redcpp · Oct 9, 2026 · Oct 8, 2026
diff --git a/src/content/blog/en/what-ai-makes-cheap.md b/src/content/blog/en/what-ai-makes-cheap.md
@@ -0,0 +1,104 @@
+---
+title: "What AI Makes Cheap, and What It Doesn't"
+subtitle: "Everything that lives on a computer is getting cheaper. What sets us apart is no longer intelligence, but what we own and how we direct."
+excerpt: "I used to bill 50 dollars an hour as a remote programmer; today I'm offered 29. What gets cheaper when AI can do everything that lives on a computer, what doesn't, why it arrives differently in Mexico, and how I use it: as a secretary, not an oracle."
+date: 2026-10-25
+readMinutes: 11
+key: lo-que-la-ia-abarata
+tags:
+  - Artificial intelligence
+  - Work
+  - Mexico
+---
+
+*This essay was written in Spanish. What follows is a translation; the [original](/blog/lo-que-la-ia-abarata/) is the one the argument is about.*
+
+## My edge disappeared
+
+It finally sank in that what I was selling, remote code by the hour, is losing its price. I was offered a remote job as a tech lead for 5,000 dollars a month. Full time, that's about 29 dollars an hour: 42% less than I was billing in 2021. And the openings I came across as a full stack engineer were junior level and didn't reach that rate either. People expect more for less.
+
+In 2021 I was a full stack engineer, a contractor for companies in the United States. My skill compared to other programmers' was large enough for a company to look outside its own country for someone with senior ability, at a slight discount for working remotely from Mexico. I competed with engineers there without trouble and was paid 50 dollars an hour, because very few people could solve LeetCode problems, the exercise platform companies use to screen programmers. There was no way around it: either you found a good one, or you found a good one. You couldn't take a sloppy one, hand them a weapon as powerful as artificial intelligence and raise them to your level.
+
+Now you can, and I see it every day. AI does the work much faster, better and more organized than I do, and it doesn't get tired. I understand what's going on and can guide it with finer moves. But someone who uses it very crudely still gets results that are good enough.
+
+Thinking through the reasons, I realized I had lost my competitive edge: the distance between me and a programmer who used to be average or bad. That difference is something companies are no longer willing to pay for, least of all for someone remote in Mexico. The rejections confirmed it, and so did the data: according to Indeed (2026), in September software development openings in the United States were 23% below their February 2020 level.
+
+So if anyone with artificial intelligence is better than I am without it, there is no longer a gap I can close. That's when a different question started to matter: what my advantages over others are. The technical part, on its own, is no longer one of them. That's what I set out to think about and research, starting with understanding exactly what is getting cheaper.
+
+## Everything that lives on a computer
+
+Remote cognitive work is all the work that can be done from a computer. And everything we do on a computer, the computer itself can do. A few years ago it was unimaginable that a computer could operate itself as generally as it does today.
+
+The difference is that today's artificial intelligence is general. The ones that existed before were specialized: they filtered email to catch spam or recommended what to watch on Netflix. And today's doesn't just build software, it also uses it. With MCP, a standard that connects AI directly to other applications, it can operate tools we used to run only through buttons and screens.
+
+Setting aside privacy and access restrictions, in principle any work done from a computer can be done by an AI, as long as someone defines the requirements well. So the real work is defining the requirements.
+
+AI is very good with anything that is text and thought; with the physical, not so much, and there it will take longer to have such a large impact. But in the digital world it is more meticulous, much faster, works in parallel and never stops. That will make cheaper, and speed up, the pace at which needs, solutions and uses are created in finance, accounting, design, communications and sales.
+
+And it's getting cheap fast. A study using Epoch AI data found that, for the same level of capability, the price of using the most advanced models falls between 5 and 10 times a year (2025).
+
+## What doesn't get cheaper: owning
+
+What has set us apart as humans is our ability to think in the abstract. But if anyone has access to a source of intelligence that guides their actions with great agility, and cheaper every year, intelligence stops setting us apart. What sets us apart is what we already owned before.
