Mike Adams here with AI predictions for 2026 and beyond.
Just quick background, you probably know if you listen to my channel, I'm an AI developer.
I built the Brightlearn.ai book creation engine that has become quite popular.
Thousands of authors have published now over 11,000 books.
They're all free to read or free to download.
And we have auto-translation coming for translating books, which are mostly all in English right now.
They're going to be translated into Spanish.
Not every one of them, but as they achieve a certain number of reads, then they will be translated into Spanish and then after that, translated into French.
And then we're going to move into Chinese and other languages as well.
So that's coming up in 2026.
And yeah, I'm an AI developer.
I did all that.
I'm the only human on that project.
And so I do a lot of vibe coding and built AI models.
BrightU.ai is our AI engine.
And you can see all the AI tools that I've built and rolled out at brightion.ai.
So the reason I say all that is to let you know, I've spent two years in the AI developer space, and I've interacted with a lot of companies, a lot of people, a lot of frontier model developers, a lot of just cutting edge people who are some of the best and brightest in this space.
And I've seen what AI can do well, what it sucks at doing, where I think it's going to fail, and where I think it's going to succeed.
So that's the background for what you're about to hear here.
The big bottom line to me is that 2026 will be a year of massive AI expansion, rollouts, and advances in terms of its core technology.
And one of the big themes of my prediction here is that those people who are saying that AI technology will plateau, that we've reached the end of what LLMs can do, they're wrong.
They're wrong.
Not only are we seeing actually advances in the basic LLMs, for example, at DeepSeek, we've seen DeepSeek sparse attention algorithms that are phenomenal, revolutionary in terms of making large model usage very fast and very cost efficient.
But we're also seeing a lot of post-training techniques being applied that will be game changers.
For example, chain of thought reasoning and using more tokens to force models to go back and recursively check their own work or to derive their own work.
In fact, there was interesting research about a very small model that can score very high on a lot of benchmarks simply by burning tokens and recursively reiterating, you know, checking its own work again and again until it gets to the right answer.
And I think that's what we're going to see.
We're going to see a lot more breakthroughs in terms of scientific research from LLMs that are doing chain of thought reasoning or recursive looping in order to arrive at the best answers and to check their own work, things like that.
And those are techniques that can be applied to the current models.
So we're going to see a lot of those techniques coming out or being refined in 2026.
So no, we're nowhere near the limit of what LLMs can do.
And it's not just about scaling up in terms of the number of parameters.
Maybe the scaling has reached some kind of natural plateau, but the post-pre-training, that is a term, the post-pre-training technology that can be applied or processes that can be applied are going to continue to have breakthroughs for many, many years to come.
And that will also include breakthroughs on model training and fine-tuning.
All right, the next thing that's important to understand is that 2026 will see more models pushed to the edge.
And so a lot of compute right now is heavily centralized because the large models require server infrastructure to be able to run.
You know, many models might have hundreds of billions of parameters or even in some cases, a trillion or more parameters.
So you've got to have some pretty hefty machines with a lot of onboard GPU RAM in order to run those models.
And so two things are going to happen.
First, you're going to have better quantization of models with some additional iterations of fine-tuning of quant algorithms that will do a better job even with a low number of bits.
And I've even seen some models run on two-bit quant with selective loading and unloading out of RAM.
I think I saw a report of DeepSeq 3.2 being run that way.
And I mean, two bits?
I can't even imagine that the model would be any good.
But apparently it retains about 80% of its functionality, even when you quantize it down to just two bits.
It's just unbelievable.
You know, two bits.
That means there's only four possible answers.
There's zero, one, two, and three for each node or each vector in the compute.
That's extraordinary.
So that's one thing.
The second thing is that edge hardware is going to continue to become more and more capable.
And so this rising edge hardware, which will be due to advances from NVIDIA and Samsung and Tesla and Intel and other chip makers.
And I say Tesla because you're going to have a lot of AI compute being pushed to the edge in terms of vehicles.
Like vehicles will be rolling servers of compute.
And some people will even tap into them when they're parked.
It's like, hey, you could be running inference from your car.
Your car's AI engine could be doing work for you while you're charging your car.
In other words, that kind of thing is going to become more common.
But even in other devices, including eventually mobile phones, they will be more and more capable of running more sophisticated models.
And the thing that's going to work for edge is the first point I mentioned, which is the post-pre-training chain of thought recursive loop reasoning type of layers that are added to models.
Those can work very well on edge devices because then it's just a question of time.
Time and how many tokens you're burning to keep looping through the problem.
Where you could have a mobile phone, you could have a decent language model loaded onto it, a reasoning model, and you could ask your phone to solve this science problem.
And it might take an hour if it's a complex problem, but it can work through it through the recursive looping.
And then eventually it will come up with the right answer.
And, you know, hey, if you do that when you're not using your phone, it's great, especially if it's plugged in because it will use a lot of power to go through those iterations.
