What happens when we get AGI?

What does AGI (Artificial General Intelligence) mean? That is the extension of the large language models, LLMs, that can control every data network in the world. Or the system can control physical tools that are connected under their dome. Normal LLM has its domain. The domain is like a state that involves certain actions. Drone control is one domain, and home appliances are one domain. Those domains can have multiple subdomains. The AGI interconnects those domains under one dome or one entirety. So how far are we from that model? 

The answer is more complicated than we can imagine. We can think that the LLM can control things like microwave ovens, but for controlling those tools the LLM requires a socket that it can use to adjust microwave ovens. So the man-shaped robot can use a microwave oven, or the other version is that the home appliances are equipped with a control system that the AI can use to command it. 

When we connect new things under AI control we can face the same thing as when we try to learn to use some new systems. When we buy something new like a microwave oven, we must learn how to use it. In the same way, the AI must learn to use those equipment. And we have two versions for making that thing. 

To use any tool the AI requires a model that it can use in that operation. The model can be in the central server that runs the AI. But where does that server get the model? That is the point. The operator can teach the AI to use the microwave oven as well as the drone. But the system that is connected to the AI can also involve that model. Things like quadcopters must involve programs that control the rotor’s positions. In those cases the operative model is in the robot, or some other thing. The LLM gives orders to robots where they must travel. 

Then the robot can use its internal systems to navigate and move to the location. But orders for autonomic operations are coming from the central systems. This kind of network-based solution is easier for programmers. In those solutions, every single machine that is connected under the LLM domain has its own operational model. The system is modular and each module is independently programmed. 

Basically, if we think that AGI is the tool that just connects multiple devices under one domain, we could do that thing immediately. We can use man-shaped robots that can do almost everything. But the key word is “almost”. 

 Let’s return to the microwave oven. The reason why it’s hard to make that precise thing is the lack of standard user interfaces. The robot must learn to use every single microwave oven independently. That means it must make an independent model for each microwave oven. If there is a system where we can put seconds and minutes separately, the systems where there are only minutes in the timer are not the same. We learn that difference in minutes. But for robots, we must make an independent model of how to adjust the timer. 

Many systems in the world are so easy to use that nobody has wasted time creating standards for them. Easy systems are easy for people, but then we must think about things like the microwave oven. There are button- or toggle timers and that makes them hard to learn. For robots and AI the difficulty is this in the fact all microwave oven models require their independent model of how to use them. 

The robot must connect images from the user manual to the microwave oven’s interface. There is a possibility that if the system does not learn independently the “teacher” or programmer takes an image of the front panel, and then puts the buttons in the right places. Then the AI can learn the rest of the task from the user manual. 

The bias in AI is a big problem.

The AI's problem is that it doesn't think. That is the thing that we usually repeat and repeat. That ability to get data without analyzing it causes bias. The thing that makes that model possible is that the AI doesn't know what the homepage involves. 

It sees words and their connections to something, but then the system doesn't realize what those words really mean. The AI, or large language model, LLM is programmed to give answers in a certain time. That means the LLM doesn't compare and search connections to every single word in the document. That saves time, but it decreases the data's trustworthiness. 

*********************************************

The study put ChatGPT through 18 different bias tests. The results?

AI falls into human decision traps – ChatGPT showed biases like overconfidence or ambiguity aversion, and conjunction fallacy (aka as the “Linda problem”), in nearly half the tests.

AI is great at math, but struggles with judgment calls – It excels at logical and probability-based problems but stumbles when decisions require subjective reasoning.

Bias isn’t going away – Although the newer GPT-4 model is more analytically accurate than its predecessor, it sometimes displayed stronger biases in judgment-based tasks.

The study found that ChatGPT tends to:

Play it safe – AI avoids risk, even when riskier choices might yield better results.

Overestimate itself – ChatGPT assumes it’s more accurate than it really is.

Seek confirmation – AI favors information that supports existing assumptions, rather than challenging them.

Avoid ambiguity – AI prefers alternatives with more certain information and less ambiguity.

(ScitechDaily, More Like Us Than We Realize: ChatGPT Gets Caught Thinking Like a Human)

*********************************************

The fact is that at least part of the results that the AI gets from the net are based on the search engine's page ranks. When we think like this: we teach AI we should give feedback on its results. 

