How AI is rewriting the rules of programming and transforming the way developers work
Programming has always been about solving problems, but the way we do it is changing at a speed few could have predicted. What once required hours of meticulous typing, hunting down missing commas and forgotten brackets, now happens in minutes with the help of Artificial Intelligence agents that write, test, and review code practically on their own. This is not some futuristic prediction or a flashy corporate presentation slide. It is the real daily life of thousands of software development professionals around the world, and the signs of this transformation are showing up at companies of all sizes, from Silicon Valley giants to startups with half a dozen people.
A recent report from The New York Times, written by journalist Clive Thompson after interviewing more than 70 developers from companies like Google, Amazon, Microsoft, and smaller startups, revealed a scenario that borders on the surreal. Experienced programmers are spending their days talking to AI in plain language, scolding agents when they mess up, begging in all caps for them to follow instructions, and discovering that this emotional tone actually seems to improve results. It is weird, yes, but it works. The relationship between humans and machines in the context of coding has gained a layer of interaction that nobody really planned for, but that is becoming a core part of the daily workflow.
The case of Manu Ebert, a machine learning engineer and co-founder of the startup Hyperspell, illustrates this shift perfectly. Ebert, who grew up in the 90s learning to program line by line, now uses Claude Code — a tool from Anthropic — to delegate most of his code writing to AI agents. One agent writes the functionality, another runs the tests, and a third supervises the entire process like a virtual manager. Tasks that used to eat up an entire workday now take half an hour. Manual coding, the practice that defined the profession for decades, has taken a back seat. Ebert describes the experience as a kind of team management, where the team members are language models that need clear context, constant feedback, and, curiously enough, a bit of human patience to deliver good results.
The strange art of scolding an Artificial Intelligence
One of the most curious details revealed by the report is how developers are communicating with their AI agents. Ebert, for instance, keeps a prompt file — basically a list of commandments the agents must follow before executing any action. One of the instructions literally said that submitting code that failed tests was unacceptable and embarrassing. When Thompson asked if this emotional language really made a difference, Ebert could not prove it empirically, but reported that prompts like this seemed to slightly improve Claude’s performance.
This practice is not unique to him. Many developers today scold their agents, plead with them, repeat commands multiple times as if they were hypnotizing the machine, and write instructions in capital letters for emphasis. It might sound crazy, but it makes sense when you remember that large language models are, as the name suggests, language machines. Words loaded with urgency like embarrassing or national security imperative probably activate patterns in the model that prioritize accuracy. Conor Brennan-Burke himself, co-founder of Hyperspell alongside Ebert, mentioned research suggesting that the ruder the tone used with the models, the better they performed. Programming, which for 80 years was a quiet, solitary activity, is turning into a lively — and sometimes heated — conversation between humans and bots. 😅
Abstraction as the engine of evolution in software development
This transformation represents the latest and perhaps the most radical layer of abstraction in the history of software development. For those unfamiliar with the term, abstraction in programming is basically the process of hiding technical complexity behind simpler, more accessible interfaces. In the 1960s, the first programmers worked directly with Assembly, a language that communicates almost directly with the machine’s hardware. Every instruction was meticulous, every mistake was catastrophic, and the whole process demanded a level of attention that made the work extremely slow.
The original New York Times article provides a practical example that helps visualize this evolution. To calculate 5% interest on 10,000 dollars over 10 years, a programmer in the 1960s would need about nine lines of fairly obtuse Assembly. In Python, the same operation can be expressed in a single line: interest = 10000 * (1.05 ** 10). The programmer no longer needs to worry about where data is stored in the computer’s memory — Python handles that automatically. That is a layer of abstraction over all that memory management complexity.
In the 2000s and 2010s, the explosion of open source libraries and frameworks added yet another layer to this cake. Instead of writing everything from scratch, developers started combining ready-made pieces as if assembling a sophisticated puzzle. Thompson mentions the example of Beautiful Soup, a Python package that lets you extract text from websites with simple one-line commands, without the programmer needing to understand how all the underlying complexity works. A significant portion of the software produced worldwide works exactly like this: developers stitching together large blocks of code that someone else wrote.
Now, with generative Artificial Intelligence taking on the heavy lifting of coding, we are facing a new frontier. The programmer no longer needs to write the code — they need to know how to ask for the right code. Agents translate human intent, described in plain language, directly into functional code. There is no longer a need to mentally juggle the nuances of Python, JavaScript, or Rust. There is no longer a need to botch an algorithm and spend hours trying to figure out where the bug is. All of that has been abstracted away. And that changes absolutely everything about how we think about the profession.
Architects instead of bricklayers: the developer’s new role
Boris Cherny, head of Claude Code at Anthropic, framed the question philosophically during his conversation with Thompson: what is computing? What is coding? The answer that echoed among virtually all the developers interviewed is that the programmer now looks more like an architect than a bricklayer. The focus has shifted from execution to the big picture — how features fit together, how the system architecture should be structured, which design decisions make sense for the final product.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
Because agents produce functional code so quickly, developers can now experiment with much more freedom. Several people Thompson interviewed said they feel like Steve Jobs, who was famous for asking his team to produce multiple prototypes so he could evaluate and pick the one that worked best. The developer’s job has become more about judging and curating than about creating from scratch.
