Transcription
Greetings. You know, it often seems that creating something with artificial intelligence is, well, some kind of higher mathematics, only for programmers with a lot of experience. Like, you know, almost magic. But what if I told you that this barrier, it's like starting to crack, that creating your own AI application is now possible in principle, even without writing a single line of code, well, in the usual sense of the word code. This is exactly what we want to dive deeper into today. We will analyze this idea of creating AI applications without such, you know, classic programming. We will rely on a recent video review, which, imagine, shows 101 examples of such applications. 101. Right away, it's probably important to clarify. We are not saying that programming is all obsolete. No, of course, for complex, custom, high-load systems, you can't do without it. It is absolutely necessary. But for a whole range of tasks, new approaches, new possibilities are really emerging. And our goal today is to understand how it all works, what the approach is, what exactly can be created with what the source author calls vibe coding. An interesting term, by the way. Perhaps a hint that you are somehow tuning into the same wavelength and more intuition, dialogue. >> Yes. Hello. This is indeed a very, very curious direction right now. Lowering the entry barrier, well, into the development of AI solutions – this is a fact that we are observing. And this opens up, well, simply colossal opportunities for people who have great ideas but lack these formal coding skills you're talking about. At the same time, of course, it's important to soberly assess both the process itself and its boundaries. That is, what can really be done like this on a vibe, and where you can't do without serious professional development, well, at all. To understand both the possibilities and, so to speak, the limitations. >> Okay, let's figure it out then. So, how do you approach creating such an application? The source suggests a rather practical five-step framework. Let's, so to speak, unpack it. Step one. Meta-prompt. A meta-prompt is the very, very early stage. Formulating the idea. Here you need to ask yourself basic key questions. Why is this application needed at all? Who is it for, what should its main functions be, what should it be able to do. In essence, this is a stage of deep conceptual reflection, so that you don't just want an application, but clearly understand what and why. >> Yes, this is the foundation. Without it, it's difficult to move forward. >> And from this first step, from this meta-prompt, comes step two. It's called a product requirements prompt, or PRP for short. This is already a detailed description of the future application. The author compares it to a product requirements document, which product managers write. That is, it's already a detailed prompt, a detailed task, which is fed to the AI tool for coding. And this, in my opinion, is the key moment in this whole story. This structuring of thought. Instead of immediately jumping into, well, the fray, right, asking the AI to make me something cool. >> This framework forces you to first clearly formulate the task, what exactly needs to be built, because the AI, no matter how advanced it may be, doesn't read minds, right? >> Uh-huh. And this PRP, it essentially becomes a technical specification, a TS, for the AI assistant, the very foundation on which it will build something. Without a clear TS, the result will be, well, corresponding, >> random. >> Understand? >> And here, if you believe the source, the very magic they talk about happens. The AI tool, having received this PRP, this detailed TS, can generate, pay attention, 80-90% of the basic functionality. >> Mmm, that sounds impressive, doesn't it? >> Frankly, 80-90%. But, as you said, this is just the beginning, it's not everything yet. Then comes step three. Incremental implementation. That is, what I didn't finish right away or did, well, not quite as I would have liked, it's refined step by step. Iteratively, we ask things like, add a button here or change the background color to blue, or can you connect voice processing? >> Uh-huh. Like, polishing it with a file. Yes, yes, exactly. And almost in parallel with this, step four goes hand in hand. Debugging. Well, it's clear. Errors, bugs, they are inevitable in any development. This is absolutely normal. You can, as the author writes, show it a screenshot or describe the problem and ask: "Please fix it." So, it's a cycle, right? Implementation, debugging, implementation, debugging. We add a feature, check it, find a bug, ask the AI to fix it, check it, add the next feature, and so on. Yes, this iterative process is very reminiscent of real development, as something rarely turns out perfectly the first time, even for experienced teams, but here, I think, it's important to understand this. Although the source author describes it quite easily, right, show it an error, ask it to fix it. In practice, this debugging can turn out to be the most difficult stage for someone without a technical background. >> Mmm, so it's not that simple. >> Well, look. Interpreting an error message that the system gives, or even accurately describing a problem for the AI so that it understands what exactly to fix. This is also a certain skill. The ability of AI for self-correction is, of course, great, and this is precisely part of this vibe coding, but it is not limitless, and it strongly depends on the quality of feedback, on how accurately you can point out the problem. So, these claimed 80-90% of code generation are, well, often the framework, the skeleton of the application. But bringing it to perfection, debugging, fine-tuning – it still requires effort, time, and patience. >> Understand? So