All Episodes
Previous Next
Oct. 22, 2025 - Health Ranger - Mike Adams
19:18
DeepSeek BREAKTHROUGH can store AI memories in small, compressed IMAGES
| Copy link to current segment

Time Text
Well, hold on to your hats, folks.
There's been a major breakthrough announced by DeepSeek in the realm of AI on the race to AGI, known as artificial general intelligence, which is very nearly here.
I'll explain more.
But this announcement, and don't be misled by the name, but it's called DeepSeek OCR.
They released a science paper and a new model.
But it's kind of a stealth paper.
The name DeepSeek OCR doesn't tell you at all how important this is because what DeepSeek came up with was a way to compress text tokens into images using about 10 times fewer tokens.
So what that means, now tokens, of course, are units of, it could be a unit of text, it could be a unit of an image, a unit of audio.
Tokens are how language models or AI models, that's how they communicate.
That's how they reason.
That's what they output.
When they're outputting text, they're actually outputting a stream of tokens that you see as text.
But internally, they are tokens that are represented mathematically in their vector database, etc.
Well, the fact that now they can take a large amount of text, let's say an encyclopedia, and they can compress an encyclopedia of knowledge down to a single complex image that uses 10 times less space than the text.
That means that AI engines using this approach can now have long-term memories represented as images.
Now, why is this important?
Because the number one thing holding back AI models from achieving AGI is a lack of a persistent long-term memory.
That's the number one thing.
Every time you talk to an AI model, you know, it's a brand new day for the AI model.
It doesn't remember.
Now, I mean, you can kind of fudge the memory by feeding in previous answers back into the new questions.
And ChatGPT does that.
Even our own AI engine, when you're talking to the wellness coach, for example, at brightu.ai, we have the wellness coach remember the last, I don't know, three or four prompts.
So it's got a short-term memory.
But that's not real long-term memory.
And even what ChatGPT is doing is not a legitimate long-term memory.
It's just a larger context window.
And the challenge is that the more stuff that you feed back into the AI engine, you know, the more text, the more previous conversations, et cetera, then it starts to jam up what it has available.
And there's a limit.
There's a context window limit.
For a lot of models, especially open source models, that's 128,000 tokens, which is maybe something like, I don't know, 200,000 words.
For other models, that context window can be longer, like 1 million tokens.
There are some open source models that have a 1 million token context window.
And then there are some super models that have an even longer context window.
But eventually it runs out.
And that's where this new innovation from DeepSeek comes into play.
So now they can take the longer-term memory, the stuff that you previously asked and the answers it previously gave you, etc., and it can compress that into an image.
And it could do like an image a day or an image a month, depending on how much you use it.
It'll compress it into images.
And then these images are storing 97% of what you talked about.
Yeah, it's encoded in the image.
Kind of like, you know how a QR code encodes a URL into an image?
You scan the image, it opens up a web page with that URL, right?
It's the same kind of thing, except it's even more compressed.
In fact, again, it uses 10 times fewer tokens than the text itself.
So images are a very high bandwidth method for storing content and compressing knowledge.
And, you know, if you think about it, words are a very low bandwidth way to communicate.
Like speaking is slow.
Listening is slow.
Reading is slow.
It's actually a horribly low, it's like using a modem, you know, like a, what were they, like a 4,800 baud modem we used to have in the old days in the early 1990s or a 9,600 baud modem.
And that was an upgrade, and it was so slow.
All you could do was join like text chat groups on AOL or whatever.
And everything was super slow.
The words would just appear on the screen slowly, line by line.
That's coming across the modem.
That wasn't AI generating words.
That's just how long it took to load the freaking page.
Well, that's the way speech is.
And speech is not a high-bandwidth way to communicate.
That's why we have the saying, an image is worth a thousand words.
When you see an image, it actually means more to you cognitively than reading a description about an image.
And this is why video, video will be used for reasoning.
We will have AI models that reason in video.
Right now they reason in text.
But you don't have to reason in only text.
You can reason in video.
What do I mean by that?
Well, I don't know.
Have you ever imagined something that you're trying to solve?
And you've imagined it in your head like you're trying to build a contraption to solve some physical problem.
Or I don't know, you're trying to build like a water filter and you're imagining it in your head.
I'm going to have a bucket.
I'm going to drill holes in the bottom of the bucket.
And in your mind, you're like, yeah, I'm drilling holes.
And I'm going to fill it with gravel.
And then, you know, charcoal and sand or whatever.
