Transcription
Listen, well, usually, when it comes to business software, a picture of a huge supermarket immediately comes to mind. >> Uh-huh. With endless shelves. >> Yes, exactly. That is, there are ready-made boxed solutions. Here's accounting software for you. Here's a service for creating presentations, here's a task tracker. >> And you just take what looks more or less suitable off the shelf. >> Uh, yes, you take it, sign up for a monthly subscription, well, for about 15-30 dollars, and just hope that it will somehow cover your current tasks. >> At the same time, you have to put up with the fact that the interface, as a rule, is terribly overloaded. >> Oh, yes, 100%. half of the functions are never used at all, and those that are really needed, on the contrary, are lacking. And we got used to it. >> But the point is that this approach no longer makes sense. That is, what if creating your own, absolutely perfect application to solve one specific pain now costs not 1,000 dollars, >> right? >> But, say, 10-20 dollars one-time, or even zero? >> Wow. Wait, so instead of months of work by an entire IT department? Yes, it takes a couple of hours on a weekend. This is truly a fundamental paradigm shift. In this discussion, we will discuss the transition from the era of forced consumption of mass software to the creation of so-called hyperspecific applications. >> Hyperspecific – that means written for absolutely unique, narrow, personal, or business tasks. >> Exactly. Such software is simply impossible to find on the mass market for one simple reason. It is not economically profitable for large corporations. >> Well, понятно, they need millions of users to recoup development costs. >> Exactly. And a hyperspecific application solves the problem of exactly one process, one company, or, you know, even one specific person. >> Okay, let's unpack this somehow. It sounds like you still need to be a professional developer with a specialized education for this. >> But no. The economics and the mechanics of the issue are now astounding; if you use open language models, that very Open Source, it can be absolutely free. >> And the coolest thing here, as it seems to me, is absolute control. That is, when a tool is created from scratch, for a specific process, you gain power over the algorithms >> and over your own data, which is critically important for any business. >> Yes. But how do you approach this? Where does the creation of such a personal digital assistant even begin? It's logical to assume that you need to choose a target. Right? So, moving on to the first step of our framework, choosing a task. It makes no sense to try to automate everything. >> I agree. >> Ideal candidates for such hyperspecific applications are divided into three categories. The first is what the materials call the curse of existence. >> Sounds dramatic. Well, it is. These are tasks that cause almost physical pain, take hours of life, but are vital to do. >> A classic example is some kind of complex cross-border accounting. Yes, >> yes, yes. Imagine a business that hires contractors all over the world, needs to reconcile invoices in different currencies, consider tax laws of different countries, >> reconcile endless bank statements. >> Exactly. And not a single standard software from that very supermarket handles this smoothly. You still end up having to do everything manually in huge spreadsheets. >> Oh, yes, that's the very routine that makes your teeth grind. And what's the second category then? >> The second category is tasks that are constantly postponed. >> That is, they are strategically important for the quality of work or life in general, but require disproportionately large efforts to start. For example, >> for example, creating a complex habit tracking system where you need to reconcile data on sleep, nutrition, productivity, and all this from different sources. >> Ah, well, yes, people usually give up on such things on the second day because filling everything in manually is just unrealistically difficult. >> Exactly. And here AI comes to the rescue. And the third category, it's probably the most inspiring, is going beyond our usual capabilities. That is, some cherished dreams, >> yes, projects that were previously simply inaccessible due to a банальное lack of specific skills. For example, there's a desire to create your own graphic novel >> or manga? >> Yes, manga, without knowing how to draw at all. >> Listen, I really like this classification. It's very similar to how, imagine, in the office, an ideal, reliable assistant with unlimited potential suddenly appeared. >> A good metaphor. And first, we delegate to him the work that drains all our energy, those very taxes, cross-border spreadsheets, >> just to finally breathe out, >> yes? And then, when the basic stress is removed, this assistant can already be entrusted with helping to realize ambitious creative ideas. But here's the problem >> what? >> The assistant, even the smartest and most advanced, needs to be clearly explained the task. And here, as I understand it, many stumble. It turns out you can't just say to artificial intelligence: "Do it well for me." >> Absolutely correct. And this brings us to the second step. Having