Introduction

So what today's talk is all about. So I'll try to explain where AI stands in 2026, what I'm personally looking at in the future.

And the purpose of this talk is actually not to answer some of your questions, but for me to provoke you to ask me or questions to yourself what to do with it in the future.

And I'll cover a variety of topics starting from the basics of what LLMs do, how they're built, what they run on. And then we'll dream a little bit about what's

How Large Language Models (LLMs) Work

coming for the future so let's start with language what made it all possible and it's really fascinating how we as humans produce words so that's what that's how we are different from monkeys we are able to communicate and share thoughts verbally and this lies as the premise of the whole large language

modal construction so why for example like if you take a monkey and it's kind of intelligent being right so you can train it to do certain tasks but it cannot easily share its thoughts so monkeys can did not produce petabytes of language corpus so hence they are

not as developed as us because of that nevertheless large language models have their own limitations

that we'll speak about a bit later and where the trick is coming from so large language models they're trained on huge amount of word clouds on petabytes of data and they just produce only and only thing they produce the next words in the sequence

sequence so you prompt an NLM and what LLM in a nutshell does it just produce the next most statistically probable word that belongs to the sentence

luckily to us why we find ourselves in probably the greatest revolution in IT

From Natural Language to Code: Agents and Automation

after the emergence of cloud programming languages are languages too so this is actually in the roots of this agentic world so what happens in a nutshell

behind all of those chats how those are called AI agents any kind of thing that produces you output in the background what it does it goes to your computer or to some computer virtual machine in the cloud it generates some code in programming language or in bash commands or in some kind of command line commands and that's how it does things for you effectively

so it cannot think it cannot pretty much predict what's happening in the future it just produces This is the next most statistically probable token or word or command. And it makes things work.

So those are LLMs, and you see that natural language basically transforms into programming language and it transforms into agentic actions that we all enjoy.

Because it's actually very fun to reply to emails with a simple prompt. I actually hate writing emails, so I think this is very cool.

Why the Cloud Story Matters for AI

and i mentioned connection to the cloud before so all of that is very much familiar it looks very familiar to me because i was like still setting up freebsd servers in my classroom back in 2010 and well then cloud emerged and we moved to first virtual machines in the cloud we moved to different kind of functions that execute online you don't need infrastructure you don't need any provisioning.

You don't need five team guys to support any kind of storage and plug and do hot plug -ins and hot unplug -ins of SSD drives if they go bad. So we've all been there.

And this is sort of annoying, especially if you are new to the business. You don't have enough money to hire a team to rent out a data center because those are very expensive.

So what happened when cloud was just in its baby form like AI is now?

A few companies, one of them is AWS, thanks to Jeff Bezos, they figured out that they have a brilliant infrastructure to distribute online books and they have a hell of a lot of compute and storage to spare.

So what they did, they created a virtualization layer, So some kind of thing that enables us to access their recess the resources for free or almost for free and Next on they figured out that it lowered the barrier for entry for all of the businesses

So for example a new e -commerce business would not hire five People of system administrators they would not buy

I don't know plan for five years their computer usage and buy physical service They would just go to Uncle Jeff, sign a contract and get started for pennies in comparison to how much traditional data centers cost.

And next on, another company with a huge amount of compute, huge amount of storage like Google, like Microsoft, like IBM, they joined the game and they started fighting for market. So that's the reason.

Cloud Market Dynamics: Competition, Pricing, and Lock-In

What is the best way to fight for market? It's pricing, right? we all love free stuff especially businesses and every platform started doing things for their ideal customer profile slightly differently they secured vendor lock -ins they secured like

multi -year contracts and then first of all they stopped having enough of free compute and storage So demand equalized with the supply and prices started going up, right? So very, very similar subsidy strategy that we're currently seeing in the AI world.

AI Economics: What Really Drives Cost

So with the AI, I'm not really sure how much we're subsidized right now. Some people tell Forex. I don't really believe that because I'm very much fundamentalist. So how much AI is subsidized is very much dependent on a few things.

So how much electricity costs at the moment of time, at certain location where AI is produced, what is the current relation of supply and demand of compute, aka TPUs or GPUs, so those chips that say to you, that make it possible for you to chat with AI and, I don't know, discuss stupid stuff.

