Artificial intelligence continues to develop. And more and more often we hear from different companies about strong AI. But what is it and should we expect its creation in the near future?
Content
Introduction
Strong or general AI (AGI) is AI that can navigate changing conditions, model and predict the development of a situation. If the situation goes beyond standard algorithms, then he must independently find a solution. For example, solve the problem “go to university.” Or learn the rules of the game of checkers, and start playing checkers instead of chess.
What qualities should such an AI have?
Thinking is the use of methods such as deduction, induction, association, etc., which are aimed at isolating facts from information and presenting them (preserving them). This will allow you to more accurately solve problems under conditions of uncertainty.
Memory - the use of different types of memory (short-term, long-term). Can be used to solve problems based on previous experience. Even if you try to communicate with ChatGPT 4, you will see that the algorithm has a small short-term memory, and after 10-15 messages it forgets where it all began. In general, in my opinion, the issue of memory and the “massiveness” of AI models will become 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 is still only in its infancy. And the deeper the planning, especially in conditions of uncertainty, the more capacity is needed. After all, it’s one thing to plan a chess game 3-6 steps deep, where all the rules are clear, and another thing in a situation of uncertainty.
Learning is imitation of the actions of another object and learning through experimentation. Currently, AI learns from large amounts of data, but it does not model or conduct experiments itself. Although, we do not fully understand how Chat GPT works. And this is one of the main problems. But learning requires the formation of long-term memories and complex relationships. And this, as we understand, is a problem for AI.
Nowadays no one has such strong AI. We are only at the stage of transition from weak to intermediate. Yes, OpenAI's ChatGPT, Google's LaMDA and other large language models (LLMs) can generate text/illustration/video through query analysis and big data processing. But they only broadcast what their creators taught. They look for the most likely combinations of words and sentences, words and images, and try to imitate human activity.
But don’t be under any illusions, this is just mathematics and statistics. And there is a lot of “marriage” and “hallucinations” in their answers. They are not yet ready for real interaction with the world.
Let's give a simple example.
As part of our work, especially in the field of project management, we decided to write a post on why Agile is not a one-size-fits-all approach to project management. Why, when implemented correctly, any company and team should move closer to hybrid or waterfall approaches.
We decided to try ChatGPT, GigaChat, YaGPT and AI Search.
To do this, we formulated a simple request: “Write a post about why Agile is wrong to apply endlessly. Emphasize that conducting retrospectives should move away from a situation of uncertainty to a situation of complex ordered or simple ordered systems.”
ChatGPT4 Answer : “Agile is an approach to software development that is based on iterative and incremental product creation, collaboration among self-organizing teams, and continuous process improvement. One of the key principles of Agile is holding regular retrospectives, that is, sessions in which the team analyzes its work, identifies problems and finds ways to solve them.
Retrospectives are a powerful tool for improving efficiency and quality of work, but only if they are done correctly. Unfortunately, many teams make the mistake of applying Agile indefinitely, without taking into account the context and changes in the environment. This results in retrospectives becoming useless, boring and demotivating.
To avoid this, you need to understand that Agile is not a one-size-fits-all 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 in the form of a matrix, where the degree of uncertainty is plotted horizontally, and the degree of complexity is plotted vertically.
!Zone matrix (chatbot edition)
Each zone requires a different approach to retrospectives and improvements. In a chaotic 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 complexity zone, a team works with unclear and dynamic requirements that require creativity and experimentation. In this case, retrospectives should be long, infrequent, 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 moderate, 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 needed.
However, it is important to understand that zones are not static and can change over time. Therefore, the team must constantly evaluate its context and adapt its approach to retrospectives. The purpose of retrospectives is not to apply Agile indefinitely, but to move the team from a situation of uncertainty to a situation of complex ordered or simple ordered systems. Only then will retrospectives bring benefit and pleasure to the team and the project.”
GigaChat Answer : “Agile is a software development methodology that has become very popular in recent years. It offers a flexible approach to project management that allows you to quickly respond to changes and adapt to new requirements. However, endless application of Agile can lead to problems.
One of the main reasons why Agile is endlessly misapplied is that it can lead to situations of uncertainty. In Agile projects, we often encounter changes in requirements and priorities, which can cause uncertainty about what to do next.
