Artificial intelligence. Freefall

What is artificial intelligence?

To begin with, there is no such definition. Let’s ask the AI itself, or rather several large language models (LLM – large language model). Below are some answers with the original spelling.

GPT4: “Artificial intelligence (AI) is the ability of artificial intelligent systems to perform creative functions that are traditionally considered the prerogative of humans. It is related to the science and technology of creating intelligent machines, especially intelligent computer programs. It’s important to note that AI isn’t necessarily limited to biologically plausible methods.

Existing intelligent systems have rather narrow application areas. For example, programs that can beat a human at chess usually can’t answer questions.

In general, artificial intelligence allows computer systems to mimic human cognitive abilities, such as learning and problem solving.

This is an important area in modern computer science and data science, and its application spans many areas, from finance and medicine to entertainment and transportation”.”

YandexGPT 2: “Artificial Intelligence – AI) is a set of computer science, mathematics, biology, and psychology techniques that develop systems that can perform tasks that usually require human intelligence. Their main task is to model the human mind.”

YandexGPT 3 Pro: “Artificial intelligence (AI) is the ability of intelligent systems to perform creative functions that are traditionally considered the prerogative of humans. It is the science and technology of creating intelligent machines, especially computer programs that can perform tasks that require human intelligence, such as solving complex problems, learning from experience, understanding natural language, and being able to adapt to new situations. AI is related to the use of computers to understand human intelligence, but is not limited to biologically plausible methods.”

GigaChat: “Artificial intelligence (AI) is the ability of intelligent systems to perform creative functions that are traditionally considered the prerogative of humans. It is the science and technology of creating intelligent machines, especially intelligent computer programs. Artificial intelligence is related to the similar task of using computers to understand human intelligence, but is not necessarily limited to biologically plausible methods. Today’s smart systems have rather narrow areas of application. Artificial intelligence was founded as an academic discipline in 1956 and has generated several waves of optimism in the decades since.”

But as a human, the simplest and most understandable definition is closer: AI is any mathematical method that allows you to imitate human or other intelligence from natural nature.

In other words, AI is a huge number of solutions, including primitive mathematical algorithms and expert systems based on rules.

And although this direction was born somewhere in the 50s of the 20th centuries, we are primarily interested in what we understand by this today, in the early 2020s. And tut has three main areas.

1. Neural networks – mathematical models created in the likeness of neural connections of the brain of living beings. Actually, the human brain is a super-complex neural network, the key feature of which is that our neurons are not limited to “on / off” states, but have many more parameters that cannot yet be digitized and fully applied.

2. Machine learning – ML) – statistical methods that allow computers to improve the quality of the task performed with the accumulation of experience and in the process of retraining. This direction has been known since the 1980s.

3. Deep learning (DL) is not only machine learning with the help of a person who says what is true and what is not (as we often raise children, this is called reinforcement learning), but also self-learning of systems (learning without reinforcement, without human participation). This is the simultaneous use of various training and data analysis techniques. This direction has been developing since the 2010s and is considered the most promising for solving creative tasks, and those tasks where the person himself does not understand clear relationships. But here we can’t predict at all what conclusions and results the neural network will come to. Here you can manipulate what data we “feed” to the AI model at the input.

How are AI models trained?

Currently, most AI-models are trained with reinforcement: a person sets input information, the neural network returns an answer, and then the person tells it whether it answered correctly or not. And so on, time after time.

Similarly, the so-called “CAPTCHA” (CAPTCHA, Completely Automated Public Turing test to tell Computers and Humans Apart) work, that is, graphical security tests on websites that calculate who the user is: a person or a computer. This is when, for example, you are shown a picture divided into parts and asked to specify the areas where bicycles are depicted. Or they ask you to enter numbers or letters displayed in an intricate way on the generated image. In addition to the main task (the Turing test), this data is then used to train AI.

At the same time, there is also unsupervised learning, in which the system learns without human feedback. These are the most complex projects, but they also allow you to solve the most complex and creative tasks.

General features of current AI-based solutions

Fundamentally, all AI-based solutions at the current level of development have common problems.

– Amount of training data.

Neural networks need huge amounts of high-quality and marked-up data for training. If a human can learn to distinguish dogs from cats in a couple of examples, then AI needs thousands.

– Dependence on data quality.

Any inaccuracies in the source data strongly affect the final result.

– The ethical component.

There is no ethics for AI. Only math and problem completion. As a result, complex ethical issues arise. For example, why should I knock down autopilot in a desperate situation: an adult, a child, or a pensioner? There are countless similar disputes. For artificial intelligence, there is neither good, nor evil, just like the concept of “common sense”.

– Neural networks cannot evaluate data for reality and logic, and they are also prone to generating poor-quality content and AI hallucinations.

Neural networks simply collect data and do not analyze facts or their connectedness. They make a large number of mistakes, which leads to two problems.

The first is the degradation of search engines. AI created so much low-quality content that search engines (Google and others) began to degrade. Just because there is more low-quality content, it dominates. This is especially helpful for SEO-optimizers of sites that simply outline popular queries for promotion.