Cherny himself is a living example of this transition. As a teenager, he learned a bit of Assembly to automate math tasks on his calculator. Today, he simply picks up his phone and dictates to Claude what he wants the agent to do. One hundred percent of his contributions to the Claude codebase are written entirely by Claude itself. During his interview with Thompson, his phone sat on the table, and by the end of an hour ten Claude agents were simultaneously working on the codebase. As he put it, it is an alien intelligence that we are learning to work with.
The new programmer profile in the age of Artificial Intelligence
The report shared a detail that caught attention: many of the developers interviewed described a mixed feeling of excitement and discomfort. The excitement comes from the absurd productivity gains. Steve Yegge, a veteran who built his own tool for managing swarms of coding agents, said he is between 10 and 100 times more productive than at any other point in his career. His analogy is simple: it is as if they had been walking their entire lives and suddenly been handed an extremely fast vehicle.
The discomfort, on the other hand, comes from realizing that the manual skill of writing code is becoming less relevant with every update to these models. For professionals who spent years refining their coding abilities, watching a language model do in seconds what would have taken them hours triggers an inevitable reflection about the future of their career.
Anil Dash, veteran programmer and tech executive, offered a perspective that helps explain why developers are more optimistic about AI than professionals in other fields. In creative disciplines, LLMs strip away the most human, most rewarding part of the work and leave the tedious stuff for the professional. In programming, the opposite happens: AI takes away the boring part and leaves the creative, human part for the developer. This fundamental difference explains why surveys show that most Americans are neutral or skeptical about AI at work, while programmers tend to be more enthusiastic.
This reflection, though, does not need to be pessimistic. What is happening is a redefinition of skills, not an elimination of professionals. Knowing how to communicate clearly has gone from being a nice resume booster to being an essential work tool. Critical thinking, the ability to evaluate system architectures, product vision, and business understanding have moved to center stage in modern software development.
Greenfield versus brownfield: productivity gains are not equal for everyone
A key point from the report that often gets lost in the headlines about productivity gains is the difference between greenfield and brownfield projects. Startups and personal vibe-coding projects are greenfield — they start from zero, with no legacy code, no complex dependencies. In these contexts, the productivity gains with AI are truly impressive. Dima Yanovsky, a 25-year-old programmer at the startup Prox in San Francisco, estimates he is about 20 times faster with Claude’s help. What used to take weeks now takes hours.
But the vast majority of developers do not work on greenfield projects. They are at mature companies, with codebases that were written over years or decades and have already accumulated millions or billions of lines. Adding features quickly in these environments is generally a terrible idea — any change can conflict with another part of the code and break something that millions of customers rely on daily. At these companies, programmers historically spent most of their day planning, discussing priorities, and sitting in meetings, sometimes dedicating little more than an hour to actually writing code.
In Google’s case, CEO Sundar Pichai mentioned a 10% increase in engineering speed thanks to AI. For Ryan Salva, senior director of product at the company, that number is an average — some simple tasks became dozens of times faster, while larger changes remain slow. At startups, close to 100% of the code is written by AI. At Google, it still has not reached 50%.
Salva demonstrated for Thompson how Gemini, Google’s LLM, works in daily practice. He typed a request for Gemini to create a feature that would let users see which account they were using in the command-line interface. Gemini processed the request, explained how it planned to implement it, received approval, and worked in the background. When Salva checked ten minutes later, the code was written and the agent was already running tests — 8,000 of them, actually, well more than necessary. But it worked. 🎯
Amazon and the maintenance of the digital world
At Amazon, the brownfield scenario takes on even more dramatic proportions. David Yanacek, senior principal engineer at AWS, works in an area where servers supporting services like Netflix, Uber, and Fortnite need to run nonstop. When something breaks, someone needs to fix it — preferably before users notice. Yanacek still keeps an old pager under his monitor, a relic from the years when he was woken up in the middle of the night by incidents. Today he gets alerts on his smartphone, but the urgency remains the same.
AI is significantly speeding up this process of diagnosis and repair. During Thompson’s visit, a demo application threw an error alert — and within 11 minutes Amazon’s AI had already identified the problem, discovered that a code change had added a timestamp field that another part of the system was not expecting, and prepared a report with the suggested fix. Yanacek reviewed the suggestion, approved it, and that was it. One customer reported that Amazon’s AI agent solved a problem in just 15 minutes, while a similar failure months earlier had required eight hours of work from an entire team of engineers.
McLaren Stanley, another senior principal engineer at Amazon, recently modernized a piece of code that he himself had written years before. The original version took a month to create. With the help of the company’s internal AI, he finished the job in one morning. For Stanley, one of AI’s biggest advantages is making experimentation easier: ideas he always wanted to test now require only a six-minute conversation and a command to execute.
