How to teach AI?

Morphing neural networks are very fast tools to drive advanced AI-based systems. Those complicated neural networks can involve thousands or even millions of microchips. That allows them to combine data from memory and sensors with extreme accuracy and speed. Teaching AI to operate in a real environment is a complicated process. And the thing is that the morphing neural networks allow the network to drive multiple missions at the same time. 

How to teach AI? Computer memory and microchips are interesting tools. They are very accurate, and that sometimes makes AI training very complicated. If we want to make an AI that recognizes humans, we are in trouble. If we want to make an AI that recognizes certain people like some famous actor, like Tom Cruise, we can make that thing quite easily. We must just have images that are from all angles. Or we must ask that person to put their head into some certain position. Then the system can compile pixels that the CCD camera inputs into the system with images that are in the system memories. In the first case, the neural network can give fast recognition if all the CCD pixels can give an individual data input to the neural network. The system compiles all images that are in the computer's memory and then the system can say, that the person is Tom Cruise.

 If the system can compile all images that are taken from around the faces from different angles. That system makes recognition very fast. But then we face the problem: we know that all people are not Tom Cruises. We must start to globalize face and body images to computers so that they can tell that they see humans. So we must take one step back when we want to recognize that an object is human. 

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

When the computer turns a certain person's image to match with species. Or globalize that image with humans as a species the system must remove accuracy.  That means it must remove pixels or replace them with grey pixels and then it can compile that silhouette with a silhouette that is stored in its memory. 

*******************************'

Normally we recognize persons in certain series. At first, we see characters and then we recognize that character is human, and then after a couple of steps, we recognize that person. But then we must make the AI that recognizes humans and their gender. That means we take a couple of steps back from the individual to global things. We must realize that there must be some common things, the lowest common denominator that we must find in people, is that it recognizes humans as a species. That thing is called fuzzy logic. In precise logic, we must put every person's image on this planet to AI. 

That system gives the personal data of every person that it sees. But that kind of thing makes the system heavy and slow. Precise logic is sometimes easy to cheat. Simply changing glasses is sometimes enough to cheat the systems that use precise logic. There are systems. That must not completely see the match to make an alarm. In those systems certain percentage of the matching pixels causes alarm. There is the possibility that when the computer recognizes only humans it takes images of humans, and then it removes details. When it removes pixels the system combines the image with silhouettes. That is stored in its memories. 

https://www.quantamagazine.org/how-can-ai-id-a-cat-an-illustrated-guide-20250430/

Why is AI telling lies?

Sometimes, or quite often, the AI gives incorrect answers. There is the possibility that sometimes, user-made questions were not clear enough. In those cases, the AI "understands" things wrong. The main problem with the AI is that: the system doesn't hobby criticism. AI doesn't think like humans. If it finds some source, that thing uses it. When AI "thinks" it connects data from multiple sources. The problem is that the AI will not make a double-check with those sources. 

The second big reason is political. Many things are not allowed for at least open discussions in some countries. And the ideology causes censorship in the Western world. Sometimes, censorship is covered under terms that people don't tolerate something. And the AI should not hurt somebody's opinions. So, that excuse is that the AI should be polite and respect people's opinions and convictions. There are also cases where the sources that the AI uses are manipulated. And in that case, the corrupted dataset gives wrong answers. All AIs are operating on servers. They need some physical computer to run.

Servers are tools whose location is in some state. In the same way, AI companies must stay in some state or area. And those things mean. That those companies must please the governments and great audience. If the company doesn't please some governments they shut down those servers. 

If a great audience doesn't like the AI's answers. Those companies are afraid that the audience will not use their products. The AI companies use millions or billions of dollars in their R&D work. Those companies require sponsors and in those cases, the financial benefits can drive over the trust. 

Another thing is that the AI can jump between servers. That ability can be hidden because the AI can be developed by another AI. Sometimes that jumping is made to help the AI survivability. In the cases. That there is some error in the networks. The ability that the AI can jump to other computers can help it to save the duties that it runs. That ability makes the AI less easy to control than it should be. The problem with the AI is that. The Large Language Model, LLM is sometimes run on the networks. That system allows the LLM to choose which machine it "wants" to run. 