If we tell that the AI gives bad answers that helps it to adjust the results that it uses. The problem is always that for giving trustworthy feedback to the AI we should sometimes be experts. Of the thing that we asked. 

The big problem with AI and other copilots is that those things are made for customers. The product must also please the user in some way. And that can cause the bias. When somebody writes about AI, that person should compare it with other LLMs. The most advanced AI "thinks" that it's the best of all. The algorithms tell it how impressive it is. The more advanced AI means that it's better than other AI's. 

When we think about AI and its operators and owners in the world we must realize that all those companies that develop AI are independent companies. That means those AIs don't have access to the competitor's statistics. So, those AIs are not under one dome. And if they should compare themselves with other AIs like "What is the best AI in the world"? 

Those systems cannot get trusted data. Like how many percent of their answers satisfy the users. That kind of data is not collected. The AI will not ask would the user like its answers or if are they good. Another thing is that if the regular user is not the same as some doctoral user. The AI requires specific commands that involve the right terminology. Without the right terminology, the AI is helpless. 

And then AIs cannot compare those percents with other AI's statistics. This makes the AI think. That it's better than it is. The AI can use Wikipedia. When it should use specific sources like CERN homepages. The AI requires that the person who uses it knows the terminology, and then the misuse of some term in the wrong place in a query can turn the AI's text into a grab. AI will not make a difference in the words "hall" and "hall effect". That thing can cause a situation in which the AI starts to talk about echoes or some aircraft hangars when it should start to talk about resistance in electric wires. 

 https://scitechdaily.com/more-like-us-than-we-realize-chatgpt-gets-caught-thinking-like-a-human/

Calculation power doesn't itself mean that the LLM turns into AGI.

When we think about large language models (LLM) and artificial general intelligence (AGI), we sometimes forget that AGI is an extended version of the LLM. The LLM can handle any mission. We can imagine that it has the right dataset. Then we face another thing. We sometimes forget that AGI just handles the data that it has. The system connects data into new forms like puzzles. To generate things the AI requires data that it puts in the new order. The biggest difference between AGI and LLM is the scale of questions that they can answer. The system is productive if it has data that it can handle. And that is one of the things that we must realize. 

The AI systems are impressive. But they are also computer systems. Those systems have two layers. The "iron" or physical layer. And software layer. The AI can run on separate programs or be integrated into the operating system.  Or the AI algorithms can operate on the kernel when the AI software is loaded into microchips. That thing can seem like "iron-based AI". But it is software, like all other AIs. 

When we think about AI and its shape. 

We must realize. That even the best systems like human brains are useless without information. 

The software sorts information like puzzles. We call that process using the name: "thinking". Human has two thinking speeds. The fast and slow. The slow is the analytic and the fast is like reflex. 

Computers are useless without programs. Those programs are algorithms that the AI uses to control data. The thing. That makes it hard to make cognitive AI seem simple to solve. 

We must make a system that learns like humans. And then mimic that process in the system. When we make a robot that reads a book and then stores that information in its memory we must realize that there are things like program code that the computer can handle quite easily. 

It simply sees the code and then compiles it with its data. The system sees the details or attributes that make the database controller search for the right database that involves the right programming language like C++ etc. But when the system must handle abstract or non-certain data there is a problem. When the system learns something by watching movies the system must put the data to match with things that it sees on streets. 

The problem is that the computer doesn't think. We can show anything like movies about some circus artists and tell that those people that the computer sees are "boxes". The computer can have details about things like boxes. But then it can connect those people in the database there are boxes. That might seem ridiculous. But that is possible. Same way if the robot gets the order to get the car. 

A robot walks to the street and takes the first car. If there is no program. That makes the robot choose only the car that its master owns. The robot doesn't itself make a difference between, a car, lorry, van, or truck. For a robot, they are all cars. So if the robot must get the car to take some sand, it can go to the nearest car and then take it. Then robot can simply put sand on the trunk. The robot must have information about what type of vehicle it needs. For carrying sand. 

That means we must also develop programs. That we can create AI that doesn't make surprises. The large language models are quite a new tool. Advancing is fast. But then we must realize that the road to the AGI can be longer than we expect. Or it can become a reality sooner than we expected. And then we must also understand that there is not a single person that can ask all possible questions in the world. The human's knowledge is limited to data. That human is stored. And there is no "General person" who knows everything in the world. 

