I’ve been thinking about how quickly the world of software development is changing. During the winter holidays, I read an article by Gergely Orosz describing his experiences with new artificial intelligence models. He personally tested how models such as Opus 4.5 and GPT 5.2 can write hundreds of lines of code based solely on instructions. He even did it from his phone while traveling—he instructed AI to make changes to code on GitHub, run tests, and prepare pull requests. All without writing any code manually himself. It feels like magic, but at the same time, it makes you wonder what it will do to our jobs.
Gergely Orosz mentioned that these models handled medium-sized tasks extremely well. For example, he added self-service group subscription features for large companies and modified the admin panel for his newsletter, The Pragmatic Engineer. He deployed all of it to production after reviewing and testing it. And he is not alone—many other developers have shared similar stories. It feels like a turning point when AI stops being merely an assistant and becomes the primary creator of code.
"Aha" moments from experienced developers
When you read what people like Jaana Dogan from Google are saying, it is clear that this is not just hype. Jaana described how she gave Claude Code a description of a problem involving distributed agents, and the AI solved it in an hour—something her team had spent an entire year building. She said it is not perfect, but it is amazing for complex tasks in a domain you understand. Similarly, Thorsten Ball from Amp admitted that after 15 years of programming, writing code manually now frustrates him. He used to love the rhythm of typing, but now he sees that AI does it faster.
Malte Ubl, CTO at Vercel, shared how he built two open-source projects, wrote a book, and fixed many things during the holidays—all thanks to Opus 4.5 and Claude Code. He said that AI behaves like a senior developer who only needs to be told what to do and then does it with minimal supervision. David Heinemeier Hansson, the creator of Ruby on Rails, has shifted from his view in the summer, when AI was not yet good enough, to optimism now. And Adam Wathan, the creator of Tailwind CSS, says that writing syntax manually is now boring, while programming with AI is more fun.
The biggest reversal came from Andrej Karpathy, a co-founder of OpenAI. As recently as October, he called AI tools "slop"—poor-quality stuff. But in December, he wrote that he had never felt so far behind as a programmer. According to him, the profession is changing because the programmer contributes only small pieces while AI does most of the work. He talks about a new layer of abstraction involving agents, prompts, contexts, and tools that must be mastered. Boris Cherny, the creator of Claude Code, added that AI wrote all of his code last month—about 200 pull requests, without a single line written manually.
Why now?
I’ve been wondering why this is happening at this particular moment. According to the article, it was caused by new models released at the end of the year: Google’s Gemini 3 on November 17, Anthropic’s Opus 4.5 on November 24, and OpenAI’s GPT-5.2 on December 11. These models are significantly better at generating code. Peter Steinberger, the creator of PSPDFkit with 20 years of experience, said that the leap from GPT 5.1 to 5.2 was enormous—AI can now usually handle tasks on the first attempt without frequently getting stuck.
Simon Willison, an expert on large language models, confirms this: November models such as GPT-5.2 and Opus 4.5 crossed a threshold that opened up more complex problems. Even Salvatore Sanfilippo, the creator of Redis, noted that AI is good at C as well. And do you remember the prediction made in March by Dario Amodei, CEO of Anthropic? He said that within three to six months, AI would write 90% of code, and within twelve months, almost all of it. Now it looks as though it is already happening.
Gergely Orosz himself says that AI generates 90% of the code for his projects using TypeScript, Node/Express, React, and Postgres. The same applies to Go, Rust, or other languages for which AI has enough data. He expects this to become common among startups searching for product-market fit or new projects without legacy code.
Downsides: What is losing value?
Now comes the less pleasant part of the discussion. Some skills that used to be essential are becoming less important. Take prototyping, for example—platforms such as Lovable or Replit allow non-technical people to build apps themselves. Gergely mentioned a Replit ad featuring basketball player Shaquille O’Neal, who "vibe-coded" an app for collecting pickup lines without any prior experience.
Being an expert in multiple languages, such as Go or Rust, will become less important because AI will make it possible to jump into any codebase and have it implement things. Specialized roles such as frontend or backend developer may disappear at startups—one good engineer who uses AI for everything may be enough. Implementing well-defined tickets from JIRA or Linear? AI can handle that on its own, just like code refactoring.
Peter Steinberger admits that he no longer reads most of the generated code—he only looks at the structure and key parts. He chooses languages such as TypeScript for the web, Go for the command line, or Swift for macOS because AI excels at them. But he warns that review is still necessary for mature software or security-related work.
Benefits: Greater value of real engineers
On the other hand, it occurs to me that real software engineers may become even more valuable. The article begins with this point, but the details are behind a paywall, so I am inferring them from the introduction. Gergely suggests that the skills of tech leaders will be in greater demand, as will a product-oriented mindset at startups. Being a solid engineer, not merely a "coder," will become the standard. AI writes code, but decisions about architecture, testing, and overall design will remain with humans.
I’ve been thinking about how this will affect product managers versus engineers. Product people can now generate software themselves with fewer engineers, but engineers, in turn, need less product guidance. The professions will begin to overlap. And what about the ugly aspects? More generated code means more problems, weak practices will become apparent faster, and developers may face a worse work-life balance. It is like an earthquake, as Andrej Karpathy says—we have to adapt so that we do not fall behind.
