マリウス . The TEMU-fication of Software, Digital Goods & Services
A hypothesis on the not-so-distant future of software, books, music, and<br>movies, in which most of what we consume gets cheaper, more abundant, and<br>noticeably worse, while the human-made variant moves into a luxury<br>segment of its own.
Disclaimer: This is an opinion piece and most of it is speculation about a<br>future that has not arrived (yet?), based on a few data points that have. As<br>usual, summary at the end.
A few years ago I would have laughed at anyone telling me that there is a<br>serious market for ten-dollar drills, two-dollar dresses, and one-dollar pairs<br>of shoes shipped from a warehouse on the other side of the planet. Today,<br>however, that market exists and it has a name, and it is even publicly traded<br>(sort of, through holdings). TEMU, Shein and a few others have built frankly<br>mind-boggling businesses around the idea that if you make production cheap<br>enough, fast enough, and just barely good enough to look right on a phone<br>screen, an enormous part of the population will buy it, even when the product<br>breaks within a week, when the materials it is made of contain worrying levels<br>of toxic substances, and when the carbon footprint of one<br>delivery exceeds that of an equivalent local purchase by orders of magnitude.<br>The key to this sort of business model is not innovation, but instead the<br>externalization and compression of cost. Somewhere upstream,<br>people work seventy-five hours a week, in conditions most readers<br>of this website would refuse to even visit, so that the rest of us can have a<br>cheap plastic spatula at our doorstep within five business days. While the<br>visible price collapses, the invisible costs get distributed onto landfills,<br>lungs, and ultimately people that we will never meet.<br>What follows is a hypothesis I cannot prove but have been turning over in my<br>head for a while, as we are watching the same thing happen to software, books,<br>music, (film-)scripts, and most of the digital goods and services we consume.<br>The cheap labor in this case is not human, it is a Large Language Model<br>(LLM), or what many people these days call “AI”, and the externalized cost<br>is, among other things, quality, which requires craftsmanship to produce, and<br>attention to perceive. And just like with physical goods, we will probably end<br>up with a two-tier market, in which we have a large and massively profitable<br>lower tier of generated slop, and a smaller, more expensive upper tier of work<br>that is still recognizably human.<br>I’d like to call this the TEMU-fication of software, digital goods and<br>services, and describe what it might look like.<br>Cheap labor<br>For decades, the global fashion industry has relied on a workforce that has<br>almost no leverage and no voice, and for which the economics work because<br>someone, somewhere far away, will sew a t-shirt for less than the price of a<br>coffee. Without that skewed arrangement, the entire fast fashion business<br>model collapses. The garment in your hand is only cheap to you because it has<br>been expensive to someone else, in ways that the price tag does not show.<br>Modern Large Language Models occupy a similar position in the economy, with<br>one important difference, which is that there is no human being in the<br>sweatshop, only a stack of GPUs trained on a corpus of work that other human<br>beings produced over the course of decades. The labor that has been compressed<br>is historical and the model is a kind of compressed copy of the work of<br>millions of programmers, writers, illustrators, and musicians, served back at<br>near-zero marginal cost. Well, at least in theory, and only if the<br>hyperscalers find a way to lower the cost per token, but that’s a different<br>topic.<br>However, the result is the same. A class of goods can suddenly be produced for<br>an order of magnitude less than before. And, just like with TEMU,<br>those goods turn out to be just barely good enough.<br>Vibe-coded software<br>The most direct manifestation of this so far is what is being called vibe<br>coding. The term refers to the practice of describing what you want in natural<br>language to an LLM, accepting whatever it produces, iterating over it with<br>more refined descriptions of the basic idea and eventually shipping the result<br>into production. Whether the developer actually understands what was generated<br>is increasingly considered an implementation detail. And while the output is<br>technically software, the question is what kind of software it is.<br>A 2025 Veracode report found that approximately<br>45% of AI-generated code samples failed security tests and contained<br>critical vulnerabilities from the OWASP Top 10, and a multi-language,<br>multi-model academic study that evaluated outputs from<br>Claude, Gemini, Codestral, GPT-4o and Llama-3 across<br>Python, Java, C++ and C, found that a substantial fraction of generated snippets<br>were either non-compliant with basic secure coding standards or actively<br>triggered classified weaknesses (buffer overflows, hard-coded credentials,...