Kitabı oku: «Artificial intelligence. Freefall», sayfa 2
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 its obvious advantages, Agile can be applied incorrectly, and as a result, its potential is not fully revealed.
One of the main problems of using Agile incorrectly is the lack of regular retrospectives. A retrospective, as an expert assessment of past work, helps teams take into account the changes, shortcomings, and improvements that have occurred when planning future iterations. This allows you to hone the work process and increase efficiency. However, some teams often ignore conducting retrospectives or conduct them unsystematically.
By not conducting regular retrospectives, teams miss opportunities to see their issues, validate or change their vision, and customize development milestones. As a result, teams may be stuck in a situation of uncertainty and unable to move forward.
One of the main principles of Agile is to ensure a fast response to changes and continuous improvement of the work process. If retrospectives are not conducted regularly, teams are not sufficiently aware of their mistakes and omissions, which makes it difficult to progress towards achieving order and efficiency.
In addition, conducting regular retrospectives helps teams move from a situation of uncertainty to complex or simple ordered systems. Starting with simple, streamlined systems where rules and processes are easily defined and controlled, teams can gradually move to more complex systems where rules are more ambiguous and require more adaptation.
So, flashbacks are an important part of the Agile process and help teams move from uncertainty to order and complex systems. They allow teams to recognize their mistakes and shortcomings, as well as see areas for improvement. Do not neglect regular retrospectives so that your team can maximize the potential of Agile and achieve high performance.
I wish you successful work in applying Agile and continuous improvement of your processes! Be flexible and persistent and considerate”.”
Well, and my post: “Why is Agile not eternal and not a panacea?”
“In recent years, everyone declares that they are Agile.
Let’s be honest: most people hide banal chaos under this. After all, Agile is not about the lack of planning. In addition, he is very sensitive to the implementation of certain rules and to the stability of the team, its motivation. If you don’t meet these conditions, don’t use Agile in the future.
And using the same Scrum, Kanban or other approaches should lead to the absence of the need to implement Agile projects.
But why?
Let’s remember that Agile was originally designed to work and implement projects in a highly uncertain environment.
There is even a special tool – model Kenevin, that helps you understand what situation you are in and what approach you need to choose, what to focus on. So, in ordered systems (simple or complex situations), Agile, on the contrary, is contraindicated, because it reduces the cost of achieving results. That is, Agile is effective in cases where you need to do – I don’t know what”. But it’s not about efficiency.
Now let’s take a look at the retrospective. All approaches within Agile involve regular retrospectives, analysis of their work, and interaction with the client / customer / partner. In other words, the very logic of these tools is to get away from an uncertain situation and learn how to predict them and become more effective.
If you constantly (every six-months or a year) change jobs or constantly launch new products, and do not replicate certain solutions (which is strange for a business), then yes, you need to be Agile.
But if you have a segment in the AC, and you have worked with experience and expertise on typical approaches / products that need to be adjusted and adapted in a small part, then sooner or later you must move away from Agile and come to an ordered system where cascading or hybrid approaches are needed. These retrospectives should lead you to understand what customers want in 90%of cases and how the organization works.
As a result, if you are Agile on an ongoing basis and everywhere, and not for the period of rebuilding / launching / adapting, then this may indicate that:
– you are not following the Agile tools;
– you haven’t found your product or niche, and you haven’t developed the necessary expertise;
– you always have a unique product / project (which should be reflected in the high price of your services);
– the organization is “sick” from the inside, and this way you mask high turnover, lack of work on processes, etc.
What do you think about it?”
The quality of neural network responses is more like a weak student who just picks up similar words, but doesn’t understand what he’s talking about. Could I get a similar post from the neural network? So that she can convey my meaning? Yes, of course, but I wrote my post in 25 minutes, and how long would it take me to try to achieve such a result from AI? And does it look like a really thinking intelligence? And in general, on Intelligence?
The fact is that our thinking is strongly tied to ideas and images. That is, basically in thinking we use an approach from the general to the particular, we think in images, and then from some vague, but understandable to us at the unconscious level of abstractions we come to the smallest details. In parallel, we conduct a bunch of associations, completely tied again to images. It is not for nothing that there are so many metaphors, comparisons, allusions and word changes in human speech.
