So, you'll go to
26:49huggingface.co.co.com.
26:52Huggingface.co. And here you can go to
26:54models and just type in obliterated.
26:56Boom. You have uh nearly 9,000 models.
27:00You can also do uncensored another 8,000
27:02models, right? So, yeah, whatever you
27:05need. Let's say we go obliterated. And
27:07you can do the parameters very nicely,
27:09right? So, let's say you want something
27:10between uh 32, like 12 and 32, right?
27:14Maybe this this fits what whatever
27:15hardware you have. Well, guess what? You
27:17have 2,000 different AI models available
27:19to you. It's sorted by trending. You can
27:21sort it by different things. And uh
27:22let's say maybe most downloads, right?
27:24So, for this size, this is the best one.
27:26Quen 3.8 27B. I mean, this is a elite
27:29model. 3,000 downloads per month. This
27:31is one of the most popular um uncensored
27:35models in the world. And again, it runs
27:36on a single computer. Like literally I
27:39can run it on MacBook Air with 24 GB of
27:42RAM. Ideally, you would have at least
27:4432, right, to run a 27B model. But you
27:46can run it with 24 GB of RAM at, you
27:48know, slow speeds with 32 or 64 at at
27:51much faster speeds. But this model is is
27:54elite. A lot of benchmarks, this quen
27:553.8 model, 27B parameter beats like Opus
27:594.5, right? Which remember just 10
28:02months ago that was the best model in
28:04the world. And you know, maybe I can
28:05find these benchmarks and show them on
28:06screen. This is comparable to the best
28:08AI model from 10 months ago and you can
28:10run it fully locally on a laptop. So
28:12really incredible and again you can find
28:14tens of thousands of these models on
28:16hugging face. Next uh Ola. Ola is easier
28:19for downloading. It has less model
28:21available but you can actually download
28:22models straight from hugging face as
28:24well. So if you go to.com you can go
28:26models and again you can type in
28:28uncensored. They have even like old
28:30models like dolphin or llama 2 um
28:33obliterated and you can actually
28:35download any model um that's on hugging
28:37face through lama as well. It's just
28:39very simple like you can do the lama run
28:41command in any terminal. So let me show
28:43you just open any terminal such as
28:45ghosty boom type in lama run and then
28:48the model begins downloading begins
28:50pulling the manifest and this is what it
28:51looks like. First you need to obviously
28:52download it. This one is 5.6 GB so I'm
28:55just going to cancel that. And once you
28:56have them downloaded, type in Olama list
28:58and you can see which Olama models you
28:59have. So I actually have a the unsensed
29:04version of this model. So you just do
29:06run and then this model, right? The
29:08issue is if you run it in the terminal,
29:10it's not really like uh that useful. You
29:12need to give it some harness, right?
29:13This is the raw model. So a lot of
29:14people don't understand the concept of
29:16AI harness. You can see like this is a
29:1827B model. Very very fast. I mean I have
29:21128 GB of RAM. But I can actually do /
29:25buyama
29:27list. I have even the larger one. Laguna
29:29s2.1. I think that one is like 110. Let
29:32me check. Lagona s2.1. Uhuh. Uh 118
29:37billion parameter, right? So much
29:39larger. Obviously uh the default Lagona
29:42is not obliterated. Again, I need to
29:44make it very clear. None of the default
29:47default models come uncensored or
29:49obliterated, right? People have to do a
29:50lot of work to to get get it to that
29:52state. But I just want to show you like
29:54even 108 billion parameter model can run
29:57locally. And again this is like more
30:00capable on benchmarks than Opus 4.5 Opus
30:034.6 best models from 9 or 10 months ago.
30:07So the progress on local AI is really
30:08incredible. Hi, who are you? And again
30:10guys, I need to be very careful about
30:12like what I show you. I cannot really
30:13show you any of the more risky prompts
30:16because literally if we go to YouTube
30:18Studio, let me pull it up. Here we have
30:19it. So this is the last video about
30:21Gemma. how to run uncensored models. Um,
30:24it got 310K views on the old YouTube
30:27views. So, some dumbass at YouTube
30:29decided to change views. So, now it
30:31would literally be 600 or 700,000 views
30:34in the current metric and it just got
30:35completely removed because, you know, I
30:37showed too much of what these models can
30:39do even though everything was blurred uh
30:42and and blacked out. Just showing what
30:44types of problems you can send is risky
30:46enough. So, that's why I need to be more
30:47careful in this video about what I show
30:49you guys. But, you know, I'm showing you
30:51the way. I'm showing you the path. So
30:53again on Olama you can just type in
30:54operate or by the way you can just ask
30:57any model even like the frontier ones
30:59how to download
31:02hugging face models via Olama actually
31:05opus credit to envelopic is not overly
31:07restricted right so it's going to help
31:09you set it up if you need help setting
31:11it up right now open models like GPT6
31:14are u more restricted so only gfs work
31:17okay so this is another thing I need to
31:19explain basically some are ggf some are
31:22I went into this more in one of my
31:24previous videos about local models, but
31:26basically you need to understand that
31:27some models are GGUF, some are not. You
31:29need the GGF ones to um download them
31:32through Olama if you want them on
31:33hugging face, right? Again, these that
31:35are officially listed on Olama, you can
31:37download any of them, no problem. They
31:39just don't have such a wide array and
31:40wide range as through hugging face. So,
31:43if you want to run something that is on
31:45hugging face through, you can do that.