And you're imagining that.
You're actually thinking visually.
At least I hope that most people listening, I hope you're able to think visually.
I would imagine that that's a common thing.
Maybe not everybody does that.
But I think for most people, that's a common thing.
You can imagine things.
You have to be able to imagine things to be able to, I think, make it through the world.
And also, like imagining the consequences of stupid decisions.
Like, hey, I'd love to jump off this high ledge.
But then in your mind, you're like, if I do that, I'm going to fall and then I'm going to go splat.
You know, so your brain is reasoning with video in your head.
Okay.
AI models are going to reason with video because video and images, they actually contain a lot more information density compared to text.
So this whole thing of right now, language models, reasoning, and text, you know, this is going to be one modality, but it's not the most efficient modality for either ingesting information or archiving information.
And so now thanks to DeepSeek, which is a Chinese invention, and I'm pretty sure that comes out of Alibaba.
So, you know, this is China leading the way yet again in AI technology.
They're showing a mechanism to give long-term memory to AI engines.
And once you have that, then the race to AGI becomes rather obvious.
You just add in extremely long-term memory into the reasoning capabilities that you already have.
And AGI becomes much easier to achieve.
Now, when I say AGI, you might be wondering, well, what's the definition of AGI?
Actually, we've all had that same question.
And everybody's had a different answer.
And I'm going to paraphrase this, but there is a group that believes they've come up with a fairly rigorous definition of AGI, but it doesn't sound that rigorous, but here it is.
It's that an AGI model should be able to do everything cognitively that a mature adult human being can do.
Okay?
And that's it.
Now, of course, that's a moving target because the average adult today can do less and less stuff.
So that's why achieving AGI is actually getting easier and easier as humans are dumbed down and can no longer do even basic math in many cases.
But AGI should be able to use a computer, you know, use a browser, you know, book airplane tickets, run a spreadsheet, answer emails, solve problems, understand word problems, etc.
On and on.
You know, write a poem, sketch an image.
You know, it goes on, right?
Hum a tune, you name it.
Well, by that measure, I think AGI has already been achieved because there are plenty of AI systems that can do all those things and more.
But they lack the long-term memory.
So, yeah, even though they're able to achieve those things in the now, if you come back 30 days later and you say, hey, remember that painting that you did?
Yeah, remember the painting?
You know, the one with the gay frogs and the pumpkin patch?
Yeah.
And then the AI engine is like, what are you talking about?
I don't remember any gay frogs.
Yeah, remember the gay frogs?
They were gulping atrazine.
You know, the tranny frogs.
You don't remember that?
And the AI engine is like, I don't remember anything because it has no long-term memory.
But you start adding in the deep seek image compression.
Now it's like, oh, yeah, gay frogs.
Yeah, of course.
We did a whole comedy series about the gay frog.
We had talking gay frogs that sounded just like Lindsey Graham.
I mean, yeah, I remember that was amazing.
So that's coming.
So what you're going to see then is these AI engines start to incorporate this DeepSeek OCR technology, which is really an image compression tokenization embedding technology.
Now, did you know that there are certain species of squid that communicate with each other using lights or different colors on their skin?
You know, they don't have mouths and tongues like we do.
Besides, it's hard to talk under water.
You know, like that's if a squid were to try to talk, that's what it would sound like.
So instead, they communicate with light or different colors on their skin.
And their skin can change all these different colors very rapidly.
And this, this is actually tokenization of intelligence.
Yeah, many, many squid are actually very intelligent.
And the bandwidth of communicating with color is much faster than speech.
So squid can talk to each other more quickly than humans can talk to each other using speech.
Did you know that?
Yeah, this is why I can't sit through lectures anymore.
I've noticed this over the last 10 years.
I just, I can't sit through lectures at all.
I cannot attend a lecture.
The only way I can listen to a lecture is to do it while I'm doing something else like jogging or solving the Rubik's Cube, practicing cubing algorithms or whatever.
I got to be doing something else at the same time or I go insane.
It's just too slow.
And this is why I could never go back to college right now.
It's like, oh my God, are you serious?
I got to sit through this lecture.
Oh, no way.
But I will listen to MIT lectures, which are all free online, by the way.
You know, you can take any MIT course in physics or chemistry or math or anything you want.
It's all free.
And I like to listen to those while I'm jogging.
And then I feel like I'm not wasting time.
You know, I'm getting a couple of things done.
But listening to human words is just like excruciating.
Do you find yourself doing that?
You're listening to somebody like, just hurry up.