decided what pain we want to solve, let's look at process mapping. >> Mapping, >> yes? >> There is a golden rule of system analysis: you cannot automate chaos. >> Logical. Before calling for neural network help, you need to thoroughly, step by step, break down how the process actually works right now. >> Otherwise, the model will just start guessing. >> Yes, it will produce unpredictable, hallucinating, and most likely, absolutely useless results. >> So we literally need to draw a map. And here, there's probably a huge difference between familiar and unfamiliar processes. >> A colossal difference. Let's take that very hated accounting. If a person has been doing it every month for, say, 5 years, the steps are crystal clear. >> They are already at the level of muscle memory, right? Step one - collect all PDF receipts from email. Step two - download transaction history from the bank. Step three - reconcile amounts. And step four - enter data into the final report. This process map is built instantly. We know exactly where we get the data and what result we want to achieve. >> But a completely different picture emerges when it comes to the third category of tasks, something completely new. >> About that manga. >> Yes, it's completely unknown territory. How do you map something you've never done in your life? >> And what to do in such cases? >> In such cases, before writing code, there is a stage of in-depth research of the subject area. You have to study the experience of industry professionals to identify a clear sequence, >> that is, to go through some mini-course or watch training videos. >> Exactly. For manga, this might be, for example, nine stages: from choosing the central theme of the plot and character design to writing the script, storyboarding scenes, drawing details, and final publication. >> And without such a rigid framework, AI will simply generate a disconnected set of pictures. >> 100%. The machine needs to understand the boundaries. >> Okay. Let's say the process map is on the table in all its details. And when the process map is in front of us, there's a huge temptation to give the machine absolutely everything. >> Oh, that's a classic trap. >> But really, nine steps. Great, let it do all nine. Why strain yourself if you can write one big prompt? Generate a complete book for me or reconcile all accounting from scratch to the final report. Isn't the essence of automation to press one button and do nothing at all? >> And here we move on to the third step, defining the role of AI. And here lies perhaps the most non-obvious and important insight of this entire approach. >> So interesting. >> The essence of hyperspecific applications is not to completely replace humans with machines and get some kind of averaged, soulless result. >> Then what? Technology should not take over the entire process if there is no pragmatic sense in it. Or, and this is important, if part of the work brings genuine pleasure. >> Mm, so we shouldn't deprive ourselves of the joy of creativity. >> Exactly. The concept that is important to understand here is the creation of a so-called prosthetic tool. >> Prosthetic tool sounds a bit, you know, medical, but the metaphor is powerful. How does it work in practice? Returning to those nine steps of creating a graphic novel, the first stages involve coming up with the plot, writing dialogues, creating characters. >> Well, that's pure creativity, >> yes? That's the joy of creativity. That's what it was all for. It makes sense to leave this part to yourself. But the next steps >> drawing backgrounds, applying shadows, >> yes, detailed storyboarding - this is an insurmountable barrier without an art education. It is precisely to overcome this barrier that a targeted application is assembled. >> Ah, I see. So, this digital prosthesis compensates for exactly the skills that are lacking to realize the idea. >> Yes, while leaving the person in the position of director and main creator. The ninth step. Publication can be completely discarded if the project is done just for fun. That is, we are not asking AI to write a book for us. We are asking it to be our personal illustrator who works strictly according to our text. >> Absolutely correct. >> This radically changes things. It saves us from such an artificial aftertaste in the result. We maintain control >> and get exactly what we wanted. >> But having understood exactly what part of the routine automation takes over, we move on to the most technical stage. How to go from a diagram on paper to a real program. And here comes the scary abbreviation PRD. >> Oh, yes, PRD >> it sounds like the name of some bureaucratic document from a huge corporation. Is it really impossible without it? >> Moving on to the architecture of our application and the assembly process, yes, PRD stands for Product Requirement Document. And you can't do without it if you need a really working program, not just a fun toy for 5 minutes. >> And if you explain in simple terms what it is? >> Simply put, it's a bridge. A bridge between the business idea in the user's head and the AI coder who will physically write this program. It's a detailed architectural blueprint, >> in which everything is recorded, >> yes? The