How did we eliminate hallucinations in our AI engine book generator?
And that's at brightlearn.ai.
And welcome to this podcast.
I'm Mike Adams.
I'm the AI developer of the BrightLearn engine.
And you may have noticed that the book text contains virtually zero hallucinations.
And I've heard this from people.
They're just shocked.
They're like, I normally don't trust AI writing, but your AI engine just nails it, gets it right.
It doesn't hallucinate.
How did you do that?
Well, in order to understand the answer to that, and that's the topic here, we have to understand where does knowledge come from, both in digital and biological neural networks.
So your brain is a biological neural network.
And I submit that you also hallucinate like crazy.
We all do because we don't remember where we learned everything.
So in other words, somebody asks you a question like, you know, do you know what a swordfish looks like or whatever?
And, you know, it evokes an image.
Yeah, I can describe a swordfish.
But then somebody asks you, well, what's your citation for that?
Where did you first learn the swordfish?
You're like, I have no idea.
That's just, I just have the swordfish in my head.
But who knows where it came from?
Well, guess what?
Digital neural networks are the same way.
They know a lot of stuff, but they don't remember where they got it from.
Neither do you.
Neither do I.
The vast majority of things that you know, you can't trace it back to any source at all.
Could have been something, you know, your mommy and daddy told you when you were five, and it just stuck with you.
It could have been something you learned in grade school.
It could have been something you observed.
You don't know where you got your knowledge.
Neither do AI engines.
And so if you're relying on AI engines to cite specific books or authors or titles or movie names or court documents or the right year of something, they're going to get it wrong a lot.
Those are called hallucinations.
And again, the important point to recognize here is that, you know, it's funny to hear humans criticizing machines.
These machines hallucinate.
Yeah, you hallucinate like crazy.
You are a hallucination machine.
Just ask any police detective who's ever had eyewitnesses try to describe what they saw.
Oh my God, humans are hallucination machines.
No two witnesses saw the same event.
So the answer to how we eliminate hallucinations in the book creation engine is that we don't rely on AI language models internal knowledge.
In other words, when our engine is writing the chapters for the book that you've requested, the AI writing agents that we use for this process, they are not allowed to use any knowledge, like specific facts, you know, names, authors, books, dates, titles, whatever, from inside their own internal knowledge base.
Not allowed.
The job of the AI agents is to put together paragraphs and pages using research citations that we provide.
Now, let's go back to being human for a second.
If you're a human writing a book, normally you would spend a long period of time doing a lot of research and finding a bunch of other books or other science papers or other transcripts or interviews, things like that that you can cite.
And you would add those to your references.
You would have a bunch of documents.
You would collect pages or paragraphs with the citations.
And in order to write a comprehensive book, you would probably have thousands of citations.
Well, we do the same thing in our book engine, BrightLearn.ai.
We do the exact same thing, except that our research library contains hundreds of millions of pages of documents spanning every subject imaginable.
And when our AI worker wants to write a chapter, the first thing that my engine does is it goes out and it does the research and it pulls in all of the related documents with authors and citations.
And those could be interviews, they could be articles, it could be books.
And this is why you can see those at the bottom of every subchapter of every book that's created at brightlearn.ai.
There's a list of references.
Now, if there's no references there, then that means there were no references for that subchapter.
But for most subchapters, you're going to see one or more references, sometimes a dozen references, depending on what's in that subchapter.
And those are provided section by section throughout the entire book.
I don't just put them at the end of the book.
I put them in each section so you know what sources are relevant to this specific section.
So then the writing workers, which are all AI, of course, the writing workers, they are handed an assignment.
It's like, hey, here's a chapter, or here's a section of a chapter, and your job is to write this chapter.
And here's all the research that you need to use.
Here's like 250 citations of books and transcripts and authors.
And you need to use this and then write the chapter and then cite these.
And that's what it does.
And that's why we have near-zero hallucinations.
And that's why every book that is created at Brightlearn.ai is a book that is well researched.
It's better researched than almost any human written book that's ever been written because our research encompasses, again, hundreds of millions of pages of documents.
And right now it's 10,000 books that we research.
And that is, we have actually indexed the full text of 10,000 books.
And that number is about to vastly increase.
The next chunk of books we're adding, or the next wave, will be approximately, I'm estimating, 25,000 additional books, full text, will be added to our index.