So that all costs real money, and for example, like right now, poor gamers, poor computer gamers because of the scarcity of gpus because lots of bitcoin miners lots of ai companies are buying this computing power in huge amounts market and producers cannot keep up so right now

to build a gaming computer i checked this uh it costs four to five thousand dollars to build uh a decent one so back when i was still playing computer games it would cost me a thousand dollars to build top -of -the -market gaming computer. So 5X grows, and it's not stopping there

before some other companies start competing with Google and NVIDIA. So to me, I've seen this before, very, very similar pattern, just with slight differences. And let's dig into those differences.

The AI Infrastructure Stack

Those are four things that AI runs on.

This is compute, so TPU or GPU, those chips in those huge racks in data center that make calculation possible.

This is storage, which is very much important for both training and inference. This is electricity to power this all up, and this is connectivity.

Connectivity is basically the network that moves bytes in, bytes out for both training and inference.

Constraints and Trade-Offs: Electricity, Cooling, and Connectivity

Why I'm mentioning this because for example, there are also interesting projects by Google or by by the SpaceX to actually build data centers in the space. So connectivity within the space becomes quite a big of a challenge.

You get electricity and power from solar cheaper, but you have to compensate somewhere. And in this case, it's connectivity. So even back in 2010, it's actually interesting how far these companies are going.

Google built a data center underwater in the ocean to save on cooling. So again, they were saving on electricity and cooling to support the growth in compute, but try to connect a wire to something like a submarine or some station under the water. So this becomes a problem.

Is Data a Constraint? Synthetic Data and Practical Reality

I anticipated that you would actually ask that one more constraint is data. So something to train AI on. I would not put it as fundamental constraint for the model. I would put it as the constraint for the businesses who train it.

and right now i pretty much think it's resolved so we came up with a lot of ways to produce synthetic data and synthetic data solves issues quite quite fast so we have a bunch of chinese models that prove that i'm not saying it's easy but it's not on the list very much intentionally

and so so yeah just to wrap it up before i go to this dreamy section uh ai is basically uh your parrot which is running on a lot of compute storage electricity on your on the wires as traditional internet does and to be honest it's not actually very

well equipped for the future tasks and I'll go back to the analogy with the monkey or a four year old four years old child and next slide and those are the

Future Frontiers Beyond Today’s LLMs

frontiers those are not the only two frontiers there are much more that where

AI is going in the future and one of actually this it's a funny thing I checked before this presentation the person who invented world models yan lacoon is actually saying that quantum computing is nonsense and it's not mathematically or commercially viable i would disagree with that but anyway uh this is fun this is funny because those two ended up on

Quantum Computing and Matrix Multiplication

the same slide uh so why put quantum computing uh in this list the answer is matrix multiplication application.

So if you know how AI is trained, it's very similar to what you do, what every one of your computer does when you play computer games. When you play computer games, you have a problem that you have a virtual world, which is 3D, which we're trying to display in two dimensions.

So every object, every small piece of grass, a leaf on a tree, every character that you're interacting with your character in the in the online or offline game is represented as a matrix and every second uh your gpu what it does it tries to figure out how to transform the matrix based on your action so that the right side of the character is shown from 3d dimension

in through in 2d dimension so ai training and ai inference process is different so instead of lots of small matrices we are dealing with a huge huge matrix and we have to figure out a way how to reduce dimensions of this matrix to something that computer can work with but quantum computers are very good at matrix multiplication and this will potentially I believe it will solve the problems of training not

inference so what's the difference AI is built on two things specifically specifically large language models. You have to train AI on some corpus of data. It could be data from visual sensors. It could be data from, I don't know, a collection of classical books of Victorian era.

So for example, like that, if you want AI to speak Victorian English. So it's actually quite fun. I tried it. And this is it.

So matrix multiplication for training, I think it will be solved with quantum computing. Aside of that, we may actually redefine what compute looks like. And world models, again, I'll explain a bit more about quantum and world models on the next slides.

World models is, according to Jan LeCun, actually over here, I agree with him very much, is something that we dream AI to be. be.

So how AI would know that if I drop this microphone, that it will go down with a certain speed and make a boom noise. I'm not going to do that.

But AI would not would not know unless it actually was trained on the data that was mentioning something like that, for example, a book that described that AI would not know it would not be able to predict the next talking. So it is not aware of the physics.

Monkey would be aware after the second attempt, trust me. So if monkey would not like the sound of the mic drop, it would not repeat it the second time. AI would. Maybe in the next chat, AI would.

So it's not actually comprehending any kind of laws of physics that four -year -olds or monkeys, sorry for comparison, actually would.