+
+When we own land, a license or a relationship of trust, that is valuable because no cognitive work can replace it. In fact, whoever owns those scarce things can use cheap cognitive work, either directly or through subcontracting: you hire a human who in turn subcontracts an artificial intelligence.
+
+That's how I'm living it. I went from selling hours to running a business with an asset, and today my focus is on assets, no longer on trading hours for money.
+
+The flip side of this idea is the *permanent underclass*, which has been widely discussed lately. According to that idea, if AI ends up doing almost all the work, whoever hasn't accumulated capital before the transition would have no way to rise through their own effort. The term was used in the sixties for the workers automation left behind, and this year it began circulating again in Silicon Valley (The New York Times, 2026). Many experts don't expect that scenario, but the idea comes up more and more. And the most exposed are those who are just starting out.
+
+## From pyramid to column
+
+It worries me a lot that the structure of companies looks more and more like a column and less like a pyramid, with fewer and fewer entry-level jobs at the base. The data already points that way. In the United States, between November 2022 and June 2026, employment of 22- to 25-year-olds fell about 11% in the occupations most exposed to AI. In the least exposed ones, it grew close to 10% (Stanford University, 2026).
+
+We went from the industrial era to the era of computing and information, and now the era of artificial intelligence is arriving: a completely new paradigm that breaks old models and knocks down entire fields. We are not prepared. Neither schools nor universities are.
+
+Universities prepared us for cognitive work, and new graduates never had the chance to acquire land, a license or relationships of trust that would put them in a good position to catch the wave. They are right where it breaks: it's going to crash on them and they won't ride it. The same happens to those who are already poorly positioned. And I worry about children, because I don't know what careers there will be in the future. Nobody can say for sure: twenty years is a very long time, and in this new era a single year is already a huge change.
+
+But it also excites me a lot. A productivity boom is coming and services will keep getting better. The cognitive part is still important for those who know how to orchestrate these artificial intelligences. Anthropic, the company behind Claude, is preparing to go public and, according to Bloomberg, investors value it between 1.8 and 2 trillion dollars (October 2026). That speaks to how much room there is to grow. There are many opportunities, but they are closely tied to AI itself.
+
+It excites me that, as cognitive work gets cheaper, cognitive solutions get cheaper too. This can bring a lot of chaos and a lot of inequality, but at the same time better medicine, more justice and more freedom. There are many good things to come.
+
+## In Mexico it arrives differently
+
+Even so, I worry especially about young Mexicans. Here, the occupations most exposed to AI add up to 2.93 million workers, 4.9% of employment, and they include exactly the jobs of a first office job: administrative support, secretarial work and data entry (México, ¿cómo vamos?, 2026).
+
+But in Mexico everything moves at a different pace. Capital and credit are less liquid: only 37.3% of adults have any formal credit (ENIF, 2024), compared with 82% in the United States (World Economic Forum, with 2022 data). Here capital is mostly institutional, it comes from funding lines, and money moves more slowly.
+
+It's ironic, because we have one of the best payment systems in the world. SPEI has been running since 2004; the United States didn't launch its equivalent public system, FedNow, until 2023. In 2025 SPEI processed more than 7.3 billion transfers, 36.8% more than the year before. And still, according to Banxico itself, between 70% and 75% of payments in the country are still made in cash. We have the infrastructure, but adoption takes longer.
+
+You see it in how slowly the credit system has grown. You see it in how we still use WhatsApp for everything, when we could be using better tools, like iMessage, very popular in the United States, or Slack.
+
+So here artificial intelligence arrives differently, because we are slower to adopt innovation. We are more cautious.
+
+If I were that 22-year-old today, I would go all in on artificial intelligence: understanding it, using it and offering my services with it. I would do the same thing I do today, but I'd put even more effort into learning, because at that age I had so much time and energy. I wouldn't be afraid: I would take big risks, start a company and bet big on niche technologies within AI, solutions that could apply to a great many companies.
+
+## Secretary, not oracle
+
+Today I use artificial intelligence as a tool that speeds up my work, but I don't let it dictate my understanding. I delegate to it what I used to find tedious, as I would to a secretary: organizing my email, analyzing the work I've done, finding weaknesses and recommending improvements. Then I, with the judgment and personal taste I've developed over the years and through my experiences, tell it more precisely what I want. What I don't let it do is decide what I should think. That's the most important part: sometimes people ask it what to think, and that's the mistake.