Conducting retrospectives is one way to combat this uncertainty. Retrospectives help the team understand what went right and what needs to be improved in the future. They also help the team move from a situation of uncertainty to a situation of complex ordered or simple ordered systems.
However, if we endlessly practice Agile without conducting retrospectives, then we may be left in a situation of uncertainty. This can lead to the team spending a lot of time fixing bugs and adapting to new requirements rather than creating new functionality.
In addition, endless use of Agile can result in a team constantly working on small tasks and not having time to create large functional units. This may result in the project progressing slowly and not meeting the customer's expectations.
Overall, Agile is a great approach to project management, but its endless application can lead to problems. Conducting retrospectives helps the team move from a situation of uncertainty to a situation of complex ordered or simple ordered systems and improve the quality of the team’s work.”
YaGPT2's answer : “Endless application of Agile can lead to poor decisions and problems with product quality. It is important to understand that Agile is not a one-size-fits-all solution for all projects and teams, and its continued use can lead to a loss of efficiency and productivity.
One of the key principles of Agile is conducting retrospective sessions, which help the team reflect on its past successes and failures. However, if these retrospectives are conducted continuously, it can distract the team from solving real problems and challenges.
Instead of helping the team develop and become more effective, constant retrospectives can lead to a situation where the team constantly discusses the same issues and fails to move forward. This can create a situation of uncertainty where the team does not know what to do next, or, conversely, it can lead to the creation of complex and disorganized systems that only complicate the work process.
To avoid this, it is important to determine when and how often to conduct retrospective sessions. They should be focused on analyzing specific problems and opportunities for improvement, rather than discussing the same issues over and over again.
It is also important to understand that every team and project is unique, and what works for one team may not work for another.
AI Search response: “Hello friends! Today I would like to share my thoughts on why Agile, despite its popularity and effectiveness, is often misapplied. I will especially focus on why retrospectives are an important tool for moving from situations of uncertainty to orderly or complex systems.
Agile is a philosophy and approach to software development that changes the traditional approach to project management, making it more flexible and adaptive to change. However, despite the clear benefits, Agile can be applied incorrectly, and as a result, its full potential is not realized.
One of the main problems with the misuse of Agile is the lack of regular retrospectives. A retrospective, as an expert assessment of past work, helps teams take changes, shortcomings, and improvements into account when planning future iterations. This allows you to refine your work process and increase efficiency. However, some teams often ignore conducting retrospectives or conduct them in an unsystematic manner.
Without regular retrospectives, teams miss the opportunity to see their problems, confirm or change their vision, and fine-tune development steps. As a result, teams may become stuck in uncertainty and unable to move forward.
One of the main principles of Agile is to ensure rapid response to change and continuous improvement of the work process. If retrospectives are not conducted regularly, teams are not sufficiently aware of their mistakes and omissions, making it difficult to make progress toward being streamlined and efficient.
In addition, holding regular retrospectives helps teams move from uncertainty to complex or simple orderly systems. Starting with simple, orderly systems where rules and processes are easily defined and controlled, teams can gradually move to more complex systems where the rules are more ambiguous and require more adaptation.
So, retrospectives are an important part of the Agile process and help teams move from uncertainty to orderliness and complex systems. They allow teams to recognize their mistakes and shortcomings and see areas for improvement. Don't neglect regular retrospectives so your team can maximize Agile's potential and achieve high performance.
I wish you success in applying Agile and constantly improving your processes! Be flexible, persistent and attentive."
Our post: “Why Agile is not forever and not a panacea?”
In recent years, everyone has declared that they are Agile.
Let's be honest, most of them hide banal chaos under this. Agile is not about lack of planning. In addition, he is very sensitive to the implementation of certain rules and the stability of the team, its motivation. That is, you are not Agile in this case.
But the very use of Scrum, Kanban or other approaches should lead to the elimination of the need to implement Agile projects.
But why?
Let's remember why Agile appeared? To work and implement projects in situations of high uncertainty.
There is even a tool - the Kenevin model. It helps you understand what situation/system you are in and what approach you need to take and what to focus on.
At the same time, in ordered systems (simple or complex), Agile, on the contrary, is contraindicated. It increases the costs of achieving results. What is called reinventing the wheel every time. That is, Agile is about efficiency, when you need to do something, I don’t know what. But he is not about efficiency.