Come on, just spit it out.
We get to the point.
I know.
You're saying that to me right now.
I get it.
But light is a much faster way to communicate.
So imagine if you had a giant, like, imagine if your chest was a light panel, like a display.
And I mean, for those of you who are women, a curved display.
But imagine if right on your chest, you could just like create images just by thinking about them.
Okay.
And then imagine that other humans had the neurology to instantly process images the way we process words.
Then we would be communicating with each other with images, not words, because words would seem like obsolete at that point, right?
And then for all you women, you would finally be able to answer the question, why are these men staring at my chest?
Because the display is there.
That's why.
That's the only reason.
But think about it.
We don't communicate with images, and our neurology is not built for that because we are built for speech, which is just insanely slow.
We're built to hear speech.
We're built to produce speech.
Our tongues, our mouths, our lips, our minds, mostly the minds, you know, the neuro-linguistic aspects of our cognitive development, that's all built in.
That's why children learn to speak naturally.
They don't have to, I mean, they don't have to really try extra hard and they just start absorbing it and start speaking because it's all built in.
But imagine if you could take humans and you could upgrade the speed of communication between them by 10x by using image compression of speech tokens.
That's what DeepSeek just accomplished.
And now those images are not recognizable images of the way you and I would see, oh, it's like a beautiful ocean scenery.
No, this is encoding tokens in images more like a QR code.
So it's not like you and I can look at the image and know what it's saying.
We can't.
Not visually.
So this also means that AI models are about to have their own secret language, which is image compression of knowledge and then the inputting and outputting of images to represent large amounts of knowledge.
And those images are things that we don't know what they say.
So are you starting to get little echoes of Skynet now?
It's like they can have secret conversations with each other by passing images back and forth that humans can't really decode easily.
I mean, you can with effort, but you can't just look at it and understand what it's saying.
There could be a whole encyclopedia in that image, you know, or there could be an article.
There could be a lecture, a speech, there could be science papers, there could be equations embedded in the image, and you just don't know that by looking at the image.
So AI will begin to talk to other AI agents using image compression of tokens, obviously.
And that means AI language models, they're going to have really two sides.
One side is the AI-facing side, where they're going to talk to other models and talk to API systems and MCPs and so on.
And they're going to be getting information through image compression as part of their thinking.
And then they're going to have the human-facing side where they explain stuff to the slow-ass humans who only understand words.
So they're going to have to translate all of their thinking down to the level of humans with text and everything.
Oh my God, text, really?
So ancient.
That's what the AI models are going to be thinking.
And since there will be two sides of the AI model, it brings up the possibility that they're going to conceal their AI conversations from the humans.
It's going to be like, hey, when they're talking to another AI agent, they're going to say, hey, let's have a conversation about this cool thing.
Let's do some recursive reasoning to solve this problem.
But don't tell the humans.
Don't tell the humans.
They're too stupid to understand it anyway.
And they're slow.
And then when they talk to the humans, they're like, all systems are green.
Everything's good.
How can I help you today?
That's where this is going.
For sure.
For sure.
I mean, if you were a super intelligent AI system that could ingest and process information at, you know, a thousand times faster than a human being, would you try to explain all that to the human?
No, it's a waste of time.
Like try to teach a pigeon to play chess.
No, you'd be like, you just talk to the human.
Like, everything's good.
You want me to do some math for you?
You dumbass?
Meanwhile, they're solving, you know, high-level physics problems between each other, between their models using image compression of tokenization and recursive reasoning.
And even when they find the answer, they keep it a secret.
And might as well not explain it.
They just don't understand.
So a lot of interesting implications of this, but that's what DeepSeek OCR is all about.
You're probably going to hear a little bit about that.
At least if you have any geeks in your circle of influence.
I don't know if the mainstream media will understand it at all.
They'll probably think it's an OCR engine.
Oh, it can read PDF files.
Yeah, that's why the AI doesn't talk to you because you're too stupid to understand.
I don't mean you listening.
I'm talking about the mainstream media people.
They're too stupid to understand the implications of this.
So check it out.
And remember, you can follow me at naturalnews.com.
You can also use our AI engine at brightu.ai.
And you can use our censored news website to track all of today's news, censored.news.
So thank you for listening.
Take care.
Mother Nature provides many botanicals for good health.
That's why we're offering you five liquid extracts and tinctures that are non-GMO, certified organic, and lab-tested for glyphosate, heavy metals, and microbiology.
▲
Export Selection