there's no magic button to make everything perfect. It won't be without difficulties. Okay. And finally, the fifth step, when the application works more or less as it should – it's deployment. The video author notes that many of these AI tools for vibe coding already offer built-in options for publishing the finished application. That is, they simplify hosting issues, server configuration, and so on. Although, of course, security issues, version control – these are still important aspects, but the basic launch process itself becomes more accessible for a non-programmer. >> Yes, many platforms are now trying to simplify this stage as much as possible. Press a button and your application is online. Well, figuratively speaking, of course. Okay. We've roughly figured out the process, this five-step framework. Now, probably, the most interesting part. What exactly? What applications can be created this way? The source divides them into seven categories. Let's start with the first one. Databases. Data management. It might not sound very exciting, but essentially, it's about how to turn the chaos of unstructured data, well, texts, videos, audio, notes, into something useful, something that can be searched, analyzed, from which value can be extracted. >> Yes, this is a very important and large area. The main idea here is to use AI for, let's say, comprehending large volumes of information, for indexing, for labeling. And most importantly, for semantic search. The typical workflow here is what? First, data is collected, loaded into the system, this is called ingestion, then AI processes and indexes it, that is, makes it available for search, possibly labels it by categories, labeling, indexing, and organizes it into some kind of database. And after that, the system can already accept user queries, and queries in natural language, and find relevant information, considering not just keywords, but the meaning of the query. This is semantic search. Then, what is found can be filtered, analyzed, and presented to the user in a convenient form: text, lists, reports, charts. Wow! The source provides an excellent example that you mentioned. A video lecture database. Imagine, right, you have hours, tens of hours of video recordings, lectures, and you can simply ask the system: "Please find me the moment where the professor explained Bayes' theorem." >> Uh-huh. >> And the AI will analyze all these hours of video, transcribe the speech, understand the context, and show you the desired five-minute fragment. This is a student's dream. >> Yes, and not only. Absolutely. This is powerful for all types of data. You mentioned lectures, but imagine corporate documents from different sources. Something in Notion, something in Slack, something in Google Drive, something in email. And AI can index all of this and allow you to search by meaning across this entire knowledge base, where we discussed the contract terms with company Y. >> Uh-huh. >> And it will find it. >> Yes, this is super useful for companies. And another example from the source. An AI assistant for data cleaning. Well, you know, when you have a messy Excel file, with missing values, incorrect date formats, text where numbers should be, you upload it, and the AI helps find and fix these glitches, bring everything into a single format. A colossal time saver for analysts. And here, the key word, I think, is transformation and structuring of data with the help of AI. It doesn't just store information like a regular database. It actively processes it, understands it, structures it according to requests or predefined rules. Automatic report generation based on data is another area. For example, collecting data from different systems and creating a weekly sales report. Or, as you mentioned, semantic search of a company's internal knowledge base – this is now a very in-demand feature that can really save employees a lot of time. No need to sift through tons of documents manually. >> And with the physical world. This is precisely the next category. Hardware Based Applications. Sounds more complicated. Here, AI connects to various sensors, cameras, sensors, other devices. Hardware, in a word. Yes, here the focus shifts to real-time operation and cyclicity. There is such a continuous cycle: data collection from physical devices, sensors, cameras, microphones, real-time processing and analysis of this data, and AI looks for some patterns, regularities, or, conversely, anomalies, deviations from the norm, decision-making based on this analysis, if something important is detected. Action, this can be notifying a person or controlling some other device or changing system parameters. And then data collection again. And this cycle repeats constantly. >> Uh-huh. A closed loop, so to speak. The example from the source is very illustrative. A road incident detector. >> Cameras on the road, speed sensors transmit a stream of data, and it's analyzed non-stop. And if it detects, say, a sudden congestion of cars, a traffic jam, atypical behavior, it can recognize it as an accident or a serious traffic jam. And what next? It can automatically notify road services or an operator in the control center. It can even try to adjust traffic lights at the approaches to this location to redistribute traffic flows. >> This is precisely a classic example of a real-time system with AI. >> And is there anything like that for home use? AI privacy filters for video are mentioned, for example. That is, a video surveillance system records, and the AI automatically blurs the faces of strangers or car license plates that appear in the frame. Interesting. Or here are anomaly detectors in energy consumption. A smart socket to which, say, a refrigerator is connected, constantly monitors how much energy it consumes, and suddenly