Modern neural networks “think” quite differently. They simply do not know the concepts of “meaning”, “essence”, “idea”, as well as the process of reflection itself. The texts generated by the neural network look logical not because the neural network really understands what it is writing about, but because it is able to calculate which words are most often next to each other in the texts available for its training on a given topic. Pure mathematics based on calculating probabilities.
However, for the sake of justice, we should also note significant progress in this direction. Now developers are trying to teach AI to think: it asks itself questions and, giving answers, draws conclusions. This is an imitation of an internal dialog to improve the quality of content, which is typical for models from Open AI at the end of 2024 and Yandex GPT 4 Pro RC.
Here is an example of the response of the Yandex GPT 4 Pro RC model to the request” Write a post about why Agile is infinitely applied incorrectly. Emphasize that conducting retrospectives should lead away from the situation of uncertainty to the situation of complex ordered or simple ordered systems.”
Yandex GPT 4 Pro RC Response: “Why Agile doesn’t Always Work: how retrospectives can help move from uncertainty to order.
Agile is a flexible approach to project management that allows you to quickly adapt to changes and improve processes based on feedback. However, despite all the advantages, applying Agile without considering the context can lead to problems. In this post, we will look at why Agile should not be applied indefinitely and how retrospectives help to move from a situation of uncertainty to ordered systems.
Problems of infinite Agile Application
1. Lack of clear structure. Agile is based on the principles of flexibility and adaptability, which can lead to a lack of a clear structure and action plan. This can make it difficult to make decisions and manage risks.
2. Complexity of evaluating results. Without a clear plan and criteria for evaluating results, it is difficult to determine how well the project is progressing. This may raise doubts about the effectiveness of the approach.
3. Risk of losing control. The endless use of Agile can lead to a loss of control over the project. Without clear boundaries and constraints, the team may lose focus and start to deviate from the original goal.
4. Inefficient use of resources. If the team is constantly adapting to new conditions, this can lead to inefficient use of resources. Instead of focusing on key tasks, the team will spend time making constant changes.
5. Reduced team motivation. Constant adaptation and changes can cause stress and fatigue for the team. This can reduce motivation and productivity.
How retrospectives help you move to streamlined systems
Retrospectives are one of the key Agile tools that help the team analyze their work and identify areas for improvement. They allow us to move from a situation of uncertainty to more ordered systems, where processes and decisions become more predictable and manageable.
Here are some ways that flashbacks can help:
– Identifying problems and their solutions. Retrospectives allow you to identify problems that have arisen in the course of work, and find ways to solve them. This helps you avoid making mistakes again in the future.
– Improve processes. Retrospectives help you understand which processes are working well and which need to be improved. This allows you to optimize your work processes and increase efficiency.
– Increase motivation. When a team sees the results of their efforts and understands that they bring benefits, it increases motivation and job satisfaction.
– Create a culture of continuous improvement. Retrospectives foster a culture in which continuous improvement and adaptation are the norm. This helps the team to be more flexible and ready for changes.
In general, retrospectives are an important tool for moving from a situation of uncertainty to more structured and orderly processes. They help the team adapt to changes, improve their processes, and achieve their goals”.”
Yes, the model has become academic and it is clear that these are excerpts from articles, and not reflections. But this is a huge progress, and you can already work with the model and use it as a tool.
However, with only weak AI available and dreaming of a more or less strong one, various researchers already distinguish super-strong AI (ASI, Artificial Superintelligence) in their classification.
This is about an AI, that:
– It can solve both routine and creative tasks;
– instantly navigates uncertainty even without a network or internet connection;
– adapts problem solving to the context of circumstances and available opportunities / resources;
– understands people’s emotions (not only through a text query, but also based on the analysis of facial expressions, voice timbre, and other parameters) and takes them into account in your work;
– able to interact independently with the real world to solve problems.
This is the kind of AI that we only see in science fiction movies so far. Even AI itself writes about ASI as a “hypothetical concept” and “a subject of science fiction and active research in the field of artificial intelligence” (YandexGPT 3 Pro). This is a certain desired point in the distant future, which is not yet possible to reach.
Chat GPT 4 describes ASI as follows: “hypothetical artificial intelligence system with intelligence superior to human. This is not just an advanced AI, but a system that can adapt, learn and develop independently, exceeding human abilities in learning, problem solving, cognitive functions and self-improvement.