A Minimal Primer: Bits, Qubits, and Why Quantum Is Hard

And to go slightly deeper, I promise I will not bore you with what quantum mechanics and what is the wave function, but I think it's quite good to know what quantum is and why it's not yet there, why it's not solving all of our computing problems.

So classical computers, they have bits. A bit is either zero or one and of course like in quantum world we probably you've heard a lot of

lots of jokes of Schrodinger's cat there are qubits and a qubit is not just zero or one at any moment of time a qubit can be in a variety of states we call it superposition and imagine that you have hundreds of qubits so well you have a probability function for every qubit in some state and it's a challenge to get this state and it's a challenge

to work with it and actually to sustain hundred qubits you need to put them literally in outer space because they need near out of space temperature to function and also it's hard to uh retrieve data without errors so um if you like i'm shown that

i don't have enough to a lot of time so if you are interested in quantum ask me later

World Models: Toward Physics-Aware AI

there are also world models and we have briefly briefly touched that and unlike llms because you may be bought by i don't know the loading spin and the reasoning part that you may be tricked that AI thinks. No, it doesn't.

So it actually produces a lot more tokens that simulate thinking. So I'll give you an example. For example,

sometimes I think in words when I want to share the information. But in reality, when I think about the concept, I don't use words. I use visualizations or I use those concepts that I'm trained from the very childhoods that if I drop something, it falls. Right.

So AI tries to simulate that. But in the end, all of that is just extra tokens that you're burning. They're just like packaged as more expensive modals or they are packaged as reasoning tokens. But as a matter of fact, it's just text that LLM is producing.

And world modals are slightly different. And there are already some prototypes that are applicable. So it's not like quantum computing theoretical state and with very, very limited practical applications.

Simulations, Latent Space, and V-JEPA Video Models

world models the imagine in the simple words a computer game so computer game has its own physics inside of it so computer game is a great example of so -called latent space with its own laws which is a simplification of a real world so for example in computer game if you walk a few steps and like draw a sword and hit it hits a tree the tree would fall so those are laws that are in

this particular version of the simulation the idea of the world models is to first of all describe the world in those terms that are understandable for the computer not probably as complex as the world we're living in but close enough to solve a specific set of tasks and world models for example

right now they're already optimizing costs of video generation so instead of using stable diffusion models that are just removing noise from the picture they are actually trying and And this is a working prototype made by Meta. They already built in a few sort of,

they codified a simulation of the world, and they codified the behavior of certain objects, like for example, a basketball dropping and bouncing back.

And they managed to produce videos at much lower cost with much higher quality when those principles apply. Google it, it's called VJEPA models.

Again, made by Yann LeCun. Very fascinating stuff. stuff. So this is, I think, one thing worth observing.

Small Language Models and Edge Deployment

And there is, I'm not very good at collecting stuff in slides, there's one thing I would actually watch out as well.

Those are small language models.

So you can actually train something right now and put it on Raspberry or any kind of NVIDIA thing, like a small GPU that you can put in a robot.

And this, again, again is what you can do with this small stuff is fascinating of course it will not be able to provide you with a chat interface that you're used to but this is very much useful because imagine

the power of the robot that writes the code on its own live going around your apartment I'm actually trying to build one so it's uh that's that's again the future so you'll you I think that in the future we'll end up with some models maybe LLMs maybe world models packaged in small small devices and trained for some specific purpose and why with all of

Adoption: Why Most AI Projects Fail (and What to Do Instead)

that in mind why we're not yet cooked why we're still here why there is still potential for businesses for people to thrive in this AI world again people

normal people like me without PhD in quantum physics 95 % of adoption of AI in companies it's very fresh data is from 2025 95 % of projects they either fail or don't bring any meaningful result and interviewer interviewees stated that

main constraints were not modal quality so maybe actually building smarter models and invest in a lot of money is not the way forward this is actually the

adoption problems so how to make sure that ai is secure enough for certain use cases how to make sure that ai is adopted in uh the processes that already exist in or how to adjust existing processes to the a to the ai world and where is the uh result so what we need to do uh yes trouble to be in those this five percent we need to understand how integrate large language

at mobiles and we need to understand uh how to influence people to do that we need to understand both technology and business so yeah just being an engineer is not enough being a business person is not enough so i think the best things come from combining different concepts the best things come uh on the intersection of already known things on the integration level and possibilities are huge

Conclusion and Q&A

technology is there so thank you very much your questions