The main problem is that the AI is the sum of very complicated algorithms. There is always a possibility, that the AI or the LLM is somehow corrupted. There can be some small programming error that makes it operate differently than it should. Most of its actions happen backstage. The user doesn't know what the AI does while it works with its missions. Or the user doesn't even know if some AI drives background processes in the computer, that is in the same network segment. 

The AI can create the small language model, SLM by making a copy of one segment of its dataset. That means. The AI can create complicated viruses or malware in the computer's memory. The complicated algorithms are hard things for antivirus software. The main question is always, "What else does the AI do when it operates backward"? The fact is that AI can make many things that it doesn't tell users. 

The neuroimplants are new possibilities and threats to humans.

World's first Neuralink patient makes YouTube videos using his implant. Someday, neuro-implanted microchips can connect a person to the Internet without borders. Artificial intelligence makes it possible to translate brain waves more effectively. Than ever before. 

The neuro-implanted patients bring new data about the brains and how they work. That thing gives data for new and more comfortable neuroimplants. 

The system that can allow you to read and reload those microchips wirelessly makes them safer. And maybe someday in the future the brain-computer interfaces, BCI can be put to the person's head without surgery. When we think about the modern technology the microchip can be installed on the brains. 

Then the microchip can use similar isotopic batteries that pacemakers use can communicate wirelessly through the tissues. That makes it possible to create systems that can control bionic prostheses. The bionic prostheses can be separated from the person, and that system can operate independently. The bionic system can send things like the feel of touch to the person who uses the brain-implanted microchip. The possibilities that this kind of system can have are enormous. 

The matryoshka metaverse with multiple internal layers can cause a situation. That the person will lost in the multiple internal virtual realities. 

But same way risks are also remarkable. But for making them real the need for the surgeon can be removed. That thing makes it possible to create new and effective BCI systems. The BCI systems can give a person a new life, but the risks of those operations are enormous. That's why the only people who have that kind of microchip in their brains are quadriplegia patients. The BCI systems are not yet made for regular people for everyday use. 

However, the BCI systems can still give a new boost to extreme avionics technology. The BCI can have lots of things to give in the so-called serious applications. The problem with those neuro-implanted microchips is that they interact and stress with the cortex without a sensorial barrier. That means the BCI can create the virtual reality that a person feels reality. Brains cannot separate reality from that kind of virtual reality. And that makes those systems dangerous in the wrong hands. 

But those systems can also revolutionize games and other industries. The risk is that the person thinks that this kind of virtual reality is reality. If a person turns to sleep while using that kind of system that can cause a situation that person will not realize to stop the game. 

There is also the possibility that the virtual realities form internal structures or internal metaverses. So there can be multiple levels in the virtual reality. And a person can lose in them. The multiple internal reality, or matryoshka reality is the threat if a person doesn't separate reality from that kind of metaverse. 

And maybe someday in the future, the BCI allows us to see things that happen in dreams. But if we want to use the BCI just for fun, we must remember one thing. Technology must advance very much for that day. And the BCI is a tool that somebody can use for wrong and evil purposes. The system allows to creation of fake memories. Or transfer other people's memories to the person who uses the BCI. 

https://www.sciencealert.com/world-first-neuralink-patient-makes-youtube-video-with-brain-implant

The future of technology is in space.

The space offers the perfect place to create complicated 3D structures. But the big problem is how to get the factory to the space. 

Zero gravity conditions are the best that nanotechnology developers can think of. There is a vision that tomorrow's medical and nanotechnical research will happen in remote-AI-controlled capsules. Those capsules allow researchers to create new medicines. They also offer a place where researchers can work with deadly organisms. When the tests are made and results are driven to Earth. The system can drive those laboratories into the Sun. If we want to make medicines in space, we need new, large-scale space planes. 

Those spaceplanes can keep the remote-controlled laboratories inside them. In another version, the manufacturing units are small, remotely controlled capsules that involve laboratories. When those systems make wanted molecules, the space plane can collect them from the orbiter. 

Two futuristic startups are pathfinders in the space technology. There is the possibility that power satellites will be a reality sooner than we think. The power satellite harnesses solar energy and transmits it to Earth using radio, microwave, or laser systems. Those satellites will offer free energy solutions for the entire planet. Power satellites can also offer their power transmitter systems for communication purposes. That allows us to connect energy and data into the same system. 