And I'm currently processing millions of books through our data processing pipeline for classification, alignment, normalization, and many other steps.
And this is a rather large task.
I currently have 48 GPU workstations working on that task.
And they're also processing over 100 million science papers.
And it takes a long time to do that.
So ultimately, although this will take many months, ultimately, our index repository, you could say, will contain many, tens of millions of science papers, hundreds of thousands of books.
I mean, who knows how many millions of pages of transcripts and interviews and podcasts and spoken word and so much more.
And then all of that is available to the writing agents that write the book chapters.
So actually, as we move forward in time, as more people use the BrightLearn.ai engine, it will actually continue to get smarter and smarter because the indexing of documents will continue to expand dramatically.
But here's a fun fact about indexing books.
You can't just, let's say, scan a book and use the text and stick it in there.
It doesn't work because it's messy and because a lot of books have their own indexes and you don't want to index another book's index because then that just looks like garbage in your research.
Like you wouldn't cite another book's index in your own book.
You would cite a proper chapter or section of another book, not an index.
You wouldn't cite the table of contents.
You wouldn't cite the publisher copyright information and all that kind of stuff.
Instead, you would cite, you know, paragraph text.
And so we have to do the same thing.
And a lot of books have artifacts, paragraph hyphenation sentences have been distributed across multiple lines, or there's a page in the way.
And so a paragraph might span different pages and then there's a page number or there's a header or a footer printed in the book, etc.
So all that has to get cleaned up.
And the way you clean that up is you use AI.
And so it's a very compute intensive process to clean up books.
And cleaning up books is only one of the steps.
I think there are something like nine steps that my system actually goes through to take raw book scans and then convert them into usable indexed text.
About nine steps, something like that.
And you multiply that across millions of books and you understand the workload.
Now, the other thing that's important here is the classification of all this content that's going into the index.
It's critical, from my point of view, that we only index books that have a worldview that is aligned with the core pro-human, pro-integrity belief system that I espouse and probably that you share as well.
So, for example, you and I are advocates of freedom, freedom of speech, medical freedom, medical choice.
We also understand that vaccines can be dangerous.
We understand that natural medicine is safer and more effective than conventional medicine, especially at preventing chronic degenerative disease.
We understand that governments lie, that institutions are corrupt, that dollars are not money, that the Federal Reserve is a massive scam that's been running since 1913.
We know these things.
And I don't want books in my index that are pushing globalist agendas or, you know, stupidity like climate change nonsense or transgenderism insanity that pretends a man can become a woman, which is, it's just really retarded, actually.
So we don't want retarded books in our index.
And so that means that we have to go through and we have to classify every book.
So imagine if someone handed you a stack of, here's a million books, and I want you to sort them into two piles.
Like pile number one, books that make sense, books that have pro-human values, books that tell the truth, books that are pro-freedom, etc.
And then in stack number two, retarded books, which probably the retarded stack would be the larger stack, actually.
And so that's your job.
Sort out all the books.
So how do you sort the books?
Well, you have to read them.
You have to read them.
And then after you read them, you have to determine how closely they are aligned with the value system that you espouse.
And so I have defined that value system in great detail.
And my data pipeline, all the 48 workstations I'm running, they go through and they read every book and then they determine whether it's whether it goes into the good pile or the retarded pile.
And then the retarded pile, of course, is just completely ignored.
We don't process those.
Also, currently, I'm only processing English language books because there's an extra step if we want to translate them from their current language into English because the vast majority of books that have ever been published are published in non-English languages.
So I'm only using English and I'm only using books that have strong alignment with our pro-human, pro-truth belief system.
And so if you start with a stack of 1 million books, you might end up with maybe 100,000 books or maybe even less, maybe 50,000 books at the end of the day.
And that entire process takes a lot of time, a lot of compute, a lot of electricity, etc.
But that's what we're doing.
And so somebody out there who's thinking, well, I can create a book writing engine.
I'll just connect to ChatGPT and I'll tell it to write the book.
Okay, great.
That's going to be a retarded book because ChatGPT is a retarded AI engine.
It thinks that men can become women.
It thinks there's unlimited genders.
It thinks that carbon dioxide is bad for plants.
It believes in sane, stupid things.
And it thinks all vaccines are safe, etc., right?
The big pharma is great.
And ChatGPT believes the FDA is awesome and the CDC is awesome, etc.