Will LLM lead to artificial general intelligence, AGI?

"Artificial general intelligence (AGI) is a type of highly autonomous artificial intelligence (AI) intended to match or surpass human capabilities across most or all economically valuable cognitive work. This contrasts with narrow AI, which is limited to specific tasks.[1] Artificial superintelligence (ASI), on the other hand, refers to AGI that greatly exceeds human cognitive capabilities. AGI is considered one of the definitions of strong AI." (Wikipedia, Artificial general intelligence)

When we think about the AGI and its relationship with humans we can say that the AGI is only an extremely large language model, LLM. That means that to turn into AGI the LLM requires only an extremely large database structure. That kind of database structure is hard to drive. But it's possible. The AGI makes "droplets" of the small language model, SML for each mission. 

We can put all our equipment under the control of the large language model, LLM. Those systems require the computer and socket that the LLM can use to control them. In the traditional model, every single device and system. that we have can involve the computer the system that drives the vehicle on the road or cleans our house. The artificial general intelligence, AGI requires those sockets to control the vehicle. When the user says "Car come here" the AGI locates the person. And gives instructions to the vehicle that drives to pick up the person. 

In that model, the AGI requires that we update all things. That we have. But then we can make another way to create things. That thing is a humanoid robot that has an extremely large database structure. That system can operate in every situation that we can. So, we can say that the AGI is only a large-scale LLM. The robot has only a small computer. However, the internet allows it to communicate with data centers. When a robot gets a new mission the data center generates the data structure or dataset that the robot needs. 

And then creates a more limited, but compact database for that robot. In that model, the LLM creates a series of SLMs to make the robot operate in situations like visiting shops. 

The robot can use three, or four datasets in that mission. 

First, the robot must go out. Then it deletes the "home" dataset and uploads the "walk to shop dataset". 

At the shop, the robot changes the dataset to "operate at the shop". 

Then it changes the dataset to "go home and carry things that you bought". And finally, the last needed dataset involves data that the robot must do when it takes its shopping to the kitchen. In this model, every skill. That the robot has a different database structure or dataset. The central computer cuts a small part of its master data to the dataset series that the robot needs for missions. The humanoid robot is the thing. That can use all kinds of stuff. 

The system can use old-fashioned cars, trucks, and hovering machines. And we can say that the humanoid robot is the socket that can connect everything to the Internet. The robot can use machine learning to find new skills. If a robot must fix the TV it must only know the model of the TV. And then it makes checks like cable checks and other things. Just like humans. 

When we talk about AGI we must realize that even humans cannot do everything. We need to practice everything. When we face some system that is unknown to us, we must read instructions. When a robot learns something it creates a new dataset. And if some data center handles thousands of robots one of them can learn new things and then the system can scale that dataset all over the network. 

Will LLM lead to artificial general intelligence, AGI? That is a good question. The answer depends on what we mean about the AGI. If we think about the situation we can use every single vehicle, that we see. From microwave ovens to taxi cars and street sweeper robots using the AI that asks when we want that street sweeper to clean our yard and robot taxi when it comes to get us we can say that the LLM can be the AGI. The street-sweeping robot can also ask if we need help with our baggage and maybe that same robot can cut our hair. 

Then we must say that the robot that we see in front of us can make everything that we ask. It can turn to cab drivers cutting our hair cleaning our homes and making pizza for us. That kind of robot has a large scale of skills that it can use in every situation. Every skill that a robot has is a database in a large database structure. The central computer can upload the right dataset to a robot. That it controls. 

When we think about the order "pick up my baggage and take them to the car and then drive car to pick me up". The fast internet makes it possible to download and change needed datasets in the robot computer's memory. In that case, the AI. That control's robot. Creates a small language model, SLM for each stage of the mission. The SLM is a compact, reflex version of the LLM. The system can use SLM in cases where the robot requires fast reactions. 

https://scitechdaily.com/quantum-computers-just-got-smart-enough-to-study-their-own-entanglement/

https://en.wikipedia.org/wiki/Artificial_general_intelligence

How far in feelings the artificial superintelligence must or should go?

Researchers say that artificial superintelligence must have the ability to feel pain. Artificial intelligence can mimic feelings. It can read our feelings by searching and observing our faces. The AI can communicate with people using virtual actors or robot bodies. In the movie, "Aliens", the man-shaped robot called "Bishop" keeps contact with the spacecraft's central computers.

