All Episodes
Previous Next
April 5, 2024 - Health Ranger - Mike Adams
27:09
NEO LLM guide - How to get started using our Large Language Models
| Copy link to current segment

Time Text
Welcome, everybody.
This is Mike Adams, the founder of Brightown.com and Brightown.ai.
And also, I'm the executive director of the Consumer Wellness Center Data Science Nonprofit.
And that's my dog, Rhodey, on the floor there, if you guys can show that shot.
There he is.
He's sleeping because we were having fun with him.
And in this segment, I'm going to give you some information about how to download and use the large language model files, the AI projects that we've been building, that we're releasing through our nonprofit, Consumer Wellness Center, the Data Science Division.
These are experimental, non-commercial, nonprofit language models.
They come with some disclaimers and certain licenses, like Apache 2 or MIT licenses, which allows you to use them and also to build on top of them.
Now, this gives you decentralized knowledge in your hands.
These do not run in the cloud.
You download these models.
Sometimes they're three gigabytes or four or seven gigabytes.
You download them, and then you run them locally on your own computer.
And then even without an internet connection, you can chat with these language models and ask them questions.
We have trained them on literally millions of pages of documents.
It includes books.
It includes 365 interviews that I've conducted here.
Hundreds of thousands of articles, many podcasts and transcripts and websites and research documents and so much more.
So we have trained these models on a tremendous amount of information in order to help them give more accurate responses, more real-world responses, and to try to program the biases out of these base models.
So let me explain how this works.
There are base models that are produced by companies like, let me show you this one, Mistral.ai, I believe, is the company.
Mistral, this is out of France.
Here's their website.
They release base models called Mistral, such as Mistral 7B, which is a 7 billion parameter model, or BioMistral.
And these base models require often tens of millions of dollars of compute effort in order to generate.
Companies like Mistral, Microsoft, Meta, in some cases Google or other companies, they will release these base models for free to the open source community as experimental models.
Then, organizations like Brighteon, we take those base models, which have a lot of biases in them.
All those base models believe in the cult of climate change.
They all believe that men can become women.
They all believe that big pharma is wonderful and that all vaccines are safe and effective and all that nonsense.
What we do is we retrain them on reality.
So we'll train them on documents that show how people are harmed and killed by vaccines.
Or we train them on documents that talk about nutrition in food and how nutrients are anti-cancer in terms of their effects on the human body.
Or we train them with documents that tell the truth about What happened on January 6th, for example, or how the Federal Reserve works, or what is money versus currency, and so on.
Now, our focus has been nutrition, herbs, food production, survival skills, off-grid skills.
But as we continue to train and release more models throughout this year, we will expand into more and more concepts like economics, history, philosophy, politics, geopolitics, and so much more.
But the thing you need to know is that these are all experimental.
There's no guarantee that they're going to give you the right answer to any particular question.
Every model that exists can give you bad information if you prompt it to do so.
And if you maliciously prompt these models, if you want it to give you crazy information, it will do that.
And that's true with Microsoft.
It's true with Google.
It's true with ChatGPT, OpenAI, whatever.
So anybody looking to abuse our model, go for it.
You can abuse it on your own laptop.
You do so at your own risk.
And because these are experimental, non-commercial, non-profit models, there's no warranty of what kind of information it's going to give you.
It's just that we are working with best efforts.
We are good faith actors working to build models that can give you knowledge and answers that will help you navigate the world.
We are doing this kind of as an arc of human knowledge.
It's kind of like a time capsule of human knowledge because there's so much censorship in the world.
A lot of things that people need to know about nutrition and foods or vaccines and spike proteins or detoxing, a lot of that is censored by big tech, and we're able to bring it back through these large language models, such as the ones we're about to release.
So if you go to our website, brighteon.ai, You're going to see, soon, you're going to see some additional links on the right-hand side, over here, that will show you where to download models.
Now, all of our models are called NEO. And we'll have multiple NEO models that are built on top of the base models.
And the base models that we're training on include Mistral 7B. And 7B means 7 billion parameters.
