Mike Adams explores "vibe coding" through the launch of brightlearn.ai, a platform boasting 53,000 generated books and 11,000 authors. He details an architecture utilizing 48 GPU workstations for document normalization and image translation, coordinated via REST APIs where humans make critical architectural decisions while AI handles code generation. Citing DeepSeek's cost efficiency over Anthropic, Adams argues that success relies on directing AI rather than micromanaging syntax, enabling unique system creation with 90% of time spent waiting for responses. Ultimately, this approach democratizes complex software development by prioritizing functionality and reliability over traditional coding aesthetics. [Automatically generated summary]
Transcriber: CohereLabs/cohere-transcribe-03-2026, WAV2VEC2_ASR_BASE_960H, sat-12l-sm, script v26.04.01, and large-v3-turbo
|
Time
Text
Millions of Dollars vs AI00:03:27
I see a lot of bad arguments about AI coding, AI developing, and the replacement of AI engineers or just regular human coders, et cetera.
I want to just explain to you briefly here what it's actually like to be an AI developer or a vibe coder, you could say, as I am.
I've been doing this for coming up on a, well, maybe it's 10 months or something now, but I've launched several very successful projects and platforms, for example, brightlearn.ai.
Which is the free book creation engine.
And oh, also, I need to tell you, I've upgraded the free tier to where you can now generate five chapter books instead of just a three chapter book.
So you can generate a five chapter book for free.
And then if you have a token, which you can get at our online stores, then the tokens give you access to much longer format books, like I think 20 chapter books now to generate them.
Anyway, we've had over 53,000 books generated, which makes Bright Learn, by far the largest book publisher in the world.
We have over 11,000 authors and we have coming up on half a million downloads.
I mean, we're not there yet, but we will get there soon.
And I had promised that I was going to offer audiobooks.
That was accomplished a few weeks ago.
We now have hundreds of audiobooks available full length audio, MP3 files, no digital protection at all.
You can just download the MP3s and you can share them and you can listen and enjoy those.
And then I also promised that I was going to create Spanish language books, auto translated books to serve the Spanish speaking communities of the world.
And so that process is what I'm working on right now.
I just thought I would share with you what's involved in that.
And it's actually way more complicated than you might suppose, but it really shows the role of an AI developer like myself.
What is our role?
Because I have not written a single line of code.
In this entire project, I intentionally do not code.
And before AI, if I were to try to build something like this, it would have taken millions of dollars and a team of, you know, at least a dozen coders, probably, et cetera.
As of today, I'm the only person doing this.
I mean, I'm the only human on the project.
And thanks to DeepSeq, the cost of the AI coding has just plummeted by a factor of almost 100 compared to.
What it used to be.
I could code all day now or vibe code all day and it'll cost a couple of dollars instead of I used to burn through hundreds of dollars a day with Anthropic and many thousands of dollars a month.
And apparently those days are done.
Anyway, I still have to make key architectural decisions and I have to correct bad mistakes, mostly architectural mistakes, by the vibe coding agents.
So In other words, the AI is really good at writing lines of code.
It's really bad at making informed decisions about what the architecture of the project should actually be.
Let me give you some examples.
Architecture Decisions Matter00:03:56
So, this book project, as you might imagine, involves a lot of database tables, a lot of relations between tables, a lot of different status fields in different tables, especially when you try to take one book, which is in English, and then you try to translate it into multiple languages, which is Spanish, French, Chinese, whatever.
Which is the goal, then you have to figure out, of course, how do I have a master ID for the book?
How do I know that if I have a Chinese language book over here, what was the parent?
Maybe parent ID is a better term.
What's the parent of that book?
And then you have to manage things like the translation of the book covers, which is by far more difficult than translating the book.
So let me tell you I have to tell the AI engine what data structure to use, what kind of path structure to use.
I have to decide and tell it we're going to put a dash ES on the end of every art name, every JPEG, every PNG, every path, every folder on storage.
We're going to put dash ES because that means Espanol.
And FR means French, etc.
There are two letter codes for different languages.
EN is English, as you might have guessed, but ES is Espanol.
It's not SP for Spanish, it's actually ES.
So, I have to tell it what to do and how to do it and how to handle the data structures.
Now, it gets even crazier when you realize that my, I have a mini data center that I manage.
I've talked about it before.