And acts as a medium between humans and computers. In "Terminator" movies the robot "Terminator" feels damage in its body, if it is shot. That is possible for modern robots. In the modern internet the AI or large language model, LLM can also interact with users using virtual actors. Those things can also mimic feelings. If a robot or virtual actor mimics pain and other feelings, that system can play human. 

When we think about the feelings and therapy the AI can give psychotherapy remotely as effective as humans. Those AI-based remote therapies had the same effect as therapies that humans gave. And that means the intelligent chatbot can be an even better therapist than humans. The AI will not get angry. Nobody can harm it, or threaten it. And the best thing about that thing is that. The AI will not try to continue the relationship after therapy. 

Does the artificial superintelligence need to feel pain? Or how can we be sure that thing feels pain? We can create a robot that says "ouch" if we put too much weight on its feet. That doesn't mean that the robot really feels pain. Its sensor tells the computer that something that weighs over some limit is on its feet. And that thing activates the "sense-and-response" circuit. 

That is one thing that makes the system able to mimic pain. In the same way, the robot can say "Don't hit" or hit back when we hit the robot. And it can yell "Be careful" or "You almost hit me" if somebody pushes it with the car. The words that the robot uses are in its database. When a robot sees something it activates the public reaction. But the robot doesn't feel pain. 

Same way we think that the robot is the system that is connected to the morphing neural network that allows it to operate through large language models. When a robot takes command it sends it right away to its central computers that can be morphing neural networks. Those morphing neural networks check how the robot should react to commands. And of course, the LLM must check if the command is in the list of allowed commands. 

The robot is the tool that acts as the intelligent or multipurpose microphone between humans and computer centers. The morphing neural networks can look like one single computer to outsiders. And that is one thing that can make the robot a tool. That can have more flexibility. From the AI. The robot is connected to central computers and maybe quite soon, to quantum computers. The quantum system is the top level of computing. The regular morphing neural networks operate as a gate to the quantum system. 

New quantum systems can send data between quantum computers in the form of qubits. The quantum networks make data transmission more secure than regular binary systems. And those robots require ultimate data security. That should deny things like attempts to hack the system. 

https://bigthink.com/mini-philosophy/if-we-want-an-artificial-superintelligence-we-may-need-to-let-it-feel-pain/

https://www.thelancet.com/journals/lanpsy/article/PIIS2215-0366(24)00404-8/abstract

Researchers noticed that AI has its Achilles heel.

"Researchers showed AI models can be stolen by analyzing device signals, risking intellectual property and security. Countermeasures are now needed. Credit: SciTechDaily.com" (ScitechDaily, AI’s Achilles’ Heel: Researchers Expose Major Model Security Flaw)

The AI is the computer program.  Or a very complex computer program. With multiple layers. That means if somebody attacks the AI server that hosts the large language model, LLM. That makes it possible to steal the AI. Basically, that means the hacker can download the LLM source code into that person's own server. 

Researchers from South Carolina State University. Made an Attack on the AI server. Using electromagnetic waves. That means they didn't make a straight attack against the server. That means the server's firewalls and other things are helpless against the attack. The technology that those people used based the electromagnetic waves. 

"Researchers used electromagnetic signals to steal and replicate AI models from a Google Edge TPU with 99.91% accuracy, exposing significant vulnerabilities in AI systems and calling for urgent protective measures." (ScitchDaily, AI’s Achilles’ Heel: Researchers Expose Major Model Security Flaw)

"Researchers have shown that it’s possible to steal an artificial intelligence (AI) model without directly hacking the device it runs on. This innovative technique requires no prior knowledge of the software or architecture supporting the AI, making it a significant advancement in model extraction methods." (ScitchDaily, AI’s Achilles’ Heel: Researchers Expose Major Model Security Flaw)

“AI models are valuable, we don’t want people to steal them,” says Aydin Aysu, co-author of a paper on the work and an associate professor of electrical and computer engineering at North Carolina State University. “Building a model is expensive and requires significant computing sources. But just as importantly, when a model is leaked, or stolen, the model also becomes more vulnerable to attacks – because third parties can study the model and identify any weaknesses.” (ScitchDaily, AI’s Achilles’ Heel: Researchers Expose Major Model Security Flaw)

"The researchers stole the hyperparameters of an AI model that was running on a Google Edge Tensor Processing Unit (TPU)". (ScitchDaily, AI’s Achilles’ Heel: Researchers Expose Major Model Security Flaw)

So, by using this method. Researchers were able to determine the architecture and specific characteristics – known as layer details – researchers and later hackers, would need to make a copy of the AI model.