Super-strong AI, or ASI, will be able to understand and process many types of data (text, images, sound, video), which will allow it to perform verbal tasks and make decisions. It will use advanced AI technologies such as multidimensional language models (LLMs), multi-bit neural networks, and evolutionary algorithms.
Currently, ASI remains a conceptual and speculative stage in the development of AI, but it represents a significant step forward from the current level of AI”.”
And if there are hundreds of weak AIS now, for each task, then there will be only dozens of strong AIS (most likely there will be a division in directions, we will consider this in the next block), and the super-strong AI will be one for the state and even the entire planet.
Limitations on the path to strong AI
To be honest, I have little faith in the rapid emergence of a strong or super-strong AI.
First, this is a very costly and complex task from the point of view of regulatory restrictions. The era of uncontrolled AI development is ending. More and more restrictions will be imposed on it. We’ll discuss AI regulation in a separate chapter.
The key trend is a risk-based approach. So, in a risk-based approach, strong and super -strong AI will be at the upper level of risk. This means, that legislative measures will also be protective.
Secondly, this is a difficult task from a technical point of view, and a strong AI will be very vulnerable.”
Now, in the mid-2020s, creating and training a strong AI requires huge computing power. So, according to Leopold Aschenbrenner, a former Open AI employee from the Super alignment team, it will require the creation of a data-center worth a trillion US dollars. And its power consumption will exceed all current electricity generation in the United States.
We also need complex AI models (orders of magnitude more complex than the current ones) and a combination of them (not just LLM for query analysis). In other words, it is possible to exponentially increase the number of neurons, build connections between neurons, and coordinate the work of various segments.
At the same time, it should be understood that if human neurons can be in several states, and activation can occur “in different ways” (biologists will forgive me for such simplifications), then machine AI is a simplified model that does not know how to do this. Simply put, machine’s 80—100 billion neurons are not equal to a human’s 80—100 billion. The machine will need more neurons to perform similar tasks. The same GPT4 is estimated at 100 trillion parameters (conditionally neurons), and it is still inferior to humans.
All this leads to several factors.
The first factor is that increasing complexity always leads to reliability problems, and the number of failure points increases.
Complex AI models are difficult to both create and maintain from degradation over time, in the process of operation. AI models need to be constantly “serviced”. If this is not done, then a strong AI will begin to degrade, and neural connections will be destroyed, this is a normal process. Any complex neural network, if it is not constantly developing, begins to destroy unnecessary connections. At the same time, maintaining relationships between neurons is a very expensive task. AI will always optimize and search for the most efficient solution to the problem, which means, that it will start turning off unnecessary energy consumers.
That is, the AI will look like an old man with dementia, and the “life” period will be greatly reduced. Imagine what a strong AI with its capabilities can do, but which will suffer from memory loss and sharp reversals to the state of the child? Even for current AI solutions, this is an actual problem.
Let’s give a couple of simple real-life examples.
You can compare building a strong AI to training your human muscles. When we first start working out in the gym and get involved in strength training, bodybuilding, then progress is fast, but the further we go, the lower the efficiency and increase in results. You need more and more resources (time, exercise, and energy from food) to progress. Yes, even just holding the form is becoming more and more difficult. Plus, the increase in strength comes from the thickness of the muscle section, but the mass grows from the volume. As a result, the muscle will at some point become so heavy that it will not be able to move itself, and may even damage itself.
Another example of the complexity of creating, but already from the field of engineering, is Formula 1 races. For example, a 1-second lag can be eliminated if you invest 1 million and 1 year. But to win back the crucial 0.2 seconds, it may already take 10 million and 2 years of work. And the fundamental limitations of the design of the car can force you to reconsider the whole concept of a racing car.
And even if you look at ordinary cars, everything is exactly the same. Modern cars are more expensive to create and maintain, and without special equipment, it is impossible to change even a light bulb. If you take modern hyper cars, then after each departure, entire teams of technicians are required for maintenance.
If you look at it from the point of view of AI development, there are two key parameters in this area:
– number of layers of neurons (depth of the AI model).
– the number of neurons in each layer (layer width).
Depth determines how great the AI’s ability to abstract is. Insufficient depth of the model leads to a problem with the inability to perform deep system analysis, superficiality of this analysis and judgments.