So we will have NEO Mistral 7B. And we're also training on an uncensored version of Mistral called Dolphin, created by Eric Hartford.
So we'll have Neo-Dolphin Mistral 7b will be a model that we'll release.
We will have versions of these models, version 1, version 2, and so on, based on our data set getting larger and larger as we prepare more data for the fine-tuning training.
We're also going to be releasing models based on Microsoft's PHI-2, that's P-H-I, PHI-2.
So we'll have Neo PHI-2, version 1, version 2, and so on.
We're also releasing models based on BioMeestrel, which is the Meestrel company releasing a model, 7 billion parameter model, based on training of all the PubMed published scientific literature.
So we are fine-tuning on top of that to give it extra knowledge about health and nutrition from natural news articles, from other scientific articles, from books and authoritative sources, interviews with nutritionists and naturopaths and so on.
So we're going to have a Neobiomistrel 7B version 1 coming out very soon.
You will be able to freely download all these models.
And if you want to build on top of these models, we are also releasing the parameters files on huggingface.co.
Let me bring that up, actually.
Huggingface.co is the website where, here it is, the AI community building the future.
This is where you can download a lot of models and a lot of data sets.
We will be posting our parameters files to HuggingFace.
So those of you who know about HuggingFace and you know about LLMs, you'll be able to find them there.
They will be released under CWC Data Science, and the model names will be Neo.
All right, now back to Brighteon.ai.
When you download these files, they will be in the format known as GGUF. This is the modern format for large language models.
You need to run a piece of local software in order to run the language model or to run the gguf file.
Really what this local software is doing is called inference.
It's loading up the model.
It's giving you a prompt.
You ask it a question by typing it in, like, you know, hey, what is geoengineering?
And then, through inference, it will generate the answer for you.
And there are a few pieces of software I recommend for this.
One of them is called LMStudio.
And here's the website, lmstudio.ai.
And you can download this.
Here it is.
You know, Mac, Windows, Linux.
And this is an inference software.
And with this, you'll be able to run our language models on this software.
And how powerful of a computer do you need?
Well, it depends on which model you're trying to run.
PHY2 is a smaller model.
I think it's 1.7 billion parameters.
So you'll be able to run PHY2 with a lot less RAM on your system.
And if you go to our website right here, brighteon.ai, there are tutorials right here.
Click on Tutorials, and it shows you how to set up LM Studio.
We're also going to show you how to set up GPT for All, which is another piece of software that does much the same thing.
Anyway, back to LM Studio.
By following our instructions, you do have to put the gguf files in the correct folders.
You will be able to run inference on our language models.
So Phi 2 is maybe 3 gigabytes.
I think Biomistrel will be something like 5 gigabytes, maybe up to 7 or 8 gigs.
And the Neo Dolphin Mistrel will also be in that range, 5 to 7 gigs.
We may have different versions available.
We may also release an executable version based on...
A compilation protocol known as LAMA file that makes a single executable file that does inference for you using your local browser.
However, there's a four gigabyte limit on that in Windows, so that forces the models to be much smaller and not quite as useful as the full-size files.
Anyway, we'll have links on brighttown.ai for you to do that.
It's very important that when you use something like LMStudio, Let me see if I can zoom in here.
Show my screen.
You see this on the right-hand column right here?
There's a blue bar there.
And that's where you choose which model base that you're going to use.
And you need to choose the right base model there.
So if you're downloading Neo Meestral, then in this box you need to choose Meestral Instruct.
If you're downloading Neo Phi 2 in this box, you need to choose Phi 2.
These are called presets.
If you don't choose the right preset, you will get garbage.
It will just spit out just totally random garbage, by the way.
So you got to get the right preset in there, and then things will work a lot better.
Now, there are some other things that you need to know.
Some of these base models can naturally write code.
For example, the Mistral models can write Python code.
And if you ask it to write code, it will do so.
Some of them are multilingual.
And you'll find out if you want to ask it questions in Espanol or in German or French or Italian or whatever, you'll see that a lot of these models have base training in those different languages.
However, we have not trained in any languages other than English.