When an electromagnetic wave travels through another field affects its strength. That means. Things like radars that send electromagnetic waves through data lines can read data that travels in those lines. The idea is that the system acts like an X-ray system. It sends a radio wave through the data line and the sensor at the opposite side detects changes in that field and it can read data. 

The laser system that makes laser rays scatter can used to steal data from optical networks. Or the system can create a hologram around the laser ray. If the laser ray oscillates that hologram can steal information from those laser rays. 

If somebody can steal the LLM that is a very bad thing. Developing those things is expensive. But there is always some other possibility. Bad boys can use those stolen computer programs to create the evil twin for the use of China, North Korea, and Russia. In the wrong hands, the LLM can create virus codes and control cyber-attacks. 

Normally the ability to make malicious code is denied from legal or official LLMs. But in illegal copies, those abilities can exist. So, hackers can use those LLM evil twins for cyber attacks and industrial espionage. The LLM can also make cyber attacks on surveillance systems. 

https://scitechdaily.com/ais-achilles-heel-researchers-expose-major-model-security-flaw/

Don't use the Chat GPT for financial advice.

The Chat GPT and other large language models are very good tools when they answer multiple-choice tasks where is strict and precise values. But when LLM should answer about things questions that are not so precise as multiple-choice questions. Those LLMs have problems. Things like fusion contracts are problematic. 

Because there are always some kind of different interpretations of those contracts. When the data that the LLM handles involves many interpretations and uncertainty the LLM has problems. The financial information is always short numeric data when we talk about things like stock marketing. The AI can operate multiple or even millions of objects at the same time. 

And that makes the AI cannot make a financial loss. But contracts and juridical texts are different. Many things are different from the precise world. The thing is that the LLM that we call artificial intelligence can handle very well. 

The problems are coming when the AI should find information that is not from the most common topics. It's possible. That those data sources are not valid. Another thing is that the terminology in marketing, trade, and court systems are not very common things outside those systems. 

That causes a situation in which the AI can start to search for things like the term "liquidate" from the internet. And we know that this term might mean something that doesn't belong to financial and trade language. Synonyms can cause problems for the AI. 

When the AI makes analyzes of some systems that system is as good as information that it can use. The AI requires information that is the same way trusted and confirmed as humans do. The thing is while the AI analyses the court decisions and advice should the lawyer start the court process the AI requires all information about the case and the precedents about similar cases. And that means the AI cannot probably predict the court order in the case. 

The AI is the ultimate tool when it uses specific and exact information. But if that information is manipulated, corrupted or something is missing. That means the AI cannot make the right decisions without the right information. AI doesn't differ from humans when it makes decisions. If we think the need for confirmed and trusted information. 

https://scitechdaily.com/why-you-should-think-twice-before-relying-on-chatgpt-for-financial-advice/

The computer requires hardware and software.

The new photonic microchips are extremely fast. They can make many things faster than traditional microchips. And the quantum computers require that thing. Quantum computers are tools. That can make calculations faster than any binary computer on Earth. But as we know. The traditional computer can make any calculation as quantum computers. The binary computers just make those calculations far slower. And that means that quantum computers make things faster. 

But both computer types must follow the orders of mathematics. Even the fastest calculation tool is helpless without formulas. They need to make solutions that we can accept. So, we can say that even the best and fastest computers are helpless without mathematicians. However, the computer requires programs that turn mathematical formulas into understandable forms for the computer. 

Hardware and software are both important. Without software, the computer doesn't do anything. The software has three layers. The programs, operating systems, and kernel. The last one interacts with the hardware like microchips and other tools like mice, display adapters, and sound circuit cooperation. 

Or then we must realize that the computer requires programs that it can do anything. The programs and operating systems are things. That is vital for computers. And of course, computers require input and output tools like keyboards, microphones, and screens. And then there must be an internet that computer programs like AI use to for search information. Here we see what kind of problems we have with computing. 