Now let's take a look back. All approaches within Agile require regular retrospectives, analysis of one’s work, and interactions with clients/customers/partners. That is, the very logic of these tools is to get away from an uncertain situation and learn to predict them, to become more effective.
If you constantly (once every six months - year) change jobs, or constantly launch new products, rather than replicating certain solutions (which is strange for business), then yes, you need to be Agile.
But if you have a segment, and you have developed with experience and expertise “standard” approaches / products that need to be adjusted and adapted in a small part, then it’s too early or you should leave Agile. You must arrive at a streamlined system where cascading or hybrid approaches are needed. These retrospectives should lead you to an understanding of what customers want in 90% of cases and how the organization / colleagues work.
As a result, if you are Agile on an ongoing basis and everywhere, and not for a period of restructuring / launch / adaptation, then this may indicate:
non-compliance with Agile tools;
you haven’t found your product and your niche or yourself, you haven’t developed the necessary expertise;
you have a unique product/project each time (which should be reflected in the high price of your services);
the organization is “sick” from the inside, and in this way you mask high turnover, lack of work on processes, etc.
What do you think about this?
The quality of neural network responses is more like a very weak student who simply selects similar words, but does not understand what he is talking about. Could we get a similar post from a neural network? So that she conveys our meanings? Yes, of course, but we wrote our post in 25 minutes, and how long would it take us to try to achieve such a result from AI?
Therefore, according to various estimates, strong AI is still 5-20 years away.
Also, various researchers highlight super-strong AI (ASI). This is AI that can not only solve complex problems, but also do it almost instantly. If there are already thousands of weak AIs, for each task, there will be dozens of strong AIs (most likely there will be a division according to directions), then there will be 1 superstrong AI per state or for the entire planet.
Limits to Strong AI
But, to be honest, we are cautious about the rapid emergence of strong or super-strong AI.
Firstly, this is a very costly and complex task in terms of regulatory restrictions . The era of uncontrolled AI development is ending. More and more restrictions will be placed on it.
In a risk-based approach, the strong and super strong would be at the top level of risk. This means there will be restrictive restrictions. Already, AI developers, including ChatGPT, are facing copyright infringement lawsuits. And this is before the introduction of strict rules.
Secondly, this is a difficult task from a technical point of view, and strong AI will also be very vulnerable.
Now, in the mid-2020s, creating and training strong AI requires enormous computing power and complex AI models. We will have to exponentially increase the number of neurons and build connections between them. If human neurons can be in several states, and activation can occur “in different ways” (may biologists forgive us for such simplifications), then machine AI cannot do this. That is, conditionally, a machine’s 80-100 billion neurons are not equal to a person’s 80-100 billion. The machine will need more neurons. The same GPT4 is estimated at 100 trillion parameters (relatively neurons), and it is still inferior to a person.
This all comes down to several factors.
The first is that increasing complexity always leads to reliability problems; the number of points of failure increases. Such models are difficult to both create and maintain from degradation over time, during operation. AI models need to be constantly “maintained.” If this is not done, then strong AI will become like an old man with dementia, and the “life span” will be very short. Imagine what a strong AI with its capabilities can do, but which at the same time will suffer from Alzheimer's? Even for current AI solutions, this is a pressing problem.
Let's give a couple of simple examples from life. The first is based on human muscles. When we just start working out in the gym and get involved in strength training and bodybuilding, progress goes quickly, but the further we go, the lower the efficiency and the increase in results. You need more and more resources not only to progress, but also to simply stay in shape. Plus, the increase in strength comes from the thickness of the muscle section, but the mass increases from volume. As a result, at a certain point the muscle will become so heavy that it will not be able to move itself, and may even damage itself.
A similar example can be given in the engineering field. For example, Formula 1 racing. So, a lag of 1 second can be eliminated if you invest 1 million and 1 year. But to win back the decisive 0.2 seconds it may take 10 million and 2 years of work. And the fundamental limitations of the car's design may force us to reconsider the entire concept of a racing car.
If you look at ordinary cars, everything is the same. Modern cars are more expensive to create and maintain, and without special equipment it is impossible to even change a light bulb. If you take modern hypercars, then after each trip, entire teams of technicians are required for maintenance.
