How I Actually Got Started With AI: Six Months of 2AM Nights

For about six months, I was up until 2am five nights a week.

My office. Headphones on. Laptop open. Usually music I'd stopped actually listening to hours ago. I'd grind until I couldn't see straight, close the laptop, sleep for a few hours, get up, do the day, come back to the desk at 9pm. Repeat.

This is the honest answer when people ask me how I got started with AI.

Not a course. Not a certification. Not a cohort. Just a lot of late nights, a lot of curiosity, and — if I'm being honest — a wife who, for a little while, was pretty sure I was losing it.

The moment it actually clicked

The first prompt I ever typed into ChatGPT? No idea. I genuinely don't remember.

The moment I remember is different. I was using it the way most people were at first — writing, summarizing, brainstorming. At some point I asked it to generate code. Not a snippet to paste somewhere for decoration. Actual working code that did a thing.

It wrote it. And it ran.

That was the oh-shit moment. Up until then I'd always looked at building software as something I'd either need to spend years learning the craft of, or outsource to someone who had. Suddenly the friction was gone. I could describe what I wanted, watch it generate something, test it, iterate. I didn't need a developer in a Jira queue. I didn't need to block out three months for a Coursera course. I could just… build.

That's what unlocked the bender. It wasn't "AI is cool." It was "I can finally make the things that have been living in my head."

The first thing I actually shipped

The first real project — not a toy, something I actually used — was a news scanner.

I wanted to see what competitors of a company I was tracking were announcing. The obvious way to do that is to read every press release, every article, every investor update. Which is a lot. Nobody actually does it. You skim headlines, miss the signal, and hope someone else surfaces the big stuff.

So I wired it up differently. I connected to a news API. Pulled ten articles per competitor. Fed everything into a summarization pipeline that produced a single cross-company TLDR — the kind of brief you could actually read over coffee.

Sixty-plus articles in. One page out.

The mechanical parts were the easy part. APIs, chunking, prompt templates. What was hard was getting the output I wanted. The first versions spit out bullet salad. Every summary sounded like every other summary. No hierarchy. No "here's what matters." It took iteration — probably a hundred passes — to get it to a place where I'd read the output and actually learn something I didn't already know.

That was my first taste of the real work. It wasn't writing the code. It was getting the model to do useful work reliably.

The 2am routine

Picture it: office. Door closed. Headphones. Some kind of music playing that I'd stopped noticing hours ago. I'd lose entire evenings just trying things — new prompts, new APIs, new ways to wire services together.

At some point I blew through ChatGPT's token limits. I remember this vividly, because it was the first time I felt like a real user. I upgraded. Then, a bit later, I started playing a weird little game with myself: how fast can I hit the max on my new plan?

I'm not kidding. I still play that game. Every time a new model or a new tier drops, the first week I'm genuinely curious how much of it I can burn through. It's a dumb little metric, but it's also an honest one — it tells me how much I'm actually shipping versus just thinking about shipping.

Where the money actually went

I want to be specific about this because I think it matters.

I never paid for a single course. Never took a bootcamp. I taught myself everything — Python, APIs, RAG, agentic patterns, the orchestration stuff that makes things work in production — by reading docs, watching models fail, and debugging until they didn't.

Where the money went was infrastructure.

Early on: ChatGPT Plus, then the higher tier. API credits on top of that. Third-party APIs — news feeds, data providers, anything I needed to plug into. Hosting. It adds up faster than you think when you're running six experiments in parallel.

Lately it's gone further. I've spent thousands standing up a home server so I can run local models. I run Mac Studios tuned for inference. I've got a setup at home that would've felt absurd three years ago and now feels like a practical requirement for doing the kind of building I want to do.

The pattern, if there is one: every dollar I spent went toward building something or running something. Not watching someone else do it.

The thing that was actually hard

Here's the part I want my past self to hear.

Most of AI is not hard. I mean that. Reading the API docs, making a call, getting a response, pulling context through RAG, calling tools from a model — all of that is mechanical. It's engineering, not magic. A weekend of focused work will get most people there.

What's actually hard is reliability.

Getting an AI system to do the right thing once is easy. Getting it to do the right thing every time, across every input, across weeks of changing data and model updates — that's the hard part. That's where most projects die. Not at "does this work at all" but at "does this work consistently enough that a real person can depend on it."

I spent a long time on that. Building agentic workflows, watching them do something brilliant on Tuesday and something nonsensical on Wednesday, trying to understand why. Where the seams were. Which part of the prompt was too permissive. Which tool was returning noisy data. Which fallback wasn't firing when it should.

That skill — debugging agent behavior, not code — is the one I'd argue actually separates people who build toys from people who ship products. And it's the skill that takes the longest to develop, because you can't learn it from a video. You can only learn it by having an agent disappoint you enough times that you start recognizing the shape of the failure.

What people get wrong about "getting started"

Every week someone asks me how to get started with AI. Every week I give the same answer, and every week I watch their face fall a little.

The answer is: stop reading about it and go build something.

The advice I hear constantly that I disagree with: "Take this course. Watch this YouTube series. Read this roadmap."

I'm sure those things are fine. I'm sure some of them are even good. But they're also a way of feeling productive without actually producing anything. You can watch ten hours of someone else wire up an API before you've written a single line yourself. You can collect certifications for six months and still not have shipped a working prototype.

The gap between knowing about AI and using AI is bigger than it looks. And the only way across is hours. Real hours, with your own hands, on your own projects, where you have skin in the game and a clear picture of what "working" would even look like.

Pick something small. Something you actually want to exist. A newsletter summarizer. A tool that answers a question you keep Googling. A personal assistant that knows your calendar. Anything. Go build it badly. Ship v1. It'll be embarrassing. That's fine. The embarrassing v1 is the whole point.

The people I know who've actually gotten good at this didn't watch their way in. They built their way in.

The Bill Gates quote

There's one moment from those first few months I think about a lot.

My wife, at some point in there, was pretty sure I was losing it. That's fair. I was up at 2am every night, staring at a laptop, obsessed. From the outside, you couldn't really tell if what I was doing was going to amount to anything. I couldn't tell you either. I just knew I couldn't stop.

What helped me trust the obsession was a Bill Gates quote I read around that time. I won't butcher it by paraphrasing it too closely — the gist was that AI was going to be bigger than the internet. Not different. Bigger.

My dad worked for Microsoft. I grew up with a baseline respect for Gates and for how he thinks about where things are going. When someone with his track record puts a stake in the ground that hard, I pay attention. That quote was the moment I stopped asking myself if this was a phase and started asking myself how much bigger I could go.

The NVIDIA trade I didn't make

I'll close with this because it still bugs me.

I remember sitting at that same desk, probably in January of that stretch, watching NVIDIA trade around $120. I don't actively trade stocks — I put money to work in other ways — but I remember thinking, very clearly: I should put $50,000 into this.

I told a buddy. He did.

I didn't.

I look at where that stock is now and I laugh, because the part of me that was grinding at 2am every night of that period knew exactly what was coming. And the part of me that wasn't grinding was what stopped me from acting on it.

There's a lesson buried in there somewhere. Probably the same lesson as the rest of this post. The hours you put in don't just teach you about the thing you're building. They teach you about what's coming next.

Still, every night, the laptop comes open. The headphones go on. The game is the same: how long can I hold out before I hit the max?

Three years in, I'm still playing.