I don’t write code without artificial intelligence help!
On March 10, 2025, Dario Amodei announced one of the most well-known forecasts — that within the next 3-6 months, about 90% of software code would be created by AI, and within a year — the entire software industry overall. Of course, many considered this to be marketing hype, and Amodei himself later joked: in the fall of that same year, he stated that approximately 70-90% of code in companies like Anthropic and others is created with AI assistance. So, his words were partly justified — and if someone still writes code manually, it’s already their personal problem.
But what is happening outside of Anthropic? I did some digging into the materials and encountered a main problem — research in this area quickly becomes outdated, often even during publication. To understand this properly, one would need to design an experiment, recruit control and experimental groups, observe for some time, then collect and analyze data, and have it reviewed by independent experts. That’s why it’s understandable that serious scientific publications often compare GPT-4 with Gemini 1.5 — but such comparisons require enormous time and resource investments.
However, there is one interesting find. In early 2025, METR conducted an experiment involving 16 experienced open-source developers. They were given tasks, and researchers randomly determined whether the tasks would be solved by humans or AI. The main tools used were Cursor with Claude 3.5/3.7. Participants were paid about $150 per hour — roughly what an experienced engineer earns.
Initially, everyone hoped for a work speed increase of around 24%, but ultimately, participants said that AI helped them increase their speed by approximately 20%. However, METR’s measurements revealed something surprising: in reality, completing tasks with AI was faster by only 1–2 minutes — a reduction of about 19% in task completion time.
By fall of the same year, they involved 57 developers: ten from the previous experiment and forty-seven newcomers from various projects of different complexity. Budget constraints limited pay to $50 per hour — much lower than average market salaries, which the organizers believe clearly affected the results.
The outcome was quite the opposite of expectations: many participants simply refused to perform tasks without AI assistance — because they found it uninteresting or didn’t want to spend much time on old-fashioned coding. Here’s a vivid quote:
“I avoid tasks where AI can do in a couple of hours what would take me a whole day.”
After analyzing the collected data, it became clear: for those participating for the first time, acceleration was about 18%, while for newcomers — only around 4%. The statistical significance of these figures has faded into the past, but the main conclusion remains important: within just half a year, experienced developers had become so accustomed to using AI for their tasks that they could no longer imagine working without its help. Here’s another quote:
“Working the old way is like walking across a city on foot when you’re used to taking an Uber.”
An interesting observation from this experiment is that assessing efficiency has become more difficult not only because measuring actual acceleration is challenging but also because many developers launched multiple AI agents simultaneously to solve different tasks. For example, I heard about Jeff Emmanuel — he pays for 22 Claude Max subscriptions and uses an entire team of AIs to automate his work: making around two to three thousand commits per week on GitHub. This approach is more of an exception than the rule, but anyone can run several agents for different projects.
In the end, experts acknowledged that they are confident AI indeed accelerates programming. But proving this with precise numbers proved impossible — technological development is so rapid and complex that measurement is nearly unfeasible. Currently, a new study on this topic is underway; however, it’s unlikely to show fundamental results — because progress is so fast and multifaceted that capturing it with quantitative metrics is almost impossible.
The most interesting aspect is that automating programming leads to a whole range of other development processes. Modern systems like Claude Code or Codex can already write almost any tool for nearly any task given good skills. Perhaps Amodei overestimated the accuracy of his figures — but he correctly guessed the direction of development.
That’s the story about neural networks and code!
Created with n8n:
https://cutt.ly/n8n
Created with syllaby:
https://cutt.ly/syllaby