+
+There is evidence that this difference matters. In an Anthropic experiment, programmers who learned a new tool with AI help scored 17% lower on a test about what they had just used. But it wasn't automatic: those who used AI to ask questions and request explanations did retain what they learned (Anthropic, 2026).
+
+We have to be very smart about what we let AI do. I do many things by hand. I talk more now, because I'm much more efficient speaking than typing, and AI transcribes what I dictate. But the emails and messages to my family, my friends and the people close to me I write myself, because they deserve to know what I really think. They deserve my connection, not a proxy made of flesh. They don't want to ask ChatGPT; they want to ask me, to talk with me.
+
+AI opens the door to productivity and lets us do more things at once. But what becomes hard is context switching: we have to carry more context in our heads, learn faster, pick up different threads, run them in parallel, orchestrate and analyze.
+
+Today AIs communicate in text, and organizing many of them to work together is a problem the world of management has been trying to solve for years. These are very human forms of organization that we are now applying to artificial intelligence. They will no doubt specialize for whatever AI needs, but that orchestration, that collaboration, already existed in the physical world. That's what we invented language for.
+
+So the cognitive work of the future is more about reflection. Less grunt work and more thinking: sitting down to think about the right direction to give one or several artificial intelligences. It's the work of directing. We have to learn to be better directors and to communicate better; with AIs and with our fellow human beings.
+
+## Sources
+
+Accessed October 6, 2026.
+
+- Indeed Hiring Lab, [software development job postings in the US (via FRED)](https://fred.stlouisfed.org/data/IHLIDXUSTPSOFTDEVE), data as of Sep 11, 2026.
+- Epoch AI and Artificial Analysis, [The Price of Progress: Algorithmic Efficiency and the Falling Cost of AI Inference](https://arxiv.org/html/2511.23455v1), Nov 2025.
+- The New York Times (opinion), [Silicon Valley Is Bracing for a Permanent Underclass](https://www.sltrib.com/opinion/commentary/2026/05/02/opinion-silicon-valley-is-bracing/), republished by The Salt Lake Tribune, May 2026.
+- Brynjolfsson, Chandar and Chen, [Canaries in the Coal Mine? (August 2026 revision)](https://digitaleconomy.stanford.edu/news/canariesaug26/), Stanford Digital Economy Lab.
+- México, ¿cómo vamos?, [exposure to generative AI using the Q1 2026 ENOE](https://expansion.mx/economia/2026/08/11/ia-amenaza-primer-empleo-formal-jovenes-mexico), via Expansión, Aug 2026.
+- Bloomberg, [Anthropic's IPO](https://qz.com/anthropic-ipo-thanksgiving-valuation-100226), via Quartz, Oct 2026.
+- INEGI and CNBV, [ENIF 2024](https://www.bbvaresearch.com/wp-content/uploads/2025/03/Tenencia-de-productos-financieros-y-acceso-a-cursos-basicos-para-finanzas-en-2024.pdf), via BBVA Research, Mar 2025.
+- World Economic Forum, [formal credit in the US with 2022 data](https://www.revistaabogacia.com/credito-formal-el-pendiente-de-la-inclusion-financiera-en-mexico/), via Revista Abogacía.
+- Banxico, [SPEI transfers in 2025](https://marketing4ecommerce.mx/spei-sigue-sin-ganarle-al-efectivo-la-paradoja-de-los-pagos-en-tiempo-real-en-mexico/) and [cash use](https://expansion.mx/economia/2026/03/19/banxico-preve-que-transferencias-con-spei-superen-pagos-con-tarjetas-en-2026), 2026.
+- Federal Reserve, [FedNow launch](https://frbservices.org/news/fed360/issues/080123/fednow-service-instant-payments), Jul 20, 2023.
+- Anthropic, [How AI assistance impacts the formation of coding skills](https://www.anthropic.com/research/AI-assistance-coding-skills), Jan 29, 2026.