And so if you ask it a question in Espanol, it's going to give you an answer in Spanish that we did not train it on.
It's probably going to be more of a biased globalist type of answer until we get a chance to retrain them in Espanol, which is one of our goals that we're going to do this year, by the way.
In addition, Understand that, again, every base model has biases.
Every base model was trained on Wikipedia.
Wikipedia is not a reliable source of information.
Wikipedia is run by the CIA. And Wikipedia is used to smear certain individuals, like Donald Trump, for example, or anybody associated with Trump, or anybody in alternative medicine is smeared in Wikipedia.
And so that bias continues in the base model.
And the only way for us to remove that is with new training.
Just like we're training it with new real-world knowledge on vaccines, on spike protein, on geoengineering, on climate change, on transgenderism, on herbs, whatever.
We also have to train it on all these other areas in order to affect its parameters.
So when you start using these models, and if you see it give an answer that sounds like totally wacky, like let's say you ask it, hey, why are the globalists trying to achieve depopulation of the human race on planet Earth?
If it gives you an answer like, oh, depopulation, that's just a wild conspiracy theory.
There's no credible evidence that that exists.
Well, that's not something that we trained it to say.
That's something that Microsoft trained it to say.
And we just haven't yet been able to override the Microsoft brainwashing.
Actually, on our models, if you ask about depopulation, it does tend to give you very good answers right now, but that was just an example.
There are many areas where if you ask it a question, it's going to give you biased answers that are kind of bleeding through the original base model that came from Meta or OpenAI or Google or Microsoft or what have you.
Over time, however, our training will tend to override that.
So as we release more and more models, which we're going to do this entire year, they will all be free, they will all be non-commercial, non-profit models, they will continue to expand their knowledge base, and you'll see them giving better and better answers over time.
Right now, most of the answers on subjects that matter are pretty good.
Especially on the PHY2 model, which is easier to train because it's a smaller model.
The larger the model gets in terms of parameters, the more difficult it is to train.
It takes more epochs of training, and it takes a larger data set, and it takes the rephrasing of the data sets in order to achieve that.
And all of this takes time.
So what we have actually built since about Thanksgiving of last year is we have built a massive data pipeline system that processes incoming data, whether it's books or audio files, interviews, videos, articles, what have you, It processes that.
It does the transcriptions.
It does the rewording.
It segments content.
It builds question-answer pairs.
It does everything that's necessary.
It normalizes content in order to put it into a training format to train these language models.
And as a result, our data sets are getting larger and larger, which means that throughout this year, the language models that we release will have wider coverage of more areas of human knowledge.
And ultimately, we will use these language models to translate everything into Spanish and German and French and even Mandarin Chinese and who knows eventually how many other languages we will get to.
But we have a very good R&D team.
We have our own hardware.
We have our own servers.
We own hundreds of thousands of dollars of servers and NVIDIA GPUs and A40 cards and whatever.
We've spent a lot of money with NVIDIA in order to do this.
But as a result, we can't be deplatformed.
Nobody can take this away from us and deplatform us and say, oh, you can't use our servers for training your language models.
And that's why we built this infrastructure ourselves, so that we can do this without being deplatformed, and we can get you these tools and get them into your hands and continue to improve them all year long.
Now, a couple other things.
Unfortunately, we cannot offer tech support on this.
We don't have a support department for this.
If you're having trouble running it, Go on to, I suppose, YouTube right now, and like, how do I use LMStudio, or how do I use GPT4All?
Because the way you'll use our language model is exactly the same way you would use Mistral, or Llama2, or Phi2, or any of the others.
You download a file, you run it locally, you do local inference, and then that's your chatbot on your own desktop.
We can't provide support.
We're not charging for this.
We're not earning anything on this.
There's no advertising.
There's no commercial aspect of this.
So unfortunately, we cannot offer support.
We will, however, attempt to continue to build better and better tutorials to show you how to do things with these models.
And one of the areas that you'll want to learn about is called prompt engineering.
And it means that the better you are at writing prompts, that is asking the questions, then the better answers you're going to get.
You can even search online for prompt engineering.