notices that consumption has consistently increased by 15% compared to the usual norm. This could be a signal that something is wrong with the refrigerator, possibly an impending breakdown, and the system will send a notification. Yes, these are examples of predictive analytics, predicting problems based on sensor data or wearable devices, gadgets. Imagine an application that aggregates data from various sources. Your fitness tracker, smartwatch, maybe smart scales, a ring. >> And it collects all this together, analyzes it comprehensively – sleep, activity, heart rate, heart rate variability – and gives you a holistic picture of your health. Not just scattered numbers, but some conclusions, trends, personal recommendations. It seems you haven't been sleeping well for the last 3 days. This is affecting your concentration. Try to go to bed an hour earlier. >> Uh-huh. >> Or smart hearing aids. They don't just amplify sound, but using AI, they analyze the sound environment and actively filter out background noise, isolating the speaker's voice. Integration with cars is also a big topic. Collecting data from dozens of sensors in the car to predict possible breakdowns. The system noticed increased vibration in the engine at certain RPMs. We recommend a diagnostic check. In all these cases, AI acts as an analytical center, as a brain that processes signals from the physical world and reacts to them. Of course, huge opportunities are opening up here, but at the same time, very serious questions arise. About the reliability of such systems, about security. What will happen if the AI makes a mistake in analysis or decision-making, especially when it comes to the physical world, cars, medicine, infrastructure? The stakes are much higher here. >> Logical, right? The responsibility is completely different. And look, many of these hardware solutions, and the data management systems we talked about, generate a lot of information. It needs to be presented visually somehow, so that a person can quickly understand what is happening. This brings us to the third category: dashboards, i.e., monitoring panels. This is about displaying information, often in real-time, and often with AI enhancements. Not just charts, but smart charts, so to speak. >> Exactly. The main task of a dashboard is visualization. To present key information in such a way that you can quickly assess the situation and make a decision – visually, clearly, with an emphasis on the most important. The workflow here is usually as follows. First, data collection. It can be from internal company systems, from external sources, it can be updated almost in real-time. Then processing and analysis. Data needs to be cleaned of junk, brought into a uniform format. Calculate key indicators, metrics. Then the most important stage – compilation and visualization. How exactly to display this data, what charts to use, how to highlight trends, anomalies, important changes. This is where AI can help, for example, by automatically selecting the best visualization method or highlighting non-obvious patterns. And the final stage is distribution. How users access the dashboard. It can be a link to a web page, regular email reports with key metrics, notifications when threshold values are reached. >> An example that probably everyone will understand is a personal finance application. An application that connects to your bank accounts, credit cards via API. Well, this is an interface for program interaction. >> Uh-huh. It collects all transactions, automatically categorizes them: food, transport, entertainment, housing. Shows beautiful charts of expenses and income, trends. And here's the AI enhancement. It can not just show numbers, but provide some insight. For example, "Caution, your food delivery expenses this month are already 20% higher than average" or "Congratulations, you've almost reached your vacation savings goal." >> Yes, a good example. The source mentions many more ideas for dashboards. For example, a competitor monitoring dashboard collects news about them, their marketing activities, price changes, or an internal KPI tracker for a company shows key performance indicators of departments in real-time. A dashboard for analyzing customer sentiment based on reviews from social media and review sites. A dashboard for predicting customer churn. A dashboard for tracking competitor prices. Very useful for e-commerce. The video author mentions that she has her own investment dashboard. Yes. Yes, these are all excellent examples. Dashboards with AI elements can do much more than just display static images. They can highlight anomalies, as I mentioned, they can make simple forecasts, for example, predict sales for the next quarter based on historical data and current trends. They can even provide some basic recommendations. This turns a dashboard from a simple reporting tool into a much more valuable decision-making tool for personal finances, business management, and monitoring complex systems. >> I agree. Let's move from data visualization to more active interaction. The next category is very large and popular right now: chatbots and assistants. Chatbot assistant. But here, it's not just a bot that answers questions from a knowledge base, but one that can perform certain actions. >> Yes, this is what is often called agent capabilities now, meaning the ability not just to process information and communicate, but also to interact with other systems, to perform tasks in the real world or in the digital space. The workflow here usually looks like this. The user gives some request or command by voice or text, and the AI determines the user's intent, what they want to do. Then the AI either searches for the necessary information or performs some action. Often, to perform an action, it uses