The width of the layers determines the number of parameters / criteria that the neural network can use on each layer. The more they are, the more complex models are used and the more complete reflection of the real world is possible.
However, if the number of o layers has a linear effect on the function, then the width does not. As a result, we get the same analogy with muscle – the size of top AI models (LLM) exceeds a trillion parameters, but models 2 orders of magnitude smaller do not have a critical drop in performance and quality of responses. More important is what data the model is trained on and whether it has a specialization.
Below are statistics for LLM models from different manufacturers.
Compare the indicators LLaMa 2 70B, LLaMa 2 7B, LLaMa 2 13B. Indicators 70B, 7B and 13B conditionally demonstrate the complexity and training of models – the higher the value, the better. But as we can see, the quality of responses does not radically change, while the price and labor costs for development increase significantly.
And we can see how leaders are building up computing power, building new data centers and hurriedly solving energy supply and cooling issues for these monsters. At the same time, improving the quality of the model by a conditional 2% requires an increase in computing power by an order of magnitude.
Now a practical example to the question of maintenance and maintenance due to degradation. Tut will also be noticeable about the effect of people. Any AI, especially at an early stage, will learn based on feedback from people (their satisfaction, initial requests and tasks). For example, the same ChatGPT4 uses user requests to retrain its model in order to give more relevant answers and at the same time reduce the load on the “brain”. And at the end of 2023, there were articles that the AI model has become “more lazy”. The chatbot either refuses to answer questions, interrupts the conversation, or simply responds with excerpts from search engines and other sites. And by mid-2024, this has already become the norm, when the model simply cites excerpts from Wikipedia.
One possible reason for this is the simplification of the user requests themselves (they are becoming more primitive). After all, LLMs do not invent anything new, these models try to understand what you want them to say and adapt to it (in other words, they also form stereotypes). It also looks for the maximum efficiency of the labor-result bundle, “disabling” unnecessary neural connections. This is called function maximization. Just math and statistics.
Moreover, this problem will be typical not only for LLM.
As a result, to prevent the AI from becoming degraded, you will have to load it with complex research, while limiting its load to primitive tasks. And once it is released into the open world, the ratio of tasks will be in favor of simple and primitive user requests or solving applied problems.
Remember yourself. Do you really need to evolve to survive and reproduce? Or what is the correlation between intellectual and routine tasks in your work? What level of math problems do you solve in this job? Do you need integrals and probability theory, or just math up to 9th grade?
The second factor is the amount of data and hallucinations.
Yes, we can increase the current models by XXXX times. But the same ChatGPT5 prototype already lacks training data in 2024. They gave him everything, they had. And with a modern AI that will navigate uncertainty, there simply won’t be enough data at the current level of technology development. You need to collect metadata about user behavior, think about how to circumvent copyright and ethical restrictions, and collect user consent.
In addition, using the current LLMs as an example, we can see another trend. The more “omniscient” a model is, the more inaccuracies, errors, abstractions, and hallucinations it has. At the same time, if you take a basic model and give it a specific subject area as knowledge, then the quality of its responses increases: they are more objective, she fantasizes less (hallucinates) and makes fewer mistakes.
The third factor is vulnerability and costs.
As we discussed above, we will need to create a data-center worth a trillion US dollars. And its power consumption will exceed all current electricity generation in the United States. This means, that the creation of an energy infrastructure with a whole complex of nuclear power plants will also be required. Yes, windmills and solar panels can’t solve this problem.
Now let’s add that the AI model will be tied to its “base”, and then one successful cyber-attack on the energy infrastructure will de-energize the entire “brain”.
And why should such an AI be tied to the center, why can’t it be distributed?
First, distributed computing still loses performance and efficiency. These are heterogeneous computing capacities that are also loaded with other tasks and processes. In addition, a distributed network cannot guarantee the operation of computing power all the time. Something turns on, something turns off. The available power will be unstable.
Secondly, it is a vulnerability to attacks on communication channels and the same distributed infrastructure. Imagine that suddenly 10% of the neurons in your brain just turned off (blocking communication channels or just turned off due to an attack), and the rest are working half-heartedly (interference, etc.). As a result, we again have the risk of a strong AI that forgets who it is, where it is, for something, then just thinks for a long time.