purpose of the application, the target audience, specific functionality, interface behavior in different situations, and error handling logic. >> You know, it's very similar to communicating with an incredibly diligent, brilliant, but absolutely literal intern. >> An ideal comparison. >> That is, if you throw him the phrase: "Make me a tax calculator," he will make it as he understands it himself. And it might not have an Excel export function, simply because we didn't ask for it. >> Exactly. The more precise the technical specification, i.e., our PRD, the less AI will fantasize. >> Writing such blueprints from scratch is difficult for a non-programmer. Can't we ask the neural network itself to help us compile this document? >> Of course, we can. That's what professionals do. Modern language models are excellent at acting as system analysts. That is, I can just say: "Here's my idea." >> Yes. You describe the idea in your own words and ask the model to ask you 20 clarifying questions about the future application. By answering them, we naturally formulate that very PRD. >> Brilliant. And then the magic of choosing tools for assembly begins. And here, as I understand it, everything depends on the complexity of this very blueprint. Yes, tools are now developing at cosmic speed. If the project is multi-layered, with deep internal logic, >> like that very cross-border accounting, where you need to parse documents and calculate complex interest. >> Exactly. In such cases, environments like CloudCode are used, or >> what's so special about them? Why can't you just paste code from a regular chat, as we're used to? >> Because complex tasks require a systematic approach. Environments like CloudCode allow for parallel work of several autonomous AI agents. >> Parallel? So they work simultaneously, >> yes? One agent can analyze the database structure, another can write a script for processing PDF files at the same time, and a third can check their work for errors. >> Wow! >> And they interact within the developer terminal, solving non-trivial mathematical and logical problems. A regular chatbot will simply get lost due to context limitations. >> Amazing. A whole virtual team. And if the task is more visual, say, a slide generator or a dashboard for tracking habits? >> Here, a platform like Bolt comes into play. And it already looks like science fiction. >> I've seen a couple of demos. It blows your mind. >> Yes, you upload the PRD directly in the browser, and the system starts building a working web application before your eyes. And most importantly, these are not just empty interfaces with placeholder text. >> They have built-in connectors. Yes. >> Yes. This means that the application can connect to real work data via API, pull tasks from Jira, read tables from Notion, extract code from GitHub. >> So, full integration. But working with an AI coder is not just a passive process, like pressing a button and going for coffee. >> In no case. There is a certain framework, a set of principles, so that this assembly does not turn into a catastrophe. First, constant reflection is required. You cannot blindly copy code. You need to at least try to analyze the logic of what the machine offers. >> Uh-huh. And secondly, >> secondly, it is necessary to insist on using existing, proven frameworks and libraries so that the neural network does not reinvent the wheel where ready-made solutions for data sorting or design already exist. >> And probably the most critical thing is a system of checkpoints. That very version control >> absolutely, >> because AI often breaks what just worked perfectly. If it adds a new beautiful button and suddenly the entire tax calculation under the hood breaks, there must be an opportunity to instantly roll back to the previous step. >> Without this, debugging will turn into an endless nightmare. Plus, the machine needs to be constantly given context, uploaded screenshots of errors, fragments of logs. The more data it sees, the faster it finds the problem. I agree. And so we've gone through this path. The PRD is compiled, the code is written, the agents have done their work. The interface is pleasing to the eye. The application is ready and working. >> But where is it physically located? >> Right. Now let's look at the vital issue of hosting. Our fifth and final step. This is the foundation that non-technical specialists often forget about. They forget until the first serious failure or >> data leak. Yes, the question of where the data lives is perhaps the most important in the corporate world. And to explain the hosting options, I really like an architectural metaphor. >> About a hotel, >> yes? The first option is a hotel. A hotel is a metaphor for large cloud providers like AWS or Google Cloud. Living in a hotel is incredibly convenient. >> There's a concierge, regular cleaning. A mini-bar >> exactly. And you can rent 10 more rooms at any moment if a crowd of guests suddenly arrives. In the world of software, this means instant scaling for unpredictable high traffic. >> But you have to pay for this comfort, and the payment is charged for the time of stay and for consumed resources. Bills can grow exponentially. >> Plus, you are firmly tied to the hotel's