So yeah, it's going to be a book.
It's going to be a retarded book, though.
It should be called like retard GPT, actually.
So the secret to Brightlearn.ai is not actually creating an app that can write books.
That part, that can be done by others who are, let's say, capable of understanding architecture and things like that.
But the really difficult part is having the repository of index content that is used by the book agents in order to create the chapters.
And that has taken me two years to put together.
And if somebody else were to try to do that, they would run into all the same interesting challenges that I ran into, which are massive.
I mean, I haven't even gone into detail.
I probably never will.
But I'm simplifying everything here.
In reality, this is way more difficult than you might imagine.
I mean, first of all, how do you even acquire a million books?
Or 10 million books or 100 million science papers.
How do you even do that is a big issue.
Not trivial.
And then once you get them, how do you even write the code to process the books?
You know, it's interesting for the first year and a half of this project, I had a team of human coders that were writing the code to do the book processing.
And then in the last six months, I've taken over using AI coding agents to write the code.
And so I've already abandoned all the human written code.
And now I just use purely AI written code to do the processing because, well, it works better.
It's faster.
It has fewer errors.
And I can create it myself just through prompt engineering or what's called vibe coding.
So you could say I used to have engineers on the data pipeline side.
other humans, but I don't now.
And in terms of the Bright Learn engine, I've never had any other person work on that at all.
I'm the only person that's ever worked on it.
In fact, nobody else even has access to it.
No one else even knows how it works.
I haven't shared it with anyone.
Maybe I will at some point, but it's pretty complex.
Anyway, that's a little bit under the hood.
That's why we've eliminated hallucinations.
And it's why the book content is so amazing.
And it's also why the book content is going to continue to get better.
And remember that we can regenerate books as we add more citations.
And I actually intend to do that as resources allow and as the cost of compute continues to fall.
One of the things that we will do in 2026, for example, is we may choose to regenerate most of the existing books.
That is being true to the title and your user prompt and the table of contents.
All of that remains the same.
We simply augment the sub chapters with new research that comes from more books.
So in essence, it's sort of augmenting the existing books with additional citations and research and making them stronger.
So that's something that we may do once we get a lot more books into the index and a lot more science papers also.
That's coming as well.
So the books that are generated at brightlearn.ai are really, they're not static.
It's a living book in essence.
It's a dynamic system that continues to expand as we have more knowledge indexed.
So it's free to use.
You can generate books right now with three chapters in length, completely free without using a token.
Just go to brightlearn.ai.
And if you have a token, you can generate books of much longer length and you can edit your author profile page and you can request book deletes and you can request cover art regeneration and things like that.
That's if you have a token.
So check that out at brightlearn.ai.
If you want to get tokens, you can currently get a free token by just signing up at healthrangerstore.com.
Sign up for our loyalty lion program, which gives you loyalty points when you place purchases at healthrangerstore.com.
Just by signing up, which you just give your email address, you'll get 300 points that you can swap for a book token.
And then if you do make purchases, then you get more points and you can swap those for tokens, you know, all throughout the future.
And you're also helping to support the project because HealthRangerStore.com is what donates the compute to our non-profit Consumer Wellness Center, which actually owns this entire project.
So if you want to help support the project, just shop with us at healthrangerstore.com.
If you would rather just use it free, that's fine too.
Just use it on the free tier and enjoy.
And remember, you can use all of our other AI tools that are also free at brightlearn.ai.
Check them all out, use them, enjoy them.
Again, they're all free and they're all very capable engines.
In fact, arguably the best in the world in each of their categories.
For example, there's no other book creation engine in the world that even comes close to what we've built at brightlearn.ai.
And it's only going to get better also.
So thank you for your support.
I'm Mike Adams, the HealthRanger.
That's why it's called HealthRangerStore.com.
And also I'm the executive director of the Consumer Wellness Center nonprofit.
And if you want to donate to our project, of course, it's tax deductible.
And you can reach us at support at brightlearn.ai.
That's the email address.
If you want to make a year-end donation to help support the project, we'll give you public credit for that.
Otherwise, just spread the word and thank you for helping to support us.
All right, use all the tools and enjoy.
Take care.
Beet juice is a refreshing beverage that can help you stay energized.
You can easily incorporate beet juice into your daily routine with Groovy B organic freeze-dried beet juice powder.