Artificial intelligence. Freefall. Dzhimsher Chelidze

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.

      Chapter 2: Weak, Strong, and Super-Strong AI

      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