We develop processors but we should also develop programs that support those new processors. Things like biological processors that involve living neurons are not solutions for all problems. The microchip must transmit data to neurons and receive it. But that means the computer must program the neuron. 

We can think about things like "iron-based AI". That term means the computer chip or microchip that has the large language model, LLM in its kernel. We have the "iron-based (A)I) in our nervous system. We can say that when we learn something we program our nervous system. 

That kind of microchip is a tool that involves miniature computers. So that means that the layer where that tool called LLM exists doesn't matter. It can be an independent program or it can be a layer in the operating system. If the LLM is the layer on the operating system all commands to the computer travel through the LLM. The thing is that the AI is the computer program that is in some layer in those computers. The thing is that the network-based solutions offer laptop users the ability to use complicated AI-based systems over the internet. 

Those complicated and hard systems exist only in the remote computers far away from the users. The code is driven by supercomputers. And the home, or office users can give orders to the LLM through the internet. That means the user can have lightweight systems, and the rough force that is needed is in some cave, maybe far away from the place where that operator sits. 

Intensive interaction over the net requires extremely fast and safe communication. The system requires highly secured data highways there information is out of the hands of outsiders. The system must pack the data into the form so that the attackers cannot see or touch those data packages. This is the difference between quantum and binary networks. In binary networks, the data is just mixed in the encryption process into the form that its reorder is difficult. In quantum networks, the system denies the attacker to see the message. 

Can the AI break the net?

Above: Image By Gemini. 

When we talk about the collapse of language models. We talk about the phenomenon where artificial intelligence starts to recycle data through it. In many models, the AI that uses certain data sources starts to raise their page rank in the web index. That means that it's possible. That data starts to travel in the ring of the most used LLMs This thing causes an effect called data degeneration. The data degeneration means that as DNA will not be completely copied in living organisms in networks data will not stay perfectly in its shape. 

There are always some kind of errors. And some turbulence in the networks. The network sometimes lost a byte or two. That thing causes the loss of information that travels in the networks. At the beginning of that process, outsiders do not recognize anything. But when there are enough lost bytes or bits that thing causes the effect, that there are lots of missing parts in data. So that means that data is degenerating in the net in the same way as genomes. Another problem is this. 

If the AIs or LLMs use or recycle only certain data sources that can make the change to put new data into the system harder. The new data is always published in new homepages that are not yet ranked. That means that the AIs can bypass the newest possible data, and use old data. That is page-ranked. The thing that can fix this problem is that the AI can have an algorithm that helps it select the right data sources. 

And then the algorithm must have the ability to select data by using the publishing date. That date should be seen in search results. And of course, the system must use trusted data sources. So it should ask if the purpose of writing is scientific or if it's made for entertainment. 

And that can help the AI select trusted sources like university databases. And trusted publishers.  That helps the user to select data that is trusted. 

The AI must handle much more data than traditional web servers. That means the AI requires lots of energy. And if there are lots of users. That data mass can block the net. In some models. The AI that searches data from the net bombs search engines so much, that the servers will fall. That is one scenario where the AI can simply break down the net accidentally. 

https://www.technologyreview.com/2024/10/31/1106504/ai-search-could-break-the-web/

The quantum computer can be the answer to AI requirements.

The combination of pressure and low temperature makes it possible to create new types of superconductors. That thing allows researchers to create new fundamental quantum computers. And as you see advancing of quantum computers advance their advance. When quantum computers participate in quantum computer research and development, that thing makes them more advanced than ever before. 

Another thing is that: quantum computers advance slower or less dramatically than the first quantum computers. The quantum computers are the most advanced tools that we can imagine. When we think about their role in high-power computing, those systems will become more and more effective. When more people and laboratories participate in those quantum computer R&D processes. That makes them more effective than before. 

The "less dramatic" advance means that. Quantum computer advances happen in places. Where the users cannot see. That thing is the increase in the qubit states. Or making more advanced AI. That can control qubits. Somebody is sometimes the next breakthrough in quantum computing.

The next big breakthrough is the room-temperature quantum computer. But to make that thing real researchers must work hard. Things like crystal technology, artificial diamonds, and extremely stable conditions can make the room-temperature quantum computer real quite soon. 

It's possible. The quantum computer's super cold heart is in the pressurized and very cold gas in miniature tubes. Then there can be the thermos tube there vacuum isolates the superconducting part of the quantum computer. The idea is that the pressurized and low-temperature part of the system is as small as possible. 