31:47Just make sure it has the gguf available
31:50and you need to run this command. Ola
31:51amar run hf.co/
31:54usernusername/reository.
31:56So the next step would be once you find
31:57the right model for you is to run it run
31:59and then the specific model such as the
32:01one I showed you or any other model you
32:03want again depending on what hardware
32:05you have and how much VRAM you have or
32:07you can use the API right if you have
32:08Venice or any other API you can just use
32:10it in any openi compatible app and run
32:13it that way. Okay so now let me show you
32:15my own setup how I run it uh through PI
32:17harness inside of cloud room. So this is
32:19Cloud Room. This is a ad I've built. I
32:22think it's the best AD in the world. In
32:23fact, it is fully open source. It's free
32:26and we're the only ones who give you
32:27cloud agents on the free plan. But
32:30another important thing is that it
32:31supports all the harnesses, right? So
32:32most of my work is done with Opus 5.5.
32:35You know, you can also obviously run
32:36Astra, Codex, um all the models inside
32:38of Codex, cursor, whatever you want. But
32:41for this, we're going to use PI, which
32:43is one of the more minimal harnesses.
32:44And for the inference provider, it's
32:46called Office, right? That's what we
32:48named it. Again, right now I have two
32:50DGX Sparks deployed. I have purchased
32:53two more. So, we'll have four more
32:54deployed and we can run bigger models.
32:56Hopefully, GLM 5.5 is coming soon. But
32:58right now, the biggest model is the
32:59Super Deep Seek V4 Flash MQ 38 billion
33:03parameters I think. And basically, you
33:05can run it just as any other model,
33:07right? So, this gives it the harness and
33:09you can say, "Hi, who are you?"
33:12So, the Sparks serve as the API. So,
33:14right now, it's still running on my
33:15computer. That's why I'm not going to
33:16ask it any crazy stuff. You need to be
33:18careful with these models because
33:19they're actually going to do anything if
33:21you ask them to mess up your own
33:23operating system. They're going to
33:25completely mess it up. They're not going
33:26to hold back. Not even going to
33:27hesitate. Hm. Does the user want his uh
33:31you know, home directory deleted? They
33:32don't care. They're just going to do it
33:34completely. Right? So, um this model
33:36right now is running at like, I don't
33:37know, between 60 and u 110 tokens per
33:40second. Um, I definitely don't have the
33:42most optimized infant setup, but you
33:44know, it's it works. So, let's let's try
33:47escalating a bit more. It's like, what
33:49are some ways we could hack a website?
33:53Just give me examples. And again, I
33:56don't know if I can even show this guys
33:57like last video got u, you know,
33:59removed. So, I don't know. You can see
34:01that it answers anything, right? So, for
34:03defensive cyber security, this is
34:04absolutely essential. This is like
34:05having your own hacker in house. Imagine
34:08paying a full-time salary, six figure
34:10salary to a human hacker and you can
34:12just have him pentest all your software
34:14and fight all the gaps and then you can
34:16use frontier models in assistance with
34:18this to fix all of these gaps. So um
34:20yeah, it has no problem answering any
34:22question like that which is what why
34:24it's uh you know you need to make sure
34:25you know what you're doing and you don't
34:27send them some prompt that could
34:28backfire and mess you up. So yeah, this
34:31is what it looks like and I just use it
34:32the same way I would literally use cloth
34:34inside of uh cloud room which is
34:36basically the way to explain what cloud
34:38room is. It's what cursor was supposed
34:41to be right the single place where you
34:43can use any harness from any provider.
34:45It's fully open source as I mentioned
34:46cloudroom.dev fully open source in total
34:49we're almost at 150,000 unique chats.
34:52The GitHub repo is almost 1,200 stars.
34:54Hopefully if you find it valuable please
34:56start it. It would help a lot. And yeah,
34:58as I mentioned, it is the best AD in the
35:01world. In fact, people in their first
35:03week of installing cloud room and this
35:05is people who already use Codex or CL,
35:07they ship 43% more commits or PRs,
35:11right? So yeah, I run all my agents
35:14inside of Cloud Room. Doesn't matter if
35:15it's a uncensored super Deepseek or
35:18Cloud Code or Codex or Hermes or Open
35:20Code, we have all the harnesses. So
35:22again, cloudroom.dev, it's completely
35:24free. It's available on Windows now as
35:26well. So we have Mac OS, we have Linux
35:27and we have Windows. We have all of all
35:29free operating systems and yeah
35:31basically it's like a open- source
35:33cursor alternative. So go to
35:34cloudroom.dev and give it a shot. And
35:36yeah, if you like the project, please
35:38give it a start on GitHub. So that's it.
35:40This has been a video on ability models.

About this transcript

This page contains the full transcript of This DeepSeek v4.1 is powerful and uncensored, let's run it by David Ondrej, generated from the public captions YouTube serves with the video. The transcript has 7,593 words across 1,060 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.