I've got 48 workstations with GPUs and mostly they're working on cleaning, you know, science papers and books and normalizing documents and sometimes searching through documents.
I've got one workstation that's doing nothing but functioning as a discovery engine.
By the way, all it does is discover new things.
I'll tell you more about that later.
That's pretty cool.
We found like 200 discoveries already.
It's pretty cool.
But I've got one workstation now that just does book cover translations of the image.
And it's incredibly difficult for a number of reasons.
But first of all, my local system has to coordinate with the hosted online system that hands out the assignments of which books need cover translations.
And then my local engine takes that book cover.
Does the translation, does its own grading system, finishes the translation, and then sends it back through an API that I built on the hosted side.
And then that image has to be saved with the correct file name and the correct path name, obviously.
And then I have a different engine that is the rendering engine that actually renders the static pages of the books and the index homepage for the book site and all the books that show up in the Espanol tab and all the books that show up in the Audiobooks tab, etc.
There's a separate engine that's dedicated to rendering all that stuff.
That's also a very smart architecture.
You don't expose your rendering engine to the world through a public interface.
So you can't reach my rendering engine, which means nobody can hack it.
Nobody can overload it with a DDoS attack or anything like that.
It's an offline hidden rendering engine.
And that way, if anything happens to The live website, well, the rendering engine just re renders and pushes it back to the file host that serves up the static web pages.
Hidden Offline Rendering Engine00:03:43
So I've got three components.
I've got my local data center.
I've got the public site.
And then I've got the non public rendering engine.
All these three have to coordinate.
And the thing is, they can't talk to each other.
They can't talk to each other because they're not on the same network.
So I actually have to do things like I have to ask the AI engine on one of these.
I say, you need to write documentation of how this API works.
So I can hand that documentation to another AI engine and tell it how to use your API.
And it does.
And then I take that text file and I paste it into this other local API and say, This is how the API works.
Well, I'm sorry, I paste it into local AI and say, Now you need to build the handling system for how to pull book covers and how to submit them using this API.
This, you know, it's a REST API endpoint, et cetera.
So I actually have to pass messages between the different engines that I'm running.
This is just for the Bright Learn project, not to mention the other projects that we run.
In any case, I've gotten pretty decent at this process and it's working, but it's slow.
It's slow.
Usually I'm waiting on AI.
I'm just waiting on tokens.
I'm waiting on it to figure things out.
I'm waiting on it to read the database or to read the schema.
I'm waiting on it to write the code or to fix the code or review the code.
Honestly, vibe coding for me is about 10% telling the engine what to do and then 90% finding something else for me to do.
Because I'm waiting on the engine.
So that's why I run a lot of projects.
But again, if you don't have high level architecture and maybe experience with databases and code bases and so on, you might not know how to direct AI to do these things properly.
And that's why I find that a lot of programmers, they don't know how to use AI.
I don't know if it's maybe they're trying to over control it or over code it, or they don't like the way AI writes the code.
Here's the secret to that, folks.
Don't look at the code.
I know some of you who are programmers are like, what?
What do you mean don't look at the code?
Yeah, don't look at the code.
I don't care what the code is.
What I care is, does it function?
Is it clean?
I mean, I can tell AI to refactor it.
I don't care what the code looks like.
I care that it works, that it's reliable, that it reduces memory usage, et cetera.
The code, I mean, it's just a proxy for the underlying executables anyway.
You know, I mean, who cares what the code looks like?
The question is, does it work?
So I don't look at the code.
I don't need to see the code.
And I do know how to talk to AI.
And that's why here I am, like more than six months into this project.
Nobody else has built anything like it.
Why?
Because they can't make AI do the things that I'm doing with AI, apparently.
So I don't know.
I just know how to talk to AI and then leave it alone.
I don't keep micromanaging it all day.
So that's part of the secret.
Anyway, those are my thoughts as a vibe coder.
Feel free to use my free AI engines.
Like I said, I've got brightlearn.ai.
Very popular.
I've got bright answers.ai, which is a deep research AI engine.
You can ask any questions you want.
And I've also got bright news.ai, and you can catch more of my work at brightvideos.com.
And thank you for listening.
Take care.
Start your day right with our organic, hand roasted whole bean coffee.
Low acid, smooth and bold.
Lab tested and ethically sourced.
Taste the difference only at healthrangerstore.com.