The second is the degradation of AI models. Generative models also use the Internet for “retraining”. As a result, people, using AI and not checking for it, fill the Internet with low-quality content themselves. And the AI starts using it. The result is a vicious circle that leads to more and more problems.

An article on this topic is also available by using the QR code and hyperlink.

The AI feedback loop: Researchers warn of “model collapse’ as AI trains on AI-generated content

Realizing the problem of generating the largest amount of disinformation content by AI, Google conducted a study on this topic. Scientists analyzed about two hundred media articles (from January 2023 to March 2024) about cases when artificial intelligence was used for other purposes. According to the results, most often AI is used to generate fake images of people and false evidence of something.

– The quality of “teachers”.

Almost all neural networks are taught by people: they form requests and give feedback. And there are many limitations here: who teaches you what, based on what data, and for what purpose?

– People’s readiness.

We should expect huge resistance from people whose work will be taken away by neural networks.

– Fear of the unknown.

Sooner or later, neural networks will become smarter than us. And people are afraid of this, which means that they will slow down development and impose numerous restrictions.

– Unpredictability.

Sometimes everything goes as planned, and sometimes (even if the neural network copes well with its task) even the creators struggle to understand how the algorithms work. The lack of predictability makes it extremely difficult to eliminate and correct errors in neural network algorithms. We are only learning to understand what we have created ourselves.

– Restriction by type of activity.

All AI for mid-2024 is weak (we’ll discuss this term in the next chapter). Currently, AI algorithms are good for performing targeted tasks, but they do not generalize their knowledge well. Unlike a human, an AI trained to play chess will not be able to play another similar game, such as checkers. In addition, even deep learning does a poor job of processing data that deviates from its training examples. To effectively use the same ChatGPT, you must initially be an expert in the industry and formulate a conscious and clear request.

– Costs of creation and operation.

It takes a lot of money to create neural networks. According to a report from Guosheng Securities, the cost of training a relatively primitive LLM GPT-3 LLM was about $ 1.4 million. For GPT-4, the amounts already go into the tens of millions of dollars.

If we take ChatGPT3 as an example, then only to process all requests from users, more 30than 30,000 NVIDIA A100 GPUs were needed. Electricity cost about $ 50,000 a day. It requires a team and resources (money, equipment) to ensure their “life”. You also need to take into account the cost of maintenance engineers.

Again, these are common drawbacks for all AI solutions. Later on, we will return to this topic several times and discuss these shortcomings in more practical examples.

Now let’s talk about three concepts-weak, strong, and super-strong AI.

Weak AI

All that we are seeing now is a weak AI (ANI, Narrow AI). It can solve highly specialized tasks that it was originally designed for. For example, it can distinguish a dog from a cat, play chess, analyze videos and improve the quality of video / audio, advise on a subject area, and so on. But, for example, the strongest weakest AI for playing chess is absolutely useless for playing checkers. And AI for project management consulting is absolutely useless for planning equipment maintenance.

Example of how AI works in pattern recognition

Strong and super – strong AI-what is it?

If the definition, of what AI, is confusing, then the term “strong AI” or “general AI” is even more complicated. Let’s start by using LLM services again. To do this, I made a request: “Give a transcript of the term AGI and its brief description.” Below I will give the answers of several LLMs in the original spelling.

YandexGPT 2: “AGI (Artificial General Intelligence) is a general-purpose artificial intelligence that can solve a variety of tasks typical of the human mind. AGI has the flexibility and adaptability of the human brain and can learn independently”.”

YandexGPT 3 Pro: “AGI (Artificial General Intelligence) is a concept for creating artificial intelligence that will be able to solve a wide range of tasks at the level of human intelligence. This is the next step in the development of artificial intelligence after highly specialized systems, such as speech recognition or computer vision systems.”

YandexGPT 4 Pro RC: “AGI (Artificial General Intelligence) is artificial general intelligence, which is a hypothetical form of AI that can understand and solve problems at the same level as humans. AGI can learn, adapt, and make decisions in various fields, such as science, art, and social interactions.

Unlike specialized AI systems that are designed to perform specific tasks, AGI strives for versatility and the ability to solve a wide range of problems. However, creating an AGI remains a complex task that requires significant research and development in machine learning, neural networks, and other technologies.”

Giga Chat: “AGI (Artificial General Intelligence) – general artificial intelligence. This is the concept of creating an intelligent machine that can solve any intellectual tasks in the same way as a human does, and perhaps better than him.”

ChatGPT4: “AGI stands for Artificial General Intelligence (Automated General Intelligence), which is a type of artificial intelligence that can perform a wide range of intelligent tasks at the human level or even surpass it. This is in contrast to highly specialized AI, which is designed for specific tasks. Building an AGI is one of the main goals of AI research and companies like Open AI. The time to reach AGI is the subject of discussions withered researchers and experts.”

However, the original query also has a significant impact on the result.

I will also give you some examples of the Q & A bundle for LLM YandexGPT 3 Pro.

Question 1: “Hi. What is strong AI or AGI?”