the APIs of other programs or services. That is, it pulls the strings of other systems. After performing the action, it informs the user about the result or asks a clarifying question. And importantly, all these steps are usually logged, recorded, so that it's possible to track what was done and how. >> Uh-huh. >> The example from the video is very illustrative. An accounting assistant for small businesses or freelancers. You simply tell or write to it: >> "Listen, please issue an invoice to company X for $500 for consulting services for this month." Uh-huh. And the AI understands the request, finds the details of company X in your client database or asks you for them. Then, through the API, it connects to your accounting software, for example, QuickBooks, Xero, or something similar. It creates this invoice there with all the necessary data. After that, it informs you: "The invoice has been created, do you want to send it to the client via email?" You say: "Yes." And it sends it. And records in the log: "Invoice number 123 for company X for $500 has been sent." An excellent example of automating a routine task. >> And the source provides a huge number of other examples of such assistants. A legal assistant can help draft a standard contract or check a document for compliance with a template. An HR assistant answers employee questions about internal company policies, vacations, sick leave. A compliance assistant monitors legislative changes and warns about risks. An invoice and receipt processing assistant extracts data from scans, photos, and enters it into the accounting system. And many ideas are related to solving, you know, everyday problems. A medical insurance assistant helps to understand policy terms, fill out reimbursement claims. A medical record organizer collects all analyses, extracts, prescriptions in one place. A booking assistant, this is interesting, which can itself call a restaurant or clinic and book a table or appointment. >> Yes, such technologies are already appearing. Voice AIs capable of conducting dialogue. >> A personal shopping assistant searches for products by description. Compares prices. A subscription manager, oh, this is a real pain, tracks all your subscriptions, reminds you of charges, helps cancel unnecessary ones. A warranty and return assistant stores information about purchases, warranty periods, helps process returns, and even a negotiator for utility bills or internet, who can try to lower your tariff. The potential here is, of course, enormous, because we all have very, very many routine tasks related to information processing, filling out forms, interacting with various services and systems. And assistants can take on a significant part of this routine. However, it is very important here to remember what we discussed in the context of hardware applications. Creating a truly reliable and secure assistant that can perform actions with real consequences, especially financial transactions, legal actions, interactions with medical data. This requires a very careful approach to design, testing, and security. Even if you use code or low-code tools, the cost of error here can be very high. You need to think everything through very carefully. >> I agree. Here the responsibility increases exponentially. Okay. And what about AI that doesn't just perform tasks, but helps us become better ourselves: learn, develop. >> The fifth category: personalized coaching and learning agents. AI as a mentor, yes, this is also a very interesting and promising direction. Here the focus shifts from task execution to providing feedback, to supporting learning and skill development. The workflow here is different. The user interacts with the system in some way. This can be text, for example, a description of a situation, voice, conversational practice, a video recording of their speech or sports exercise, or even some action in a simulator. The AI evaluates this interaction, analyzes speech, text, movement, comparing it with some standards or best practices. Then the AI provides feedback, what was good, what can be improved, suggests specific steps or exercises. It can provide support, motivate, encourage, indicate the next steps in learning. And, importantly, it can track the user's progress over time. >> Sounds great. >> An example from the source: a relationship coach. You describe some conflict with your partner, and the AI asks clarifying questions to better understand the situation from both sides. It analyzes communication, gives advice on how to express your feelings or needs more constructively, may even suggest more empathetic phrasing, may even offer a role-playing game where you practice having a difficult conversation, and the AI plays the role of the partner, and then tracks how your communication changes. >> Uh-huh. Such a safe space simulator for practicing skills. Or a public speaking coach. You record a video of your presentation rehearsal, upload it, and the AI analyzes your speech, pace, intonation, filler words, gestures, eye contact, if the camera allows, and gives specific recommendations. "Try to speak a little slower, avoid the word 'um', look at the audience more." >> A language learning coach is already quite common. You practice speaking, and the AI understands, responds, corrects pronunciation, grammar, gives vocabulary hints. All in real-time. The video author mentions that she herself created such a coach. Another idea from the review. A drawing coach analyzes your sketches, gives advice on composition and perspective. A sports coach, for example, analyzes a video of your golf swing or swimming technique. A career coach helps prepare for interviews, practices answers to tricky questions, and even highly specialized coaches for pottery or playing a musical instrument. >> This