And if everything comes to the point that a strong AI will need a mobile (mobile) body to interact with the world, then this will be even more difficult to implement. After all, how to provide all this with energy and cool it? Where do I get data processing power? Plus, you also need to add machine vision and image recognition, as well as processing other sensors (temperature, hearing, etc.). This is huge computing power and the need for cooling and energy.
That is, it will be a limited AI with a permanent wireless connection to the main center. And this is again a vulnerability. Modern communication channels give a higher speed, but this affects the reduction of range and penetration, vulnerability to electronic warfare. In other words, we get an increase in the load on the communication infrastructure and an increase in risks.
Here, of course, you can object. For example, the fact that you can take a pre-trained model and make it local. In much the same way as I suggest deploying local AI models with “additional training” in the subject area. Yes, in this form, all this can work on the same server. But such an AI will be very limited, it will be “stupid” in conditions of uncertainty, and it will still need energy and a connection to the data transmission network. That is, this story is not about the creation of human-like super-beings.
All this leads to questions about the economic feasibility of investing in this area. Especially considering two key trends in the development of generative AI:
– creating cheap and simple local models for solving specialized tasks;
– create AI orchestrators that will decompose a request into several local tasks and then redistribute it between different local models.
Thus, weak models with narrow specialization will remain more free and easier to create. At the same time, they will be able to solve our tasks. And as a result, we have a simpler and cheaper solution to work tasks than creating a strong AI.
Of course, we leave out neuromorphic and quantum systems, but we will discuss this topic little later. And, of course, there may be mistakes in my individual figures and arguments, but in general I am convinced that strong AI is not a matter of the near future.
In summary, strong AI has several fundamental problems.
– Exponential growth in the complexity of developing and countering degradation and complex models.
– Lack of data for training.
– Cost of creation and operation.
– Attachment to data centers and demanding computing resources.
– Low efficiency of current models compared to the human brain.
It is overcoming these problems that will determine the further vector of development of the entire technology: either a strong AI will still appear, or we will move into the plane of development of weak AI and AI orchestrators, which will coordinate the work of dozens of weak models.
But now strong AI does not fit in with ESG in any way, environmentalists like it commercial. Its creation is possible only within the framework of strategic and national projects financed by the state. And here is one of the interesting facts in this direction: the former head of the US National Security Agency (until 2023a), retired General, Paul Nakasone joined the board of directors of Open AI in 2024. The official version is for organizing Chat GPT security.
I also recommend reading the document titled “Situational Awareness: The Coming Decade”. Its author is Leopold Aschenbrenner, a former Open AI employee from the Super alignment team. The document is available by QR-code and hyperlink.
A shortened analysis of this document is also available using the QR-code and hyperlink below.
To simplify it completely, the author’s key theses are:
– By 2027, strong AI (AGI) will become a reality.
I disagree with this statement. My arguments are given above, plus some of the theses below and risk descriptions from the authors say the same. But again, what is meant by the term AGI? I have already given my own definition, but there is no single term.
– AGI is now a key geopolitical resource. We forget about nuclear weapons; this is the past. Each country will strive to get AGI first, as in its time an atomic bomb.
The thesis is controversial. Yes, this is a great resource. But it seems to me that its value is overestimated, especially given the complexity of its creation and the mandatory future errors in its work.
– Creating an AGI will require a single computing cluster worth a trillion US dollars. Microsoft is already building one for Open AI.
Computing power also requires spending on people and solving fundamental problems.
– This cluster will consume more electricity than the entire US generation.
We discussed this thesis above. More than a trillion dollars is also invested in electricity generation, and there are also risks.
– AGI funding will come from tech giants – already Nvidia, Microsoft, Amazon, and Google are allocating $100 billion a quarter to AI alone.
I believe that government funding and, consequently, intervention is essential.
– By 2030, annual investment in AI will reach $8 trillion.
Excellent observation. Now the question arises, is this economically justified?
Despite all the optimism of Leopold Aschenbrenner regarding the timing of the creation of AGI, he himself notes a number of limitations:
– Lack of computing power for conducting experiments.
– Fundamental limitations associated with algorithmic progress
– Ideas are becoming more complex, so it is likely that AI researchers (AI agents who will conduct research for people) will only maintain the current rate of progress, and not increase it significantly. However, Aschenbrenner believes that these obstacles can slow down, but not stop, the growth of AI systems ' intelligence.