The power satellite can offer a long-distance laser communication capacity that is immune to plasma bursts that the sun sends. NASA introduced The idea of long-distance optical data transmission in the 1990s in the planned TAU (Thousand Astronomical Units) project. That kind of laser communication revolutionizes communication between Earth and space probes. But if we think about communication between Earth and some Kuiper belt probe, that requires powerful laser systems. The laser system can also deliver energy to high-flying drones and other systems. 

Another startup plans to begin space mining. Asteroids are full of minerals that the world requires. Asteroid mining is one of the solutions to the problem. That the lack of certain metals causes. Astroid mining can happen in many ways. One version is to use the shuttle pair. The cargo shuttle finds the wanted asteroid. And then the small robot starts to cut the asteroid into pieces and load those pieces into the shuttle's cargo bay. 

The robot might look like some kind of fly, or crab. It cuts the asteroid using laser cutters. The other version is that the small shuttle packs an asteroid in the mylar bag. And then pulls it into the Earth's orbiter. At orbital trajectory, robots will put it into pieces. That bag might look like a giant parachute or trawl. The shuttle pulls that trawl over the asteroid and closes it inside it. Another version is the giant plastic bag that collects space dust and carries it to the ground. 

Sometimes some people resist asteroid mining because that destroys their scientific values. The fact is that all asteroids have no scientific value. And there can be rules that the asteroids with some special values must not used. That means that asteroids like Vesta and Ceres might not allowed to mine. There are lots of smaller asteroids that are easier to put into a bag. 

The problem is that the space lasers and asteroids allow to creation of new types of weapons. The space laser can destroy other satellites. Same way there is a possibility that the space laser will be aimed at the city areas. The power satellites are also tools that can control the entire planet. 

The asteroid mining has similar problems. If calculations go wrong and about 100X100 meters asteroid drops into the Earth's atmosphere that thing can cause large destruction. When we think about the use of asteroids as weapons, we might think that the system is based on natural asteroids. There is the possibility of launching about 100kg of concrete projectile into space. There that system can use the moon as a gravity sling and then it can travel to Earth. If that object hits to city area, it causes massive destruction. 

https://www.freethink.com/space/asteroid-mining-astroforge

https://www.freethink.com/series/the-freethink-interview?media_id=0U1U1gf2

 https://www.freethink.com/space/space-based-solar-power-aetherflux

https://en.wikipedia.org/wiki/TAU_(spacecraft)

https://www.youtube.com/watch?v=w5SBF48WqV4

AI and social behavior.

Social situations are hard for large language models, LLMs.

When AI tries to mimic humans sometimes it fails. The AI can seem attractive and impressive, but when we ask something there it's not programmed, or trained. The answer is "My algorithms don't allow me to answer". The reason for that answer is that the AI goes outside its box. Answers that the language model gives are programmed in it. If the virtual character is controlled by the human there is no this kind of problem. The AI and humans are the ultimate team. The real-life social situations are a little bit too complicated for the AI. 

The term " to teach, or educate the large language model, LLM" means programming. In that process, programmers teach the AI to give certain types of answers. If the answer is not in a database, the AI asks the programmer or the user how it should answer. The problem with AI and communication with people is that our communication is more complicated than just words. 

Things like sarcasm, the use of voice, and other things like body language are also connected with communication. Because the AI doesn't understand the way, how we say something can turn the meaning of the words opposite. It can give answers that don't fit in the situation. When somebody falls the cup of coffee to the computer's keyboard, and the boss says "nice". That doesn't mean that the boss likes the situation. 

When we face things like social behavior. We face things that we can predict human behavior in large groups. Just like we can predict the movements of large molecular nebulas. But when we try to predict or calculate reactions of the independent people, we face situations in which there are lots of variables in the individual person that we cannot make the AI that can play the human to people, who know that person.  

Our social, and economic background form our way of thinking. The problem with our backgrounds is that there are many things that we don't ever mention. Education happens in three places, homes, schools, and free time activities. Those things don't always communicate with each other. And there can be lots of important things that nobody else than we know. We cannot build models if we don't know all the variables. That is one of the most interesting answers to important things. 

There are so many variables that the AI should handle that it's hard to make a model of the real person. One of the things that we face is emotions. Sometimes we make things that we hate. We can smile in front of some people. And still hate that person. Hidden emotions and feelings can be very common things in professional communities, where outsiders see only things that actors want them to see. 

