For many of us, ChatGPT was the first time we became acquainted with AI. By now, I think it's safe to say that we have all felt the impact. Does anyone remember all the quirks of the initial OpenAI GPT-3.5?
The hallucinations? The endless apologetic tone? How people used ChatGPT to turn everything into a rap - and all the interesting jailbreak methods involving hilarious roleplay scenarios?
Despite ChatGPT being less than 4 years old, we have lived through 3 eras already.
The auto-complete era
Remember just prior to coding agents taking over, we had the auto-complete period powered by Cursor & GitHub Copilot? It was a short-lived period - from 2023 to 2024 - but we all thought it was amazing.
Especially when you had a block of code suggested that was identical to what you were just about to write - and it was only a tab away!
The agentic coding era
Then in 2025 we got agentic coders - Anthropic released Claude Code in February, followed by OpenAI releasing Codex 3 months later.
They started as apologetic assistants. They felt like junior developers that displayed a remarkable inconsistency in their abilities to deliver. Sometimes they would suggest an event-sourced system with a complex reducer as an elegant solution to a complex problem. Other times, they would try to run npm run dev without installing the dependencies - and spiral into increasingly elaborate fixes, instead of simply running npm install.
They were entertaining to work with (although frustrating from time to time) - but overall, a fascinating new way to code.
But they quickly evolved, and I think Claude Sonnet 4 was a major milestone for quality of agentic coding. It still had its flaws, but overall it felt more senior in the way it would approach issues.
The loop engineering era
While many people are still using the agentic coding approach in 2026, the new trend is the construction of agentic loops. Boris Cherny (Anthropic) helped popularize this approach.
Being able to prompt an AI to solve tasks is amazing - but it still requires prompting. If you could bypass this by building autonomous loops that are triggered by events or scheduling - you would never have to prompt an AI directly.
The potential for scaling is enormous - provided you (or the AI!) can verify the results.
The next era?
It's wild how far this technology has come in such a short time. If this trend continues, I think in a year we will all look back at the models that are currently considered "state of the art" from Anthropic, OpenAI and Kimi, and wonder why we were ever impressed with them at all.
Personally, I'm excited to see what is next to come. Fingers crossed that it is not the total extermination of the field of software engineering. Despite my enthusiasm about AI these days - I still really like to write, read and understand code myself.
If this post left you feeling nostalgic, you have a final chance to play with GPT-3.5 via OpenAI's API - but not for long. The gpt-3.5-turbo-1106 model (which to be fair is the later 2023 version of gpt-3.5) is shutting down on September 28, 2026!
Top comments (0)