But edge devices also include desktop computers.
And desktop systems are becoming way more capable because of NVIDIA breakthroughs, among other things.
Desktop computing is getting less expensive, but the chipsets are getting larger.
And NVIDIA is partnering with companies like Asus and Dell and whoever else, sort of the systems integrators.
And they're going to be rolling out what are called the Spark stations in 2026.
I'll probably get one of these.
In fact, I'm certain I will, just because I need to see what it's capable of doing.
But these Spark stations will be running the GP300 chipset, the Blackwell-class chipset from NVIDIA, combined with up to, I think, 784 gigs of unified RAM, which is shareable by the graphics engine, you know, or the GP300 microchips.
But importantly, very high memory bandwidth in the unified RAM.
So it's not the slow memory that they use in the GDX Spark sort of consumer devices.
This is going to be much higher speed memory bandwidth that you can have a computer workstation sitting on your desk that 10 years ago would have been considered a supercomputer from the future.
And I don't know how many teraflops per second it's going to produce.
We can look up the specs, but it's going to kick ass.
It'll be like having a small data center on your desk.
You'll be able to run very large models with very high throughput.
Even you probably, I've done some math on this.
You'll be able to get over 100 tokens per second on large models like DeepSeq 3.2 or models that have hundreds of billions of parameters.
You get over 100 tokens per second, which means that you can set these up as a central AI server in your company or in your business or in your basement.
If you've got enough money to buy one of these to just have it at home, because it's probably going to be 40 grand.
So, you know, it's a pretty significant, it's like the cost of a car, or at least an entry-level car.
But you'll be able to serve, you know, hundreds of workstations within a company or a call center, not all at the same time, but a shared resource that can queue up AI requests and can be your on-site AI server.
So this is going to move a lot of services out of the cloud into the local environment.
Although I will still use both cloud-based and local at the same time.
I found that's the best combination.
Plus, in the winter, I get to heat my buildings with GPUs.
So that's a little bonus effect right there.
But the bottom line here is you're going to see what used to be called enterprise-level compute.
It's going to be distributed at the more local level.
And it means machine cognition will become much more widespread.
And that will have a lot of benefits for society.
A lot of automation of certain roles like customer service and marketing and marketing design and things like that.
And when we get to the point where we can run very competent coding models locally, and actually, Mistral has a model that already does that.
There are some coding models you can run locally that are pretty good.
Not as good as Anthropic.
But when Anthropic level coding models can be run locally, then you're going to have an explosion in app development.
And then things are really going to change.
Okay, also in 2026, I'm expecting we're going to see much better video production using AI with longer duration video segments that can be generated.
Right now, most of the video services only will generate 10 seconds.
Some of them will do longer, you know, 30 seconds or even a minute.
But the longer duration of video generation combined with sound effects and audio with full lip sync of any on-screen characters or avatars, that's going to become much more commonplace in 2026.
And I hope I'm right about this because I'm counting on it.
Because my own project, BrightLearn.ai, is going to be offering not just books, but also the audio books and then mini documentaries, which I'm envisioning would be three or four minutes in duration.
And they would be a short documentary about the book.
And that mini documentary is going to be, of course, rendered as video with narration and on-screen talent, you know, the whole thing.
This is going to be a fun project.
And these will be auto-created.
So if you don't want to read the book, you can see the movie version, right?
Or in this case, a short summary documentary.
It's basically a summary.
That's going to happen, I believe, in 2026.
And I'm anticipating the second half of the year.
So maybe, maybe by the fall of 2026, we might have that feature rolled out.
Just depends on what technology is available and how much it costs.
By 2027, sometime, I'm anticipating more full-length documentary capabilities, but we'll see how long that takes to happen.
The other issue is going to be cost.
Now, costs are going to continue to fall dramatically for compute.
This is for every kind of generation, whether it's text generation, image generation, video, etc.
Token costs are going to plummet by at least a factor of 10x every year, but by some estimates, that number is actually 40x.
So you could take a guess and just say, oh, let's say 20x on average is how much cheaper compute is becoming.
On an annualized basis, compounded.
So that's a big deal.
That means if it's 20x cheaper this year and then 20x cheaper the next year, that means in two years it's 400 times cheaper.
I mean, you know, my goodness, that's a lot cheaper, which means that a lot more tokens can be burned on things like video creation rather inexpensively.
So right now, video creation is very costly and some image creation is also rather costly.
That's going to get dirt cheap in the next couple of years.
And another interesting factor in all of this is that the cost of inference that is online is going to be largely determined by the cost of electricity.
And if you consider that China's power costs are 40 to 50% lower than the United States on average and substantially lower even more compared to Europe, well, that means that China can become a leading exporter of AI inference.
You know, China can set up servers and massive server farms to do all the inference of text models, image models, video models, and they are.
