Field Notes #001 — The first entry in a public lab notebook. Building an AI operating system across three businesses, in real time, with all the wrong turns left in. — Dakota

You can’t learn AI by watching videos about it.

You learn it by using it. Daily. Every release. Breaking things. Switching stacks. Playing.

Most people are still treating AI like a smarter Google. I spent 18 months figuring out it’s something completely different — and yesterday a bunch of stuff finally connected.

I get one of these realizations maybe once a month and usually let them slip. This one I want to actually write down.

Quick caveat before I get going: some of this is going to be day-one stuff for the people deep in this. Some of it is going to sound like alien tech to people who haven’t touched Claude yet. AI is funny that way. Everybody’s on a different mile of the same road.

I’m not talented. I’m just passionately curious.

That’s something I figured out about myself a while ago. I’m not the smartest person in any room I walk into. Never have been.

What I’m actually good at is asking questions. Why is it this way? Why can’t we do it like that? What would happen if I tried this? That instinct is the only thing that’s pulled me through anything I’ve ever been decent at.

I don’t really know how to code. And honestly, for the longest time I didn’t even really know what “code” was. Here’s the version I finally landed on: writing code is just typing very specific instructions that tell a computer exactly what to do, step by step. That’s it. The reason it looks intimidating is the language — every kind of computer has its own dialect — but the idea underneath is just do this, then do this, then if this happens do that.

What I’m actually decent at is poking at things out loud until they make sense. And that turns out to be the exact skill AI rewards right now. Because if you can describe what you want clearly enough in plain English, the AI writes the code for you. You just supervise.

People say I don’t know anything about AI. Dude — nobody did. Just ask the AI. That’s the whole game.

The voice AI rabbit hole

Here’s a small example of what I mean.

About a year ago I was thinking about voice AI. Honestly, I’d written it off. Everything I’d heard sounded robotic — that uncanny press 1 for billing energy. I didn’t think we were anywhere close to something that could actually sound like a real person.

So one night I just asked. If I wanted to build a voice AI that sounded exactly like me, what would that even look like?

It walked me through the whole thing. You’d clone your voice in ElevenLabs (a tool that learns the sound of your voice from a short recording and can then say anything in it). Wire it to a phone number through Twilio (a service that lets software send and receive calls and texts). Hook it to a language model for the brain — that’s the part that actually thinks of what to say. And then — here’s the part I didn’t realize — the whole thing has to round-trip in under a second or the conversation feels broken. The latency is the wall. Not the voice. Not the intelligence. The speed at which all three can talk to each other.

That was the moment I stopped wondering why isn’t voice AI here yet and started understanding why it’s been so hard. And now, a year later, watching that latency wall basically fall apart in real time, I keep thinking — oh my god, we’re so close.

That’s the whole pattern of how I learn this stuff. Ask a dumb question. Get a real answer. Now you understand the actual problem instead of being confused about why it isn’t solved yet.

The lead reactivation era

About 18 months ago I was using ChatGPT. That was it. I’d open the app, type something, get an answer, close the app. Useful, but mostly a glorified search engine for me at that point.

Then I saw people building these AI lead reactivation setups. The pitch was — take all your old dead leads, fire AI texts at them all day, have the AI hold full conversations, surface the warm ones back to a human.

The mechanics: a text comes in, gets passed through a tool called Zapier (which is basically a middleman that connects apps that don’t natively talk to each other), then to ChatGPT to write a reply, then back through Zapier to send the response. The whole chain held together with duct tape.

I bought a little course. They told me to set it up inside something called GoHighLevel. I had genuinely never heard of it. I was clicking around in there with no idea what I was looking at. Took me a minute to even realize what I was in. Like — oh, this is a whole category of software. CRMs, automations, pipelines. HubSpot kind of stuff. I just didn’t know that world existed.

So I taught myself. Stumbled through it. Watched videos. Broke things. Eventually got the AI texting flow working. And it did work. People were selling these as a service for real money. Some of them still are.

But two things bugged me.

One — it felt clunky. Four tools to send one text reply. Every new use case meant another connection, another setup, another thing to babysit. Rube Goldberg machines to do something basic.

Two — I couldn’t see myself being proud of selling that. Spraying AI texts at a pile of cold leads to manufacture conversations? Felt off. Worked for some people. Wasn’t the business I wanted to be in. Lost interest pretty quick.

The Gemini moment

So I kept playing. Reading. Watching releases. Testing things on my own businesses instead of trying to sell anything.

Then November hit. Gemini dropped their actual pro model.

The way I noticed was kind of random. I’d been pushing ChatGPT hard on long PDFs and dense contracts — I look at a lot of contracts. I knew exactly how it handled them. Threw the same contract into Gemini and it just… knew it. Every clause. Every detail.

Wasn’t obvious because I’m smart. Anyone with a different use case might’ve missed it. It was just a random moment where I went oh — oh god.