Almost every day we can read about some fundamental advances in quantum computing. The next-generation materials require very complex models. The manufacturing systems must react immediately if something in the environment changes. Those next-generation materials are things that quantum computers require. 

Quantum computers require new types of encryption. But they can also run the new AI models. The new large language models, LLMs are things that are more and more advanced. Every new skill that the LLM has spreads its size. And LLM requires lots of machine power. The new LLMs are more multitasking and versatile. That means they are larger. And those larger LLMs require new types of database structures and new types of code control. 

Quantum computers are the answer to the computer requirements that the new large AI models require. And even if the LLM does not turn larger or more versatile. The growing number of users means that the physical systems, or iron must increase its power. And that is one of the things where quantum computers can give a response.  The quantum computer itself can be energy efficient. But all its support systems like large-size cooling systems require lots of energy. 

When requirements grow, programs grow.

The main problem with computing is when developers make something. There is needed new solutions almost immediately. Those new solutions make the application interesting. And many times. Those new solutions mean new abilities for applications. New skills require new libraries. And that thing cumulates the size of the program. 

The new libraries require more space and more effective computing. The thing is that AI is going to more detailed program code. When we think about the first AI "chatbots" the simple programs that asked people some questions, and then the program gave some pre-programmed answers, we must realize that those programs looked intelligent. 

One example of those programs is a program that asks: 

Are you a boy or a girl?

A: Boy

What's your name"?

A: John

How old are you?

A: 23

Where do you live?

Bristol

Then the program said: 

Please to meet you: 

You are a boy, whose name is John. You are 23 years old, and you live in Bristol. And I'm your servant. 

Those things are quite easy to program by using C, Java, or C++. But that program puts answers to the right points in the programmed text. And the form of the text doesn't require us to notice things like "she" or "he".  That simple code is an example of pseudo-intelligence. One "else" command makes it possible to create an answer that if the answer is something else, the computer answers that "you must answer "boy" or "girl". That makes those programs look intelligent. If the programmer has time, it's possible to add all British first names and all the world's city names to the list of allowed worlds. And that gives an appearance of intelligence. 

The thing is that when we think about the modern AI that can control things like cars, we must realize that in the middle is a large language model, LLM. The LLM itself does nothing. But it transfers the data to the part of the code. That controls the self-driving cars. That LLM stays at a tolerable size. But the libraries that the system needs grow. That makes it hard to control data in the system. 

The network-based architecture means that the system can have multiple, limited AIs or LLMs that are connected into the entirety. That thing makes the system more flexible. Because the system has many separate operating cores. It allows the developers to create those cells as modules. Every module is an independently operating part of the structure. 

Above: Mesh network. In network-based systems, every computer runs its own LLM and modules connected to it. So every computer involves certain action series. And, each computer is responsible for its part of the system's skills. 

The structure might look like some kind of mesh topology. That means that there is a central LLM. That is called the specialized LLM to make the mission. The system operates like a mesh topology. When as an example a robot car drives a robot to the shop it sends a mission to the next operator. 

That operator is the robot that goes into the shop. The mesh model means that the system must not bother the central actor all the time. That means that if the central processor or LLM is busy or the system is out of connection the sub-system can continue its mission without the central actor's support or control.  

When a user gives orders to the central LLM. That can transfer the mission to the sub-LLM that has data to complete the mission. The networked structure makes it possible that one mission doesn't keep the entire system busy. When one module operates with some mission. Other parts of the system can perform other duties. 

In modern neural networks, the middle of the system is the LLM. The LLM is connected with sub-LLM systems. The system looks like a neural network. When an owner calls, the car to the door. The central LLM gives a mission to the LLM that controls the car. The car's AI system can interact with surveillance cameras and the surveillance system tells if there are some dangers in that area. 

Or if there are no surveillance cameras the owner can send a drone that looks if there is something behind corners. The drone can hang over the vehicle and send the image to the computer where the car's corners are going. The car requires a 3D camera system and radar. That it can drive safely. 

And then it requires large-size databases that tell what the car should do in all situations that it can face on the road. There are lots of things that the system must avoid. When we think about the models of the complex systems we must create the systems using the cell- or neural-based networked architecture. For these reasons, I told you before.