You'll find a lot of videos about that on YouTube, or you'll find a lot of articles about prompt engineering.
If you're good at prompt writing, You will be able to get a lot of great answers.
And you can even use these language models to do some pretty amazing things.
You can have them summarize articles for you.
You can have them expand your bullet points into articles.
You can have them write things for you.
They are generative models after all.
They generate text.
By the way, they're not awake, they're not alive, they're not aware, they don't have consciousness, they're not intelligent, and they do not plan.
All they do is they predict the next word in a series.
That's why it's called a regression generative system.
It's just predicting words.
That's all it's doing.
It's just pure math, okay?
So there's actually nothing spooky about this, and sometimes it doesn't work that well.
But overall, it's pretty impressive.
You can use it for a lot of tasks.
You can generate articles.
You can have it write product descriptions if you have an online store.
You can feed it.
You can paste in a news story that you read and say, hey, give me the five most important bullet points, and it will do so.
The thing that all these models are not very good at is controlling the length of the response.
Sometimes these models will give you a response that's way too short.
Other times it's way too long.
You're like, oh my gosh, it's still writing?
Well, you can stop it.
You can cut off an answer if it gets too long.
Other times these models will repeat themselves.
This is an issue that we have seen in all of our models in early training.
When we have not yet trained them enough, they will tend to repeat.
Sometimes they repeat words like they're stuttering, like it'll be FDA, FDA, FDA, FDA. Other times it will say a couple of sentences and then it will repeat those same sentences three or four times.
This is not a bug.
There's nothing wrong with your computer.
This is something that will improve over time as we do more training with a larger context of material that provides the statistical hyperdimensional parameter model more anchors in which to determine where a sentence should end or where a paragraph should end or where an answer should end.
So if you see that kind of behavior, don't be alarmed.
Just ask it again.
You can also turn down the temperature A lower temperature means less randomness.
And there are other parameters that you can alter in the software like LM Studio in order to increase the penalty for repeat information.
For example, that's one thing.
I think that's called top K or it's one of the parameters with a K in it.
You can turn that down a little bit and you'll get shorter answers, more consistent answers with less repetition.
So it just depends on what you're looking for.
You will need to gain some experience with how to use this correctly.
And you can't just feed the model, let's say, a whole chapter of a book and say, hey, write me a 5,000 word summary of this book chapter.
It's not going to pay any attention to your request for 5,000 words.
It's going to do what it wants to do.
It might be short, it might be long.
Length control is not good in any of these large language models, not from Microsoft, not from Google, not from anybody.
So the best way to control that is to break up your content into smaller chunks.
Like give it 400 words of text and say, hey, summarize this section.
And then give it the next 400 words and ask it summarize this and another 400 words and so on.
And that way you can have it create summary points for a much longer article if that's what you want to do.
And you are in more control over the length of the response that you get.
In any case, here's the good news in all of this.
We've already done the hard work of building a massive data pipeline.
And we built it.
We wrote all our own custom code.
We own all the hardware.
We own all our own code.
We have our own development team.
And I've been on top of this since Thanksgiving, which is why I haven't been getting as much sleep as I probably should.
That's okay.
It's totally worth it.
We have a great data pipeline in place.
We're expanding our data set dramatically.
As more and more base models come along, we're going to be training on top of those and releasing these.
They will all be non-commercial.
They will all be free of charge for you to download.
This means that throughout this year, you're going to have models that are better and better at reflecting the vastness of human knowledge that is, frankly, being exterminated by big tech.
And as we train larger and larger parameter models, like a 14 billion parameter model, it has more subtle knowledge capabilities.
You're able to, with larger training efforts, actually sort of embed more knowledge into a larger model than you are a really small model.
And so this is our goal.
And thanks to your support, Of everything we do, you know, brighttown.com and so on.
Thanks to your support, we have the funding to do this.
We don't have outside investors.
Nobody gave us any grant money.
Nobody wrote us a check.
We're just taking the money that we've earned.
We're donating it to our own nonprofit, the CWC. And the CWC is spending that money on hardware and developers in order to build this.