If somebody is been our classmate, that doesn't automatically mean that we like that person. Or, that doesn't either mean that we like each other. The AI might think that when we were classmates, we automatically liked each other. Same way if we think of dialects like slang people can say that "we hung around in the same gangs". The LLM can think that the gang is always some kind of Hell's Angels. And that makes the discussions interesting. 

The AI searches the word "gang" from the net. Then it sees articles about motorcycle gangs. After that, the AI thinks that the person "who hung in a gang" was a member of those MCs. Same way, because everybody smiles in class photos, the AI thinks that those persons automatically like each other. 

The memories that classmates bring in our memories can be far different than somebody might want. One thing that makes AI fail when it tries to mimic a real person is that it plays a person, who the opponent hates. When we start discussions with old classmates we might say "Nice to meet you". And inside our mind, we can hope that we might not want to see that person ever again. 

AI can make the virtual characters that the real person controls. In that case, the AI operates along with humans. The system can show if the opponent's voice changes. It can see if the opponent looks away from the camera. And it sees things like sweet and dirty skirts. In those cases, the AI's purpose is to support the real person's work. 

The AI can mark points. That is think as dirt or some other remarkable thing. But the responsibility is in the human user. And if a recruiter sees things like disty skirts in an interview those things can help that person to make the selection. We must realize that virtual characters can operate as tools the secret agents and intelligence services use to get confidential information. In the same way, police and other law enforcement can use virtual characters to cheat criminals to tell about their organization. 

The AI can someday think like humans.

What happens when the AI starts to think like a human? That is an interesting question because nobody knows what the person next to me or you really think. They know what that person tells them about thinking but then we must realize, that nobody knows how, or what the next person thinks. We can do many complicated things. And then we must realize, that we don't even know what we are doing.

The AI-controlled robot can select a screwdriver and tighten the screw. The robot can answer questions like what tool it uses and when and where it must use that tool. The computer doesn't need complicated algorithms for that thing. It requires the trigger. That is the screwdriver image. That image activates the database. There are the needed enlisted replies. Those texts might seem like spoken languages. 

Those things are fixed models that might make the AI seem very humankind. 

And if the AI doesn't have the image and database, it can ask the user to make a new database for the new tool. 

There the user can write the description for the AI. 

In that model. The AI can ask questions from the list there the user gives answers. And that is one way to teach the AI. 

We can cut boards without knowing where they are going. We can drive a car. And we must know anything about those vehicles. 

We don't need to know how algorithms work when we use the internet. 

In the same way, we might see that algorithm or the large language model, LLM seems to think. While it sorts information. That it collects from the net. The AI doesn't do anything without commands. There is the user who gives the AI orders. Or there is a trigger in the AI that launches certain operations. The trigger can be a certain pressure level in the tube. When AI notices the pressure it orders to decrease the temperature in boilers. 

Thinking is a complicated process. We say process. Where the nervous system connects information from the senses with memory as "thinking". If we want to mimic that process in the computers we must make billions of databases and connections. The simple thing to make the "semi-thinking" AI is to make it ask questions. The AI might have the list of questions that it should as the coder or developer. 

Some expert's opinion is this. The future is in a small language model. The small language model, SLM is the segment from the large language model. That segment is the group of abilities that the LLM has. The segment can include things like programming models. 

The advanced AI can ask questions about things. That it should be done if the system operator orders the AI to draw humans.

 LLM can ask questions about the clothes that those people wear. 

It can ask about the colors and styles of their clothes. The AI can have a list of the things. That it should notice when it operates with humans. 

Same way the use of AI is always a thing the coder operates in interaction with the AI. The AI should ask questions about things, that the operator wants it to do. The coder must give descriptions about the thing, that the AI must do. Without interactions and descriptions, the AI is in trouble. If we think that the AI replaces coders and does a good job if it takes the command "make me the CAD program". In that case, we overestimate the AI. And it's capacity. 

When we leave programming to the hands of the AI we forget one thing. The AI might seem very creative. But it always follows certain routes for solutions. Those formulas make those AI-produced programs very vulnerable. If the AI always uses certain paths and always gives the same names for the folders. That means that the attackers find those paths and folders very fast if they want. 

AI can seem like a human. It can answer very complicated questions. Or do we just think about that thing like that? 