direction makes personalized learning and skill development potentially much more accessible. Not everyone can afford to hire a personal public speaking coach or a golf instructor. And AI can provide basic practice, initial feedback, help master some fundamentals. Of course, it's important to understand that AI, at least for now, will not completely replace a live expert. It doesn't have the same empathy, intuition, ability to see the whole picture, understand a person's unique context, but as a tool for regular practice, for obtaining objective, albeit algorithmic, feedback, for lowering this initial barrier of fear – it's a very promising thing. The entry barrier to learning is lowered. Yes, indeed, an exciting opportunity to get a personal trainer for the price of a subscription. So, we've already mentioned working with different types of data several times: text, video, audio, images. The sixth category focuses more deeply on this: multimodality. Here the emphasis is on generating and remixing content using these different modalities. Yes, modern large language and generative models are becoming better at working with different types of data simultaneously; they can understand and generate not only text, but also images, code, sound, video. And, more importantly, they can connect these modalities. The workflow for content generation here often looks like this: the user provides some idea, topic, prompt, may upload some source materials. And based on this, a draft of the content is generated. This can be the text of an article, the structure of a presentation with text and images, scripts for a video, a set of posts for social media. Then comes the collaborative refinement stage. The user and AI edit, refine, personalize this draft together. The user makes edits, asks the AI to change something, add something, redo something. And this continues iteratively until the result is satisfactory. And finally, the final content is published or exported in the desired format. This is a process of co-creation, of human-AI collaboration. >> The example from the source is precisely about this. A presentation generator. You set a topic, for example, "The History of Video Games from Pong to Today," and the AI selects information, structures it into slides, generates text, finds or creates relevant images. This results in a first draft, and then you go through it, add your own emphasis, change phrasing, choose different pictures if needed, and export it to PowerPoint or PDF. It significantly speeds up the process, >> undoubtedly. Or generating posts for social media. You give the AI a link to your blog article and ask: "Make three posts for Twitter, one for LinkedIn, and an image for Instagram from this." And it generates adapted content for different formats: generating short video clips from a text description, creating newsletters based on a selection of articles, even generating music or sound effects on demand. >> And it can not only generate but also analyze multimedia content, for example, find trends in photos in a specific niche or comment on videos in real-time. The source provides an example of AI commenting on a football match. Funny. Format converters also fall into this category. Video to text, transcription, text to image, image generation from description, text to audio, voiceover. Interactive stories that combine text, images, sound, and the user can influence the plot. In general, a huge field for creativity and automation of content tasks. Yes, multimodality opens up truly great opportunities for content creators, marketers, educational projects, and media. AI's ability to understand, generate, and connect different types of data allows for the creation of completely new formats of communication and interaction. And, of course, to automate many routine processes related to content preparation and adaptation. >> Very interesting. And finally, the last, seventh category. It seems to me it's all-pervasive, because many examples from other categories can also be attributed here. This is automation and macros. This is about entrusting AI with the execution of routine repetitive tasks that we do every day on a computer or in cloud services. All these little things that eat up time. >> Exactly. And these automations can be broadly divided into two main types. The first are cloud automations. They run on servers, often using special platforms for integrating services, such as Zapier, Make (formerly Integromat), or built-in automation tools in large systems like Salesforce or HubSpot. They often utilize APIs of different services and AI models for decision-making or data processing. The second type is local automations or macros. They run directly on the user's computer. These can be scripts or small utilities that automate actions in the operating system or specific applications, browsers, text editors, spreadsheets. The general principle of automation usually works like this: an event occurs – a trigger. This can be receiving a new email, adding a row to a table, downloading a file, pressing a hotkey, a specific time arriving. This trigger launches the AI or script, which performs a predefined sequence of actions, extracts information, processes it, launches some process, sends data to another system. And often the result of this work is simply logged, recorded somewhere, or passed further down the chain. The goal here is maximum autonomy, so that it works on its own without constant human intervention. >> Let's use examples. An example of cloud automation from the source: a customer review classifier. Trigger: a new review appears in your CRM system, well, in your customer management system. >> Uh-huh. >> This trigger launches an AI model, which automatically analyzes the review text. It determines the sentiment: