In March 2023, ChatGPT was barely four months old and everyone — including me — was trying to figure out what it meant. Three years later, I've written code with it, built tools on top of it, and watched it turn from a curiosity into the default way I work. This is the story of that change, told through the posts I wrote at the time and what I'd tell my 2023 self now.
March 2023: "Why ChatGPT Matters" — it's not a brain, it's an interface
Back then the debate was whether ChatGPT was a step toward AGI. I argued the opposite: the real paradigm shift wasn't an all-knowing AI, it was the interface. For the first time, a digital system understood natural language. Not "press 1 for invoices" — actual language, in many languages, with context and follow-ups.
That take has aged well. The models got smarter, but the thing that changed everything was that you could finally tell a computer what you wanted. Every assistant, every copilot, every agent since then is a direct consequence of that moment. The "AI as interface" framing turned out to be the boring, accurate version of the story — while the AGI hype was the exciting, wrong one.
April 2023: "Developing with ChatGPT" — learning by building, with AI as teacher
That same spring I decided I had to learn this stuff seriously, not with courses but with a project: a personal assistant CLI in Python — a language I didn't master at the time. The rules were simple: use ChatGPT instead of Google whenever possible.
It was the ideal use case: a small greenfield project, a CLI (no frameworks or architectures to fight), a language I wasn't expert at (perfect for asking questions), and a domain that was ChatGPT's specialty. ChatGPT wrote my first README, a working Dockerfile, the first TTS integration, refactors, tests. I learned Python by reading its answers critically — accepting, editing, or discarding them. The meta part delighted me: using ChatGPT to build a tool that uses ChatGPT, to eventually replace my own ChatGPT.
At that moment explaining what you wanted and selecting the right context required more work, you had to add everything to the prompt by hand, so if you know what to do, it was often easier snd faster to do it yourself, the lesson of the moment: LLMs are force multipliers when you're out of your comfort zone, and near-useless when explaining is harder than doing. In my day job, where I knew the codebase and the language, I barely used it. The magic wasn't the model — it was the combination of an unfamiliar problem, a small scope, and a language model as instant context.
November 2023: "LLMs One Year After" — the hype cooled, the tools got real
Nine months later, no AGI — but GPT-4 had gained multimodality, a code interpreter, plugins, web access. Every company from Adobe to Notion had bolted an assistant onto their product. Nobody had lost their job to a chatbot in visibly angry hordes. But the tools were quietly becoming genuinely useful.
What struck me then — and what I'd still defend — was the noise. Most content about LLMs was (and is) hype or filler. The signal was in a handful of serious sources: Jeremy Howard's "A Hacker's Guide to LLMs", Data Skeptic's Machine Intelligence series, Machine Learning Street Talk. The same is true today: the field moves fast, but the compounded knowledge comes from people doing the work, not from hot takes.
2026: where we landed
Fast-forward to now. The trajectory I guessed at in 2023 — everyone ending up with a personal assistant that knows their projects — happened, just with better names: Copilot, Claude Code, Cursor, and the quieter family of agent harnesses. I now orchestrate sub-agents for real work and the "assistant that knows my projects" is no longer science fiction.
Some things I got wrong, or at least underestimated: I thought the bottleneck would be model capability. It's not — it's context, trust, and workflow design. Knowing when to delegate to an agent and how to verify its output matters more than the model's benchmark score.
And the job question I wrestled with in 2023? My conclusion hasn't changed: The developers who treat LLMs as a multiplier, not a threat, are the ones writing the next three years of this story.
What I'd tell my 2023 self
- The interface insight was the right one. Natural language as a programming interface compounds; model intelligence alone doesn't.
- Build something with it. Courses teach you tools; a project teaches you judgment. My CLI taught me more Python than any tutorial.
- The hype is noise. The useful stuff is boring: context, verification, workflow.
Top comments (0)