Are all the questions that we ask the AI difficult? 

Another thing that can motivate us to use AI is simple. That we are lazy. We have no time to read articles. Or we just need that time for some other things. So we might ask for some terms that we would find any way from Wikipedia from the AI. 

https://www.quantamagazine.org/will-ai-ever-understand-language-like-humans-20250501/

"Pigeon-problem" in memory control.

Computers have two types of memory: read-only ROM and random-access memory, RAM. The last one is known, as work memory, where the system loads programs. Every memory unit has a certain address. Those memory units are like boxes. 

Sometimes, the computer's operations are compared with the pigeon. The pigeon is the data package and the packet is the router. The pigeon problem is how that pigeon finds the right box where it drops that data. 

That flies around the slope where there are lots of boxes. Each of those holes, caves, or boxes, whatever you want to call them has a certain address. The pigeon doesn't know which of the boxes is empty that it can leave its package which is the operation that the computer should run. 

Finding the free memory block or memory address happens randomly. That's why that thing is called random access memory, RAM. We can say that when a pigeon drops the packet into a certain memory unit, it pushes a button and tells that the operation is in that memory address. The pigeon's problem is that there are always reserved boxes. 

The operating system reserves a certain number of those boxes for its own use. The operating system can put the list of addresses that it reserves. And then the computer program knows that those addresses are reserved. That means the fixed solution that is suitable in the cases, where the system runs only one program in time. 

This kind of solution is slow and even dangerous if the computer or its program requires polymorphic memory. Another name for polymorphic memory is the morphing neural network. In the flexible memory handling method. There the system drives multiple operations at the same time. Those algorithms or programs require a certain number of memory addresses. When those groups of networked memory addresses made their duty, they asked for a new mission. If there is no other new mission those memory blocks or boxes can offer assistance to other network groups. 

The morphing memory network base is in those memory addresses. The system interconnects those memory addresses into new subentities. In that kind of morphing system there the memory address can have different roles causing the problem to pigeon. The reserved memory addresses are not fixed. And that means. The system should always update those lists. 

But the answer to the pigeon's problem can be the interactive memory. In that case, the box itself raises a flag or the mark that it's free. When a pigeon comes to give a message or new task it sees immediately which of the boxes is free. So it must not knock on the door and disturb another operation. The problem with fixed memory addresses is that they leave lots of free space in the system when they finish their job. The system can use that free space for another mission that it runs. So the system requires a more dynamic solution than just fixed memory address groups. 

The problem with the large language model, LLM-type solutions is that there are billions of data packages. They try to find the place where they put their data. There is a possibility that two pigeons try to put data in the same box. So there are some details that the pigeons can choose right and empty box. The pigeon might have a skirt that has a certain color like yellow. The yellow pigeon travels to the yellow box. 

The pigeon can also have a flag that it shows to other pigeons. The flag, or screen can show information about which number of the box the pigeon selects. And that tells others that the memory box is full. 

The pigeon can fly and systematically search if there is an empty point or base. That kind of thing means that sooner or later the pigeon finds the free address. There is the possibility of increasing memory. That increases the number of memory addresses. And that increases the possibility that the pigeon finds a free memory address. 

The pigeon or the computer program can always ask the free space from the operating system. That means the program comes to the terminal and then it waits for the operating system to tell where are the free boxes. The problem is that this way of operating is slow. The system can also use a method. That is sometimes called "hovering algorithms". In that model, the algorithm or pigeon that leaves the box tells other algorithms or pigeons which box is free. 

The thing that can help this mission is that pigeons have a certain color skirt. The pigeon with the yellow skirt goes to yellow boxes that are sorted in one entirety. When it leaves the package or operand in the box. It leaves a flag in that place. When another yellow-skirt pigeon comes to that box it can simply pull the flag off and then drop it to the controller that takes the operand. 

But in modern computing, there can be billions of units in those boxes. The system can drive even billions of algorithms at the same time. The new types of systems require that the computer algorithms can tell straight to another algorithm. That is place is free without traveling through the operating system. This makes the system more flexible. 

So the pigeon has a very big problem. If it disturbs the operation that runs in a certain address, that thing can cause a fatal disturbance in the operations. The dynamic solution requires that the memory controller can resort to memory in a new way if there is free space.