I Built a Program from Scratch with Kimi: An Experience That Made Me Understand a Lot About AI and Who Will Be Able to Afford It

A few days ago, I decided to build a program from scratch. Not a quick proof-of-concept, not a few lines of script: a real Windows application with a graphical interface, capable of recursively analyzing the filesystem, calculating how much space the folders and installed programs occupy, displaying them as a percentage of the total available space, and, should one decide to do so, uninstalling them directly from the interface. A useful application, well-made, that has practical sense in the real world, inspired by ncdu (NCurses Disk Usage), the command-line utility for analyzing disk space in Unix-like operating systems, or, returning to the Windows environment, by applications like WizTree.

I decided to develop it with the help of Kimi, which is not a Formula 1 driver but an artificial intelligence model that perhaps many still do not know, but which deserves serious consideration. I must preface that I have no, or very little, programming skills.

This story has many levels. There is the technical one, which tells what it means to build software with an AI assistant as a collaborator. There is the economic perspective, which concerns how different models on the market manage free access and usage limits. And there is the one that, to use a strong but precise word, relates to the future: who will be able to afford the best artificial intelligence? And what will happen to those who cannot?

What Is Kimi and Where Does It Come From

Kimi is the AI assistant developed by Moonshot AI, a Chinese startup founded in April 2023 in Beijing by Yang Zhilin, a former researcher at Carnegie Mellon and Google, along with two other ex-classmates. The name "Moonshot" is not random: according to what the founder himself has reported, it is a homage to the 50th anniversary of The Dark Side of the Moon by Pink Floyd, with the "dark side of the moon" as a metaphor for what has yet to be achieved but is worth pursuing. The company started with an initial investment of $300 million and about 40 employees, with financial backing from big names like Alibaba.

In just a few years, Moonshot AI has built one of the most interesting models. The Kimi K2 model, released in 2025 as open source, is a Mixture-of-Experts architecture with a trillion total parameters, specifically designed for advanced reasoning, code writing, and long-context management. Kimi K2.6, the latest available version, has a much more aggressive API pricing compared to competing frontier models: $0.95 for a million tokens in input in cache miss mode, $0.16 in cache hit, and $4.00 for a million tokens in output. For a fair comparison, the closest high-end model from OpenAI is GPT-5.5, which costs $5.00 for a million tokens in input and $30.00 for a million tokens in output. However, I should clarify that, in my case, I used only the free version of the chatbot, without the APIs. Kimi can also orchestrate up to 100 agents in parallel, an agentic architecture that, as we explored in our editorial Why Agentic AI Is True Artificial Intelligence, is currently the most promising frontier in this field.

The Program: What It Does and Why I Built It

The starting idea was simple: to have a tool that allowed me to immediately understand where disk space goes. It should show folders sorted by size, with a proportional bar that made the relative weight of each visible at a glance, and that would allow uninstallation of applications directly from the interface without going through the system settings. A single, essential, and practical control panel.

The result, as you see in the image, is an application with two main sections: one dedicated to the File System, which analyzes and sorts folders by size showing a relative usage bar; and one dedicated to Installed Applications, which allows checking which programs are installed on each disk present in the system. In the File System, the program must also identify simple folders, which can be deleted without the uninstallation program, and programs that must be uninstalled only with the uninstallation to avoid compromising system integrity. The interface is built in WPF with a dark theme, with colors that immediately communicate the size hierarchy and action buttons placed where they are needed.

Developing with Kimi: Iteration as a Method

The development process with Kimi follows an iterative approach that has become typical of any work done with LLMs: it starts with an initial request, the model generates a first version, it is tested in the field, errors or parts that need improvement are identified, and corrections are made. The point is not to get perfect code right away, but to gradually refine results, add context at each step, and let the model rework the problem in light of subsequent feedback. It is precisely this mode of working, rather than a single definitive answer, that makes effective the use of Kimi in real development.

The extraordinary thing about this process is that it works even when the problem is opaque. If a compilation error, an unexpected interface behavior, or a crash under specific conditions is reported to Kimi, the model often manages to pinpoint the cause in the code without it being explicitly stated. This is not magic: it is the direct consequence of what we discussed in the editorial on Agentic AI. The model has seen millions of code examples, bugs, and related fixes. When a broad context is shown, such as a part of code, an error message, or a description of the observed behavior, it effectively recombines these patterns, moving towards the solution through successive approximations. Each iteration adds information to the context, and with more context, the model works better. That is exactly why a broad context window is not a minor technical detail, but a condition that makes this type of collaboration possible.

The Challenges: The Installation File and Windows Permissions

The technically most complex part was not the logic of the application, but the creation of the installation file. Windows is an operating system with decades of layers in permissions management, and indeed, the main issue revolved around administrator rights: the application must correctly request privilege elevation, handle cases where the user refuses, access protected system paths without generating unhandled exceptions, and sign the installation package so that Windows does not block it or flag it as potentially dangerous. And, of course, work on different PCs. Each of these nodes required various iterative cycles with Kimi: behavior was verified, problems were reported, the model identified the cause in how permissions were requested or in the structure of the installation script, and proposed a specific fix. None of these issues were immediately evident: they emerged one by one, as testing became more in-depth and edge cases were explored, even on different PCs.