If we return to the issue of degradation, then in addition to technology there will also be the influence of people. Any AI, especially at an early stage, will learn based on feedback from people (their satisfaction, initial requests and tasks). An example here is the same ChatGPT4. For example, ChatGPT uses user requests to further train its model. And at the end of 2023, articles began to appear that the AI model had become “lazier”. The chatbot either refuses to answer questions, interrupts the conversation, or simply responds with excerpts from search engines and other sites. Moreover, by mid-2024 this had already become the norm, when the model simply cited excerpts from Wikipedia.
And one of the possible reasons is that users themselves began to ask more and more simple and primitive queries. After all, LLMs (large language models) do not come up with anything new, these models try to understand what you want to hear from them and adapt to it. She is looking for the maximum efficiency of the labor input-result link. This is called maximizing the function. Just math and statistics.
In general, here is an example of an anecdote:
"I type at 300 characters per minute!
- Really?
“Yeah, it’s just complete bullshit.”
The second factor is the amount of data. Yes, we can scale up current models by XXXX times. But the same ChatGPT5 prototype already in 2024 lacks training data. A strong AI will require a huge amount of high-quality data for initial training.
The third factor is that the AI model will be tied to its “base”. It will require huge and complex data centers to operate, with powerful energy sources and high-quality cooling. Thus, according to some estimates, 5 - 50 requests for ChatGPT 4 require up to 0.5 liters of water for cooling. For more powerful models, this figure will also increase, although this may be offset by the technological development of AI accelerators.
And no matter what the Internet bandwidth is, the “brain” will still be in these centers; there will be no distributed data processing networks. Firstly, distributed computing still loses performance and efficiency. In addition, a distributed network cannot guarantee the operation of computing power all the time. Secondly, it is vulnerable to attacks on communication channels and the same distributed infrastructure. Imagine that suddenly 10% of the neurons in your brain simply turned off (blocked communication channels or simply turned off), and the rest are stupid and work at half capacity. As a result, we again have the risk of a strong AI that either forgets who it is, where it is, or something else.
And if everything comes to the point that strong AI will need a body to interact with the world, then it will be even more difficult to implement this. Then the AI model will be limited, otherwise how will it all be powered and cooled? Where to get data processing power? That is, it will be a limited AI with a constant connection to the main center via wireless communication. And this is vulnerability again. Modern communication channels provide higher speeds, but this affects the range and penetration. In addition, such channels are easier to suppress using electronic warfare. That is, we get an increase in the load on the communication infrastructure and an increase in risks.
Here you can, of course, object. For example, the fact that you can take a pre-trained model and make it local. In much the same way as we propose to deploy local AI models with “additional training” to the subject area. Yes, in this form all this can work on one server. But such AI will be very limited, it will be an “intermediate” AI, and it will be “dumb” under conditions of uncertainty and will still need energy. That is, this is not a story about the creation of humanoid superbeings. This will be a large number of strong AI, but expensive and with limited capabilities, which is not very interesting to the market.
All of these factors will lead to an exponential increase in the complexity and costs of creating, developing and maintaining strong AI. Research and prototyping costs could reach between US$0.5 and US$1 trillion.
At the same time, weak models with narrow specialization will remain more “free” and easier to create.
All this leads to questions about the economic feasibility of investing in this area. Moreover, taking into account two key trends in the development of AI :
creation of cheap and simple local models for solving specialized problems;
creating AI orchestrators that will decompose a request into several local tasks and then redistribute this between different local models.
As a result, we have a simpler and cheaper solution to work problems than creating strong AI.
Of course, we are taking neuromorphic and quantum systems out of the equation, but we have covered this topic here . And, of course, there may be errors in our individual figures and arguments, but overall we are convinced that strong AI is not a matter of the near future.
Summary
As a result, to summarize, strong AI has several fundamental problems.
1. Exponential growth in the complexity of development and degradation of complex models.
2. Lack of training data.
3. Cost of creation and operation.
4. Attachment to data centers and demands on computing resources.
5. Low efficiency of current models compared to the human brain.
It is the overcoming of these problems that will determine the further vector of development of the entire technology: either strong AI will emerge, or we will go into the development of weak AI and AI orchestrators who will coordinate the work of dozens of weak models.
But now strong AI is not about ESG, ecology or commercial success. Its creation can only be within the framework of strategic and national projects that will be financed by the state.