Whoever has tried it knows it. Programming with AI can be wonderful. Especially if you have (almost) no idea about programming. This is where generative AI models have seen their first and probably only revolution.
The developers were the first to be able to embrace this new technology. The appearance of GitHub Copilot in 2021 It showed us that it was no longer necessary to chop so much code, because the machine was already doing it for you, and since then the advance of generative AI in the field of programming has been overwhelming.
The question is: has it been positive? The answer is not at all clear. It is evident that AI has allowed:
- That millions of people who were not programmers could turn their ideas for applications and games into a reality.
- That millions of professionals can save time by not having to write repetitive code (boilerplate) to focus on other more important and productive parts of your work
The industry, of course, has been especially insistent with this vision of the transformation of this segment. Satya Nadella (CEO of Microsoft) and Sundar Pichai (CEO of Alphabet/Google) already boasted months ago that about 25% of the code generated by their companies is generated by AI. Meanwhile, Jensen Huang went further and made it clear that At this point no one should learn to program anymore because the AI would do it for us.
These are very forceful statements, but behind them lies another reality: that All that glitters is not gold in the world of AI for programmers. At MIT Technology Review they have spoken with more than 30 developers and experts in this field and have reached interesting conclusions.
AI is a better programmer than ever. At least, according to the benchmarks
In August 2024 OpenAI made a unique launch: presented SWE-bench Verifieda benchmark intended to measure the ability of generative AI models to program. At that time, the best of the models was only capable of solving 33% of the tests proposed by that benchmark. A year later the best models already exceed 70%.
Current ranking of the best models according to the SWE-bench Verified benchmark. Several already pass 70% of the tests. Source: SWE-bench.
The evolution in this area has been dizzying and we have witnessed the birth of that new modality programming called “vibe coding” and all the big ones have developed powerful programming tools to take advantage of the pull.
We have OpenAI Codex, Gemini CLI, or Claude Code, for example, but they have been added startups like Cursor either Windsurfing who have also known how to take advantage of this fever for programming with AI.
All of these tools promise basically the same thing: that you will program more and better. Productivity theoretically skyrockets, and while more code is certainly being written than ever thanks to AI, programmers They have gone from writing their own code to reviewing what machines generate.
Recent studies reveal that veteran developers who believed they had been more productive actually they weren’t. Their estimate was that they had been 20% faster by being able to move forward without blockages, but in reality they had taken 19% longer than they would have taken without AI, according to the tests carried out.
There is another problem too: code quality is not necessarily goodand as we say, developers must review that code before being able to use it in production. In the latest survey from Stack Overflow, one of the largest developer communities in the world, there was a notable fact: The positive perception of AI tools had decreased: it was 70% in 2024, and 60% in 2025.
There are limitations, but even so everything has already changed
Those interviewed by MIT Technology Review generally agreed with its conclusions. Generative AI programming tools are great for producing repetitive code, writing tests, fixing bugs, or explaining code to new developers.
However, they still have important limitations, and the most notable is his short memory. These models are only capable of handling a fraction of the workload in professional environments: if your code is large, the AI model may not be able to “consume” it and understand it all at once. For small projects, great. For large developments, probably not so much.
The problem of hallucinations also affects the code, and in repositories with a multitude of components, AI models can end up getting lost and not understanding the structure and its interconnections. The problems are there, and they can end up accumulating and causing exactly the opposite of what they wanted to avoid.
Several experts, however, explained in that text how it is actually difficult to go back. Kyle Daigle, COO of GitHub, explained that “the days of coding every line of code by hand are likely behind us.” Erin Yepis, an analyst at Stack Overflow, indicated that although this unbridled optimism towards AI has fallen somewhat, that is actually a sign of something else: that programmers embrace this technology, but they do so assuming its risks.
And then there is another reality. One that is repeated day after day and that seems undeniable. The AI we have today is the worst of all those we will have in the future. It may not be tomorrow or next week, but it is clear that the AI you program will end up getting better and better. And there may come a point when those limitations disappear. Whether they do it or not, what is clear is that AI has changed programming forever.
Image | Mohammad Rahmani
In Xataka | OpenAI has turned ChatGPT into mainstream AI. In the business world the game is being won by its great rival
GIPHY App Key not set. Please check settings