I'm writing this on a quiet December evening, looking back at what has been, without exaggeration, the most transformative year of my career. Twelve months ago I stayed up until 3am talking to ChatGPT for the first time. Today, AI is woven into every part of how I build software. I want to capture what this year felt like before the memory fades into routine.

January: The Lightning Strike

ChatGPT had launched in late November 2022, but January is when it hit me personally. I wrote about it at the time and re-reading that post now is fascinating. The excitement was raw and unfiltered. I wasn't wrong about the significance, but I was wrong about the timeline. I thought it would take years for AI to become a practical daily tool. It took weeks.

By the end of January, I'd replaced Stack Overflow as my first stop for coding questions. That shift happened faster than I expected and it stuck. I still use Stack Overflow, but it's my fallback now, not my default.

February-March: The Experimentation Phase

This is when I tried everything. Midjourney for image generation (I shared my first Midjourney experiments around that time). Copilot for code completion. The OpenAI API for building custom tools. I was like a kid in a candy store. Every week there was something new to try, and most of it was genuinely useful. Not gimmicky. Not "cool demo." Actually useful for real work.

I built my first AI-powered project during this period: a code review bot that reads GitHub PRs and leaves comments. It was 300 lines of code and it worked. The barrier to entry for building AI applications was shockingly low.

April-June: Finding the Workflow

The initial excitement settled into routine. Not boring routine. Productive routine. I found my workflow: ChatGPT for planning and debugging, Copilot for writing code, AI for tests and documentation. Each tool found its place and I stopped constantly experimenting and started consistently producing.

This was also when I got more critical. I wrote about AI-generated code quality and hallucinations. The tools are amazing, but they're not magic. They write bugs as fast as they write good code. The skill is knowing when to trust them and when to verify. I got burned enough times to develop good habits around this.

LangChain was the hot framework and I spent time with it. My conclusion: powerful but overly complex for most use cases. Sometimes the best approach is just calling the API directly. Not every problem needs a framework.

July-September: Deep Integration

By summer, AI wasn't a tool I used. It was part of how I think about development. Planning a feature? Talk it through with ChatGPT first. Writing code? Copilot is always there. Stuck on a bug? Paste the error into ChatGPT before doing anything else. Writing tests? Generate the scaffolding with AI.

I built the automated code review GitHub Action during this period and it's been running on every PR since. I explored vector databases for a RAG project. I compared Midjourney v5 with DALL-E 3. The tools kept getting better and my usage kept deepening.

The productivity improvement over my pre-AI workflow is real. I estimate 30-40%. Some of that is writing code faster. But most of it is spending less time stuck, less time on tedious tasks, and less time context-switching between coding and searching for answers.

October-December: The New Normal

The hype cycle for AI has passed the peak for the general public. People who never used ChatGPT are back to not caring. But for developers who integrated these tools into their workflow, this is just the beginning. The tools are better in December than they were in January, and the rate of improvement isn't slowing down.

I fine-tuned a model on my codebase. Mixed results, but I learned a lot about when fine-tuning makes sense versus when better prompting is enough. I've been tracking what AI tools I actually pay for and the total is about $70/month. Best money I spend.

What I Got Wrong

I thought AI would mostly help with writing code. The bigger value turned out to be in everything around writing code: planning, debugging, refactoring, testing, documentation. The code generation is nice but it's maybe 30% of the total value.

I thought the tools would plateau quickly. They didn't. GPT-4 is meaningfully better than GPT-3.5. DALL-E 3 is dramatically better than DALL-E 2. Copilot's suggestions have improved noticeably over the year. The improvement curve is steep and ongoing.

I underestimated how much prompt engineering matters. The gap between a lazy prompt and a carefully crafted one is enormous. This is a real skill that takes practice. I'm still getting better at it.

What I Got Right

My January post said "everything about how we write software is about to change." That was right. Not in the dramatic, science-fiction way that some people predicted (no, AI hasn't replaced developers). But in a practical, daily-workflow way that's just as significant. The process of building software in December 2023 is fundamentally different from the process in December 2022.

I also correctly predicted that the developers who benefit most would be the ones who started building habits early. The learning curve for effective AI-assisted development is real. People starting now are months behind people who started in January. Not because the tools are hard to use, but because using them well requires intuition that only comes from practice.

Looking at 2024

I think AI coding tools will get codebase-level context. Not just the current file, but understanding of your entire project. That will be the next big jump. I think the distinction between "code completion" and "conversation" will blur into a single integrated experience. And I think we'll see AI handle more of the development lifecycle: deployment, monitoring, incident response.

What I don't think will happen: AI replacing developers. The tools are making good developers more productive. They're not turning non-developers into developers. The skills that matter are shifting (prompt engineering, code review, architecture) but the need for skilled humans in the loop is as strong as ever.

2023 was the year AI went from a curiosity to a daily tool for software development. I'm grateful I was paying attention when the lightning struck. Whatever comes next, this was the year the ground shifted, and I felt every tremor.