The impact on the job market and the next generation
From an economic standpoint, the impacts are already starting to show in concrete ways. Silicon Valley has already gone through a major wave of layoffs. During the 2010s, tech companies hired aggressively, adding an average of 74,000 new employees per year. Job openings surged in the early pandemic years and then plummeted. More than 700,000 tech professionals have been laid off over the past four years, according to data from Layoffs.fyi.
Most industry observers believe that AI was probably not the main cause of those initial layoffs, since at the time the technology was not yet good enough to replace programmers. Factors like rising interest rates, overhiring, and Elon Musk’s example of drastically shrinking Twitter’s workforce were likely more relevant. But there is evidence that AI is now eating into entry-level jobs in the field.
Erik Brynjolfsson, an economist who runs the Stanford Digital Economy Lab, analyzed sectors based on age range and AI exposure. Computer programmers were among the most exposed, and junior-level ones were hit the hardest. The number of job openings for professionals between 22 and 25 years old has dropped 16% since 2022, while older programmers have not seen a significant reduction.
Pia Torain, a software engineer at Point Health AI, felt the side effects firsthand: after just four months of intensive use of GitHub Copilot, with hundreds of prompts per day, she noticed she was losing her ability to code manually. She decided to step back for a while and now uses AI more cautiously, carefully reading the output to absorb how the code works.
Rachel Gollub, co-founder of Point Health with nearly 40 years of experience, keeps a calmer perspective. She remembers that when Python and JavaScript came along and abstracted away memory management, old-school programmers complained that it was not real coding. But reliable companies like Dropbox were built on Python and work perfectly fine. Her expectation is that the same transition will happen with AI tools.
When anyone can create software
Perhaps the most transformative effect of this new era is the democratization of software creation. Maxime Cuisy is a prime example. A production manager at a printing house in Paris that serves clients like Dior and Louis Vuitton, he has a liberal arts degree and wrote his master’s thesis on French graphic novels. He knew nothing about programming and only started paying attention to AI after ChatGPT helped diagnose an infection in his cat — something the vet had mistaken for terminal cancer.
When the company bought new printers and ran into problems with the margin adjustment software, Cuisy decided to use Codex, OpenAI’s coding tool, to build a solution. He spent a few hours detailing the specifications, and by the end of the day he had a working application for Mac and Windows that processes up to 2,000 images at once. He has no idea how the code works — it is written in Python, which to him might as well be ancient Greek.
This kind of situation illustrates what Brynjolfsson calls the Jevons paradox applied to software: when something becomes cheaper to make, we do not simply save — we make more. Small businesses that would never have been able to afford a team of programmers can now hire a single AI-assisted professional, or even solve internal problems on their own with vibe-coding. The world is likely to be flooded with far more software than ever — built by individuals, for individuals.
What to expect going forward
The abstraction movement we are living through right now is different from previous ones in one fundamental way: speed. Earlier transitions in the history of programming took decades to consolidate. The migration from Assembly to high-level languages happened over the course of 20 or 30 years. The widespread adoption of open source frameworks and libraries took at least a decade to become an industry standard. Generative Artificial Intelligence, on the other hand, is reshaping software development practices in a matter of months. Tools that did not exist two years ago are now an integral part of entire teams’ workflows, and the models are improving at a rate that makes any long-term prediction risky.
This speed brings real challenges. University curricula and programming bootcamps need to be rethought to prepare professionals who will work with tools fundamentally different from what exists today. Companies need to invest in continuous training to keep their teams up to date, and professionals themselves need to develop a mindset of constant learning that goes far beyond mastering a new language or framework.
Not everyone is happy with this transition. Some developers object to the energy consumption required to train and run the models. Others question the ethics of training LLMs on copyrighted works. Some fear that AI’s speed will result in mountains of poorly written code that will not perform well in the long run. And there is the legitimate concern that bosses will use agents as a threat: do not complain, or we will replace you with a bot.
Thomas Ptacek, a developer in Chicago and co-founder of Fly.io, described the current landscape as a civil war between those who embrace AI and those who reject it. He positions himself in the middle, acknowledging that the resisters are kidding themselves when they claim the technology does not work, but also admitting that the impacts on the profession could be deeply negative for many people.
Kent Beck, an industry legend who has been programming since 1972, offers a more optimistic perspective. He had stopped programming ten years ago, frustrated with the available tools. LLMs brought him back, and now he is producing more projects than ever. For him, even AI’s unpredictability has its charm — asking for the same code twice and getting different answers is addictive, like a slot machine. 🎰
At the end of the day, the story of software development has always been a story of increasing abstraction, of pulling the professional away from mechanical details and closer to the decisions that truly matter. Artificial Intelligence is the next chapter of that story, and all signs point to it being one of the most transformative. What remains for the human in this equation is precisely what machines still cannot replicate with consistency: the ability to imagine what needs to be built, evaluate whether the result makes sense, and make decisions based on context, experience, and intuition. As Clive Thompson observed at the end of his extensive report, this mix of exhilaration and anxiety that programmers feel today may be a preview of what awaits professionals in many other fields. In any job that involves language and information, this combination of rhetoric, systems thinking, and skepticism toward a bot’s output could become the fabric of intellectual work. Abstraction may be coming for all of us. 😉