positive, negative, neutral, understands the essence of the problem, assesses urgency. >> Based on this, it can, for example, automatically create a task in a project management system, say, in Jira or Asana, assign a responsible employee from the relevant department – support, sales, development – and set a priority. And then it logs this action. "Review X processed, task Y created for employee Z." That's it, a person doesn't have to manually read every review and decide what to do with it. Yes, this saves time and speeds up response to customer inquiries. And an example of local automation. Macros, a file organizer on your computer. Trigger: you download a file to the downloads folder. This trigger launches a local script, and it looks at the file name, its content. If it's a document, it determines the file type, for example, it understands that it's an invoice, it extracts the date, counterparty number, amount from the invoice. It renames the file according to your standard, for example, "Invoice YYYY, MM DD Counterparty PDF." And automatically moves it to the correct folder on your disk, for example, "Documents/Invoices/2024." The video author jokes there that her downloads folder is just chaos. And she would definitely need such a macro. Oh, yes, familiar problems. And there are just tons of such ideas for automation. If we talk about cloud-based, an automatic responder for standard email requests, automatic processing of incoming invoices, extracting data and entering it into an accounting system. A smart meeting scheduler that itself finds free slots for all participants and coordinates the time. Automatic enrichment of data about potential clients, leads. Found a new lead. The system itself searches for their LinkedIn profile, company website, adds this information to the CRM. Automation of the recruitment funnel, initial resume screening by key skills, sending test assignments, a code review assistant for programmers that analyzes code and suggests improvements or finds potential errors. Task generator from chat discussions, if something was agreed upon in Slack or Teams, and it can itself create a task in a task manager. Automatic generation of draft commercial proposals on demand, checking documents for compliance with corporate standards or regulatory requirements. >> An impressive list. And what else can be thought of for local macros? A PDF or long document summarizer with a right-click of the mouse: select the file, click, get a summary. A clipboard assistant: you copy text, and it offers to expand it, rephrase it, or translate it. Voice commands for computer control. "Open browser, create new document, find file X." Macros for spreadsheets, for example, inserting frequently used formulas or formatting data according to a template with one command. Screenshot analyzer. After taking a screenshot, AI can itself highlight text or important elements on it, add annotations. Local semantic search of all files on the computer, searching not by words, but by meaning. Autotagger for photos, it recognizes objects in photos and assigns tags for easy organization. An assistant for presentations, working locally for data security, suggests talking points, checks spelling on slides. A scanner for files for confidential information, passport data, card numbers, before you upload them anywhere on the internet. >> Yes, essentially, it's using AI to create such small personal assistants that automate these small, but very numerous and often annoying routine operations that make up our digital day. And the cumulative effect of such automation can be very significant. So, we've had a, well, truly deep overview, I think, from understanding data and creating knowledge bases through interaction with the real world using hardware and visual dashboards to smart assistants, personal coaches, multimodal creativity, and finally, to automating everything and anything. The spectrum of possibilities is indeed very wide, even if you don't delve deeply into code. I hope our discussion today has given someone a couple of ideas or perhaps a new perspective on what is already possible with AI today. Yes, I think the main conclusion from all this is that the barriers to creating, let's say, relatively simple or prototype AI applications are indeed lowering, technologies are becoming more accessible, tools are emerging that allow ideas to be implemented quickly and with fewer technical skills. This is great. But, as we've emphasized more than once today, it's important to remember the limitations. Complex, high-load, critically important systems, where the cost of error is high, still require deep expertise, professional development, thorough testing, an engineering approach. This vibe coding we talked about is an excellent tool for prototyping, for testing hypotheses, for solving personal tasks, for creating simple services, but it does not replace serious development where it is objectively necessary. It's important to understand the possibilities and limitations of applicability. >> I completely agree. And in conclusion, as usual, a thought to ponder that we want to leave with our listeners. >> Think about this. What routine, perhaps even slightly annoying task in your daily life or work is just begging to be automated with a simple AI application that you might even be able to try to create yourself using these new tools. And the second question, perhaps a bit more philosophical. What new questions, not only technical but also ethical and social, arise before us when AI becomes not just a tool for data processing, but begins to act as a decision-making assistant, or a personal coach influencing our skills and views, or even becomes our eyes and ears in the physical world through hardware applications. I think this is worth thinking about.