Answer 1: “Hi! Strong AI, also known as AGI (Artificial General Intelligence), is the concept of creating artificial intelligence that can perform any intellectual tasks available to humans, but not necessarily in the same way as a human does. Unlike weak AI, which is designed to perform specific tasks, strong AI tends to mimic human intelligence in general, including analytical thinking and learning abilities”.”

Question 2: “AGI decryption”.

Answer 2: “AGI – Artificial General Intelligence) is the concept of creating artificial intelligence that will be able to solve a wide range of tasks at the level of human intelligence”.”

Question 3: “What is the explanation of the term AGI and its description?”

Answer 3: “AGI (Artificial General Intelligence) is a strong general – level artificial intelligence capable of thinking and acting like a human. They can make independent conclusions, learn, and possibly become self-aware at some point in time. Unlike Narrow AI, AGI is the entire brain capable of solving a set of tasks”.”

As you can see, a lot depends on the request. And this is one of the key limitations of generative AI (GII), but we’ll talk about it in the next chapter.

Now I will give the definition that, in my opinion, most accurately defines its essence.

Strong or general AI (AGI) is an AI that can navigate changing conditions, model and predict the development of a situation. And if the situation goes beyond the standard algorithms, then find a solution yourself. For example, to solve the problem “go to university” and learn the rules of the game of checkers, and instead of chess, start playing checkers.

What qualities should such an AI have?

Thinking is the use of methods such as deduction, induction, association, etc., which are aimed at extracting facts from information, their presentation (preservation). This will make it possible to solve problems more accurately in conditions of uncertainty

Memory is the use of various types of memory (short-term, long-term). That is, the tasks should be solved taking into account the accumulated experience. Now, if you chat with ChatGPT 4, you will see that the algorithm has a small short-term memory and after a while forgets where it all started. In general, in my opinion, the issue of memory and the “massiveness” of AI models will be a key limitation in the development of AI. More on this below.

Planning – tactical and strategic. Yes, there are already studies that claim that AI can plan its actions and even deceive a person to achieve its goals. But now it’s still only in its infancy. The deeper the planning goes, especially in conditions of uncertainty, the more capacity is needed. After all, it’s one thing to plan a game of chess 3—6 steps deep, where all the rules are clear, and another in a situation of uncertainty.

Training – simulating the actions of another object and learning through experimentation. Now he is learning from large amounts of data, but he does not model or conduct experiments himself. Although we do not fully understand how the same Chat GPT works, and this is one of the main problems, learning requires the formation of long-term memory and complex relationships. And this, as we understand it, is a problem for AI.

And now no one has such a strong AI right now. And the statement about the imminent (in the2024—2028horizon) appearance of a strong AI, in my opinion, is erroneous or speculative. Although maybe, I have too limited knowledge…

Yes, ChatGPT from Open AI and other LLMs can generate text / illustrations / videos through query analysis and big data processing. But they only broadcast what the creators gave them for training. They search for the most appropriate combinations of words and sentences, words and images for the query, and in general, associative combinations. But don’t be under any illusions, this is just mathematics and statistics. And in their answers, there are a lot of “marriage” and “hallucinations”. They are not yet ready for real interaction with the world.

Here is a simple LLM-based example.

I – come from a project management background, and as part of my work, I decided to write an article about why Agile is not a universal approach to project management, and why, if applied correctly, any company and team should move closer to hybrid or cascading approaches.

My assistants in writing text were all the same ChatGPT, GigaChat, YandexGPT and AI Search.

I formulated a simple request: “Write a post about why Agile is infinitely wrong to apply. Emphasize that conducting retrospectives should lead away from the situation of uncertainty to the situation of complex ordered or simple ordered systems.”

ChatGPT4’s answer: “Agile is an approach to software development that is based on iterative and incremental product development, collaboration between self-organizing teams, and continuous process improvement. One of the key principles of Agile is to conduct regular retrospectives, i.e., sessions where the team analyzes its work, identifies problems and finds ways to solve them.

Retrospectives are a powerful tool for improving the efficiency and quality of work, but only if they are conducted correctly. Unfortunately, many teams make the mistake of applying Agile indefinitely, without taking into account the context and changes in the environment. This leads to the fact that retrospectives become useless, boring and demotivating.

To avoid this, you need to understand that Agile is not a universal solution for every situation. Depending on the degree of uncertainty and complexity of the project, the team may be in one of four zones: chaos, complexity, order, or simplicity. These zones can be represented as a matrix, where the degree of uncertainty is plotted horizontally, and the —degree of complexity is plotted vertically.

!Matrix of zones (the chatbot’s editorial staff, which is exactly how he designated this area)

Each zone requires a different approach to retrospectives and improvements. In the chaos zone, the team faces unpredictable and extreme situations that require rapid response and adaptation. In this case, retrospectives should be short, frequent, and focused on survival and stabilization. In the difficulty zone, the team works with unclear and dynamic requirements that require creativity and experimentation. In this case, retrospectives should be long, sparse, and focused on learning and innovation. In the order zone, the team works with clear and stable requirements that require discipline and optimization. In this case, retrospectives should be average, regular, and focused on efficiency and quality. In the simplicity zone, the team works with obvious and simple requirements that require minimal effort and attention. In this case, retrospectives can be omitted or conducted as necessary.