I mean, they're running a lot of data centers.
There are still more data centers in the United States compared to China at the moment, but that's going to change.
China is going to see massive construction and massive scaling up of data centers because their cost of electricity is very, very low.
And that's going to make compute more cost-effective globally because everybody has access to China's computers, you know, through the internet.
And there could be other very clever ways for companies to save money on hosting.
For example, in some of the Nordic countries, I understand they're running pilot programs to turn data center excess heat into a heating utility for the local residential buildings.
So, you know, the heat is heating people's homes or apartments.
And then that subsidizes the cost of inference.
So in those cases, in those more northern latitude countries, heat is a benefit.
Whereas in places like Texas or California or what have you, heat is usually most of the year, heat is a problem that you try to get rid of.
And that costs you money to get rid of the heat.
So compute is going to shift globally to where heat is a benefit and where electricity costs are very, very low.
And that would be northern China.
And northern China just did a big pipeline deal with Russia that's going to bring 50 billion cubic meters of natural gas annually into northern China through Mongolia from the Yamal gas fields of northwestern Russia.
So that's going to make compute even less expensive coming out of northern China.
It should be very interesting.
Okay, shifting gears here in terms of job replacements, we're going to see millions of jobs just in the United States replaced by AI in calendar year 2026.
And that will continue to increase.
But a lot of companies aren't going to admit this.
What companies will say, because there's been a lot of backlash against Amazon, for example, when they got caught admitting they're going to replace 600,000 jobs over a period of a few years.
So there's a lot of backlash.
So these corporations are never, or not from this point forward, they're not going to admit that they're replacing humans with AI.
What they're going to say is, we're just not hiring people.
And we're just going to let the current employee base a trit while we replace them with AI gradually.
But we don't want to have a big announcement that we're laying off a bunch of people and replacing them with AI.
That's bad press.
And in many cases, it might be premature anyway because the AI tools, especially the agentic AI, may not be really ready for the jobs that it's being tasked with.
You know, it can do a certain portion of customer service, for example, maybe 80 or 90% of customer service.
But there's always a small percentage of customer service that needs a human to intervene, something well outside the training of the AI models.
So there will always be some level of humans in those tasks.
It just won't be as many as it used to be.
And companies will just stop hiring.
And as a result, millions of jobs that would otherwise have been created will not be created.
And that's going to result in a glut of human workers who have some level of skill, you know, some cognition.
Doing customer service takes brain skills, right?
And language skills and empathy and things like that.
Or you would imagine so.
And so those people are going to find they're going to discover a very difficult job market because nobody needs entry-level cognition skills at this point.
They just don't need it.
They still need high-level people, experienced people who have a lot of years under their belt.
But the younger generations are going to have a very hard time getting that experience if they can't get hired on into the entry-level positions because the AI is doing all the entry-level work or most of it.
So that's going to be a multi-year challenge of what do we do with all these young people that are graduating out of the university system with degrees in victimhood 101 and social justice, this and that.
What are we going to do with those people?
Well, they're not employable in a rational society.
Anyway, so there are going to be a lot of jobless people who realize they have the wrong college degrees or they just they made a wrong decision to go to college at all and they still owe a lot of money on student loans.
That's going to get interesting.
In any case, these are some of my predictions for AI in 2026 and beyond, and you can check out all of my AI projects at Brighteon.ai.
And by far the most popular project is the book creation engine i've built at Brightlearn.ai, where you can generate a book on any topic in minutes, completely free of charge, an entire book that's researched written edited, fact checked packaged cover art pdf, everything it's sent to you and you pay nothing for it.
So that's a very cool engine and it's becoming incredibly popular.
So check that out again.
That's at Brightlearn.ai, and i've got some other cool surprises coming up for you in 2026.
But the next thing i'm working on is a remote file synchronization system for the, the book engine, so that you can run a little utility locally on your computer on your desktop, let's say and it will sync thousands of books from our servers to your local workstation without you having to download and unzip a zip file or anything like that.
You don't even have to install any software.
You'll just be able to do a remote sync and then you'll be able to have a local copy of all those books and all that knowledge completely free of charge.
So that's coming.
You're gonna love it.
Just sign up at Brightlearn.ai, you know, create a book and then sign up as an author, and then we'll know how to reach you via email when we've got these new features announced.
But yeah, just go there right now.
Just create a book, any book, any topic.
Keep it uh, non-fiction though, if you would please.
We're really not that interested in fiction books.
We're more interested in how-to type of books and knowledge books and things like that research books.
But check it out and get ready, because 2026 is going to be a very exciting year in terms of AI technology advances and rollout and so much more.
I'm Mike Adams, the health ranger, Brighteon.ai and Brighteon.com.
Thank you for listening.
Stock up on HealthRanger's nascent iodine.
Highly bioavailable, shelf-stable, non-GMO, and lab-tested for purity.