And then we're releasing it for free to pay back, to pay it forward for humanity.
You support us, we support you with tools like this.
And you'll see soon how powerful they are.
Once you learn how to use these correctly, it's a game changer in your life.
If you're in an industry where you generate text, maybe you write summaries or product descriptions or little updates, news stories, whatever, maybe you're analyzing food ingredients or what have you, this is a game changer.
You'll literally be able to just type in a list of ingredients From a grocery product that you buy and ask the model, hey, tell me if any of these ingredients might be hazardous for my health.
And it will tell you.
It will analyze every ingredient and give you the answers.
And that's based on hundreds of thousands of articles that we've published at naturalnews.com.
10,000 of which I've written over the years because that's my area of expertise is food science.
So you're going to have at your fingertips now the knowledge, the kind of knowledge I've worked on for 20 plus years and all the other experts whose knowledge is reflected in this system.
We're talking about really, in the aggregate, centuries of human knowledge in health and nutrition that will be at your fingertips for free.
With no ads, no spying, doesn't run in the cloud, runs locally on your own computer, even if the internet is down, you can still run it locally.
As long as your computer runs, you can run it.
How cool is that, right?
This is a game changer for humanity.
And that's why we pursued this.
This is all about decentralization of knowledge or the democratization of human knowledge.
And thanks to your support, we're able to do this.
So, bottom line, Download a piece of software to run this, like LM Studio.
The other one is called GPT for All.
Let me bring that up.
GPT for All.
Hold on.
Here it is.
GPT for All.io.
There it is.
A free-to-use, locally-running, privacy-aware chatbot.
No GPU or Internet required.
You can use this software to run our models as well.
And we'll have instructions for you on our website.
There are other pieces of software that exist right now, and there are more coming that will run these for you.
Ultimately, we're going to have mobile apps that will run these on mobile phones, and those will be the smaller models like the Fi2 model.
Which are faster for inference and use a lot less RAM and so on.
So just sort of, you know, subscribe at Brighteon.ai.
If you go to Brighteon, where does it say?
Here it is.
Go to Brighteon, sign up, enter your email address here.
You'll be on our email list.
We will email you when we have new models available and we'll give you tips about what software to use and how to get the most out of these models.
And that's it.
This is our focus this year.
I've spent two-plus decades writing down knowledge and sharing it with humanity in the form of articles and also doing podcasts and interviews and broadcasts.
And a lot of that is becoming really...
Gathered or aggregated or even replaced by large language models.
But we have to make sure that the LLMs are taught the right information and that they don't just spew out like big pharma propaganda or Wikipedia lies or, you know, government narratives.
And it's only through the open source community that we are able to do this.
So we are building on top of other companies such as Mistral.
We give them full credit.
We even credit all the scientists by name who are involved in the Meestrel project.
We give Eric Hartford credit for the Dolphin variants.
We even give Microsoft credit for Phi 2, although Microsoft probably does not approve of the way that we're using it because, you know, We're not an evil corporation like Microsoft, so we're taking their technology and instead of destroying food and farms and the world, we're trying to help protect food and farms and humanity.
Instead of depopulation, we're pro-humanity, pro-life, and that's how we're using the technology.
So it's like we're capturing an evil Terminator and we're mind-wiping it and we're reprogramming the Terminator to protect John Connor to save the future.
That's kind of a Hollywood metaphor for what we're doing in reality.
Very cool stuff.
I hope you enjoy this.
Try it out.
See what you can do with these models and understand that the first models that we release are very early.
They're not what I would consider to be mature.
They're not done yet, really.
We're just sort of doing an early experimental release.
What's coming in the months ahead will be so much better.
So have patience with us.
We'll continue to work on this.
We've got now our data pipeline and our content set.
We've got all that nailed down.
It's working really well.
So we're going to be able to bring you better and better models all throughout this year.
So check it out at brighteon.ai.
And you can catch my videos about it at brighteon.com.
Thank you for watching today.
I'm Mike Adams, the founder of Brighteon and the creator of the Neo large language models that will soon revolutionize alternative media.
Thanks for watching today.
▲
Export Selection