The Advantage of Kimi: Broad Context and Generous Access

All this work composed of hundreds of messages, thousands of lines of code shared, and debugging sessions lasting hours was made possible mainly thanks to Kimi's large context window, which allows working on very long conversations while maintaining a huge amount of material at hand. This aspect is distinct from usage limits, which instead depend on the service's business policy and rate limits of the adopted plan. It is precisely the combination of extended context and relatively generous access that makes Kimi particularly interesting for complex projects like software development.

The comparison with the other main models is striking. ChatGPT in the free version limits access to the more capable models with relatively low daily message counts and requires a subscription of about 20 euros per month to use GPT-4 without significant restrictions. Claude from Anthropic has the same price for the Pro plan, with usage limits that are felt in intensive work sessions. Gemini Advanced from Google also costs about 20 euros per month. In contrast, Kimi, on its web interface, allows working for long periods on complex sessions without payment, up until the point where the chat becomes too long to manage, as shown by the message visible in the following image, which appears when the model indicates it has reached the limit of the current session. It is the signal that it is time to open a new chat, not to pull out the credit card.

In this case, the difficulty was transferring all the work from one chat window to another because, like all these chatbots, Kimi only remembers conversations that occurred within a single chat (a restriction imposed more by limits determined by subscription plans than by actual LLM limits). We found a way to transfer material from one chat to another, always through interactions in natural language (which is obviously fundamental for someone like me who does not know programming well), but, again, this happened after dozens, if not hundreds, of interactions with the chatbot.

It is not the only model to follow this logic, definable as "permissive". Google AI Studio with Gemini 2.5 Flash offers up to 1,500 requests a day for free via API, with a context window of one million tokens. DeepSeek, another Chinese model, has similar generosity in its access conditions. Groq allows using Llama 3 at high speed in the free version, with daily limits but without requiring payment. The landscape of models available for free (or almost) exists, but it requires knowing where to look and understanding the limitations of each, within a world that will see the ability to ration and manage the available tokens as increasingly important, perhaps distributing parts of the work to one model or another appropriately.

Promotion or Strategic Choice?

An open question remains, and it is no small matter: is Kimi's generosity a temporary launch strategy, destined to decrease as the service matures, and the company needs to find a sustainable business model? Or is it a structural choice, related to the desire to rapidly penetrate the market by offering something that American competitors do not provide under the same conditions?

The honest answer is that we do not know. Moonshot AI is a startup, and startups must sooner or later deal with the real costs of infrastructure. Training and maintaining a model the size of Kimi K2 requires hundreds of millions of euros a year. Broad free access can be sustained as long as funding comes in, but AI business models are solidifying in a fairly clear direction: a free tier with acceptable limits for occasional personal use, and a paid tier for intensive professional use. The question is not whether this transition will occur, but when and with what intensity.

The AI Digital Divide

This is where the reasoning becomes most important. In the world of work, the ability to use artificial intelligence effectively is already becoming a real competitive factor. Those who know how to use these tools well to build software, analyze data, write documentation, and automate processes work faster, produce more, and access skills that would normally require years of specialized training. This is already true today, with models available for free.

But the issue does not simply end up being "who pays more has more." Many capabilities currently considered cutting-edge, such as long contexts, advanced reasoning, agentic coding, and part of multimodality, will progressively tend to be available in open, local, and self-hostable models. The real gap, therefore, will not only be between those who can afford the best cloud services and those who cannot, but also between those who will be able to immediately exploit the most advanced solutions and those who will have to wait for those same capabilities to become available in more easily reachable forms.

It is a new kind of digital divide: not the one between those who have access to the internet and those who do not, but the one between those who have immediate access to the most powerful AI and those who must rely on a slower or more limited version. In a job market that is being reconfigured around these skills, this difference risks becoming structural. Those who start at an advantage produce more, earn more, and can invest more in their tools; those who start with fewer resources risk arriving late even when the technology eventually democratizes.

What This Experience Tells Us

Returning to the program I built: it exists, it works, and I built it without writing a single line of code. This fact, in itself, says something important about where we have arrived. I am not a professional programmer. I have a sense of how code works, I know how to read what the model produces and understand if it does what it should, I can describe a problem accurately enough to get a useful solution. This, today, is enough to build real software.

Artificial intelligence did not write the program instead of me: it wrote the code while I defined the problem, checked the result, identified friction points, and guided the process towards the solution. It is a collaboration, not a delegation. And it works because the model has enough context to understand the project as a whole. In other words, I managed to achieve this result (in just 2 days of work) thanks to a generous token window that made it possible to work on long sessions without losing track.

If that window were to shrink or become paid, this story would have been different. Slower, more fragmented, perhaps unfinished. This is why the question of who will be able to afford the best AI is not an abstract question. It is already a question regarding what we can build today and what we will be able to build tomorrow.