rules. >> Exactly. And the most acute issue here is privacy. What happens in the hotel doesn't always stay in the hotel. >> Sounds sinister. They are housed on someone else's servers. >> The business owner never knows for sure whether their confidential financial reports are not being used to train someone else's corporate AI models. >> Therefore, logically, the second option appears: a rented apartment. This is a virtual private server, or VPS. >> Yes, here the payment is fixed once a month. No surprises. The tenant will do the cleaning and setup of all processes themselves. But there is significantly more privacy. >> No one walks the corridors. >> However, it's worth remembering that the keys to this rented apartment are still in the landlord's pocket, i.e., the hosting provider. >> So, theoretically, in case of a failure or a legal request, they can enter. >> Yes, access to information can be compromised. And this brings us to the third, most radical, but most reliable option. Your own home, >> local hosting. >> Company. Buying your own home is a one-time investment. >> Yes, you buy it once, and you only have to pay for electricity and internet afterwards. Yes, all repairs fall on the owner's shoulders. >> Installing updates, network security, troubleshooting, all by yourself. >> But in return, something absolutely priceless is achieved. One hundred percent privacy. Data never, under any circumstances, leaves the physical device. >> And here a very strong emphasis needs to be made. For serious work processes, your own home becomes simply the only choice, >> especially for that accounting case. >> Absolutely. Returning to non-standard cross-border accounting. It involves employee salaries, passport data, account numbers, tax declarations of several jurisdictions. Entrusting such hypersensitive information to someone else's clouds is a colossal risk. Changes in the provider's privacy policy or a hacker attack. And that's it >> it can cost the business its reputation and huge fines. >> But listen, there's a practical question here. You can't just take and run a powerful neural network to process thousands of invoices on a regular office laptop, can you? >> Of course not. It will simply melt or go to sleep as soon as someone closes the lid, and the entire automation process will be interrupted. >> Correct. For serious tasks, a regular laptop is generally not suitable. The industry now offers elegant solutions for practitioners. Using open models that run locally on dedicated powerful hardware, >> for example. >> For example, using machines like Mac Studio M2 Ultra or powerful custom builds based on modern video cards. This is a relatively small box, >> yes, which quietly stands in the corner of the office, but it has enough computing power to ensure the operation of complex local AI agents 24/7. >> Wow. So, this creates an absolutely closed, sovereign digital circuit for the business? Yes, the business gets incredible data processing speed because there are no delays in sending information across the ocean to third-party data centers. >> And most importantly, it's independence. No one will cut off access due to external factors. No one will triple the subscription price from next month. >> An application created as an ideal prosthesis for a specific business task runs on equipment that is entirely owned by its creator. >> Sounds like an ideal scenario. And if we summarize all this evolution, the scale of the changes happening becomes obvious. We see how the need for endless compromise disappears. >> Yes, you no longer need to adapt your unique processes to the rigid framework of purchased commercial software. >> Any specialist who deeply understands their subject area now has the tools to create personal, free, and 100% secure solutions. We started with the idea that the familiar world of technology is a supermarket of ready-made solutions, but now the rules of the game have changed definitively. It's enough to identify a bottleneck in your work, break down the process in detail, >> find that very point where a digital prosthesis is needed, >> entrust the AI coder with assembly according to a clear architectural blueprint, that very PRD, and host the result on a reliable local server. The toolkit has become more democratic than ever. And if we combine all this with the broader economic context, a very serious question arises, which, I think, is worth considering for everyone involved in business or technology. >> What question? >> If today a professional without programming skills can deploy ideal, absolutely secure software for automating their work on a local server in one weekend, then what will happen to the SaaS industry in the next 5 years? Yes, that really makes you think, won't giant, unwieldy B2B services with their endless paid subscriptions become just a relic of the past. >> Possibly, >> in this new world, the diagnosis is no longer made by a doctor using a standard factory scanner. Now we assemble the device ourselves that sees exactly the problem we want to solve. There are no more limitations. There is only the limit of our understanding of our own processes. Well said. >> Something to think about in your free time. Until the next dives.