That sent me down a rabbit hole. I started actually looking at what Google was building. Nano Banana. AI Studio. Vertex AI. Google Cloud. Their own TPU chips (basically Google’s homemade version of the specialized chips needed to run AI — everyone else, including ChatGPT, was scrambling to buy them from Nvidia).

And it hit me. We’d all assumed Google was sitting the AI race out. They took so long with Gemini that the narrative had basically written them off.

They weren’t sitting it out. They were quietly building the whole stack while everyone else was scrambling. Models, infrastructure, chips, the workspace layer to put it all in.

Once I saw that, the next move felt obvious — Gemini was going to be the AI layer for Google Workspace. Sitting inside Gmail, Docs, Sheets, Drive. Microsoft only had Copilot and they weren’t really trying yet.

So I convinced my partners to move our six-person company off Outlook and onto Google Workspace. Not because of email. Because of where AI was going to live.

Claude Code and the death of middleware

Now here’s where it’s going to start sounding more technical, so quick disclaimer — if words like terminal or database make you want to close the tab, don’t. A year ago I didn’t know what any of this meant either. The AI literally taught me as I went.

Then Claude came along and I started using it more. Then Claude Code.

Claude Code was the next real shift. It runs in what’s called a terminal — basically a plain text window on your computer where you type commands instead of clicking buttons. Sounds intimidating, but using it with Claude Code feels more like texting an assistant than hacking. You tell it what you want in plain English, it writes the code, runs it, and shows you what happened. I’m not coding. I’m supervising.

Suddenly I wasn’t just chatting with an AI — I was working with one. Editing my files. Running things on my computer. It could see what I was seeing.

Then came MCP connectors. MCP basically lets AI talk directly to your apps — your CRM, your email, your calendar — without that duct-tape middleman in between. No Zapier translating, no webhooks (which are just little messages one app sends another when something happens). Just a direct line.

Instead of stitching tools together with five different services, I could just connect them. CRM. Email. Calendar. Drive. Ads. Books. Property management. All native.

Almost overnight, Zapier started feeling… optional. For a lot of things, irrelevant. The whole middleware layer I’d spent a year learning to stitch together was collapsing into one interface that could just talk to everything natively.

Yesterday’s lightbulb (which is actually a few months old)

Here’s the thing — I had a version of this realization a few months ago and didn’t follow through.

I was trying to build the world’s smartest real estate brain. Pull Fed data, interest rates going back 100 years, every market signal I could find, dump it all into a database called Supabase (a database is just a giant organized spreadsheet that software can read and write to), query it, turn it into a daily newsletter.

I got pretty deep. Then Claude Code dropped, MCP connectors got better, I chased the next shiny thing. The newsletter faded off.

Now I’m back at the same wall. And this time I actually see it.

Every single thing that happens in your business is an event.

A text comes in. A lease gets signed. A form gets filled out. A call gets answered. An invoice gets paid. A lead status changes.

Each one of those is just a row of data. A line in a database.

So what if every event from every tool you use just… flowed into one database? Continuously. Forever. Text comes in → row gets written. Lease signed → row gets written. Form submitted → row gets written. The database just keeps growing, capturing everything that happens across your whole business.

That’s the back end. The foundation.

The middle layer is where AI sits on top of that database and can read it. What happened today? Which leads went cold? Which tenants haven’t paid? Draft replies to everyone who texted in the last 24 hours. The AI doesn’t have to guess. It’s just reading your actual data.

The front end is the dashboard. Or the daily briefing in your inbox. Or the Slack ping. Or the pre-drafted reply waiting for you to hit send.

Three layers. Database underneath. Brain in the middle. Surface on top.

Why this is the real edge

I’ve talked to a couple people who’ve worked at Palantir, and this is basically what they say too. At the end of the day, the AI is the same for everybody. The models are the models. Some people prompt better than others, sure.

But the real edge is structured data. Your data. Organized.

Without that, you’re just talking to the internet’s AI. A really smart stranger that knows the world but knows nothing about you. Every conversation starts over.

With it, you have your AI. It already knows. It doesn’t need to be re-briefed.

The wild part — almost none of this requires code I couldn’t have written a year ago. Mostly because I’m still not writing it. The AI is. The tools caught up. The connectors caught up. Claude Code caught up.

The only thing actually lagging was my understanding of what was possible. And honestly, my ability to stay focused on the boring foundational work instead of chasing the next release.

What’s next

I don’t fully know how I’ll build it yet. Cloud, local, hybrid — still figuring it out.

Some of what feels like a revelation to me today is probably day-one stuff for the people who’ve been at this longer. That’s fine. I’m writing it down while it’s still hot.

Going to build it in public this time. Not let it fade off like the newsletter did.

If you’ve gone down this road, or you’re going down it now, or you tried something like it and learned what doesn’t work — I’d genuinely love to hear it.

This feels like the actual shift. Not the AI itself. What we finally let it see.

Originally posted on X

This piece started as a thread, then grew into something bigger. The original post lives here on X — drop your reply there or hit me directly if you’ve gone down this road yourself.

— Dakota · @xovionai