Every time the little speaker on my desk powered on, it made a loud pop. An electrical engineer would have fixed it without thinking, but I've spent fifteen years in growth strategy and commercial partnerships, and until recently I couldn't have told you what an audio amplifier does when it wakes up.
The fix turned out to be twelve milliseconds of silence. The amplifier needs a moment to settle before sound goes through it, so I wrote a few lines of firmware in C that make it wait before switching on, and the pop went away.
I didn't learn that from a textbook or a forum. I learned it by asking a long string of mostly basic questions to a machine about the size of a thick hardcover book that sits next to my monitor. The whole time, it kept reminding me of something I'd read about old factories.
When factories first switched to electricity in the late 1800s, most of them barely changed. They pulled out the steam engine, put one big electric motor in its place, and kept everything else. That motor turned long shafts running along the ceiling, and every drill, saw and lathe on the floor hung off those shafts by leather belts. If the motor stopped, the whole floor stopped.
Motors small enough to sit inside each machine came along fairly quickly. What took much longer was for owners to realize that once every tool had its own power, they could lay out the floor around how the work actually moved instead of around a shaft. Economists who've studied the period found the big jump in factory productivity didn't show up until the 1920s, about forty years after electric power arrived. The motors were ready long before anyone reorganized the work around them.
I think AI is somewhere around the 1890s right now. The big breakthroughs come out of enormous data centers, and those really are necessary for training frontier models. Right now, almost everything we do with AI lives inside a browser tab on a remote server. And plenty of companies are doing the equivalent of swapping the steam engine for a motor, bolting a chat window onto the workflows they already had and waiting for the productivity to show up.
None of this is an argument against the cloud (I use cloud models every day). I just think the more interesting change happens when the power sits inside the tool, right in front of the person using it.
A few months ago I wrote about building a local AI assistant for my website that runs on a 35-watt mini PC in my home office, with no cloud APIs and no monthly bill. It handled one well-defined job really well, which left me wondering what would happen if I stopped treating local AI as a side project and used it to learn things I'd never been able to do.
If you follow tech news, you've probably picked up the idea that AI is mostly about raw computing muscle. For the everyday stuff most of us use it for (chatting, coding help, reading documents), what slows things down usually isn't the processor, it's memory.
Picture an assembly worker at a workbench who can snap a part together in a second, but has to check a huge instruction manual for every step. The manual is kept in a warehouse across the street, and the only way across is a narrow footbridge that fits one person at a time. You can hire as many runners as you like, and the worker will still spend most of the day leaning on the bench with their hands in their pockets, waiting.
That's roughly what happens in a regular computer when an AI model is too big for the graphics card. Part of the model ends up on the far side of a slow connection, and everything crawls while data shuffles back and forth.
The machine on my desk is an ASUS Ascent GX10, built on NVIDIA's DGX Spark platform. Instead of splitting memory between the computer and the graphics chip, it has one 128-gigabyte pool they both share. It isn't the fastest memory around, and a high-end gaming card moves data quicker, but a model that would never fit on that gaming card fits here in one piece, so nothing has to cross the bridge.
Models that used to need a multi-GPU workstation or a rented cloud server now run on a box that costs about what a well-specced MacBook Pro does, draws around 150 watts, and is quiet enough that I forget it's on.
I've always liked mapping out complicated problems and figuring out how to solve them. What I couldn't do was build the software myself.
Computers are ruthlessly literal. Miss a comma and the whole thing stops with an error message that might as well be in another language. You can understand exactly how a system should fit together and still lose a Saturday because one command didn't work, and developer forums often treat beginner questions like personal offenses.
I also didn't want to just copy and paste blocks of code I couldn't read and cross my fingers that they ran. Even when the AI was helping write the code, I wanted to understand what it was doing, and I needed a teacher that could connect the way I already think about systems to the rules for actually building them.
The cloud models I mentioned are still better at long, multi-step reasoning, so I lean on them to plan a project and check my thinking before I start. For the actual building, local models hold their own. Most of the learning happened on the local machine, though, and a lot of that came down to how freely I could ask questions.
When you're working with an API key or hitting a usage cap on a frontier model, it's easy to ration your questions without noticing. You skip the follow-up that feels too basic, or you don't ask a third time when you still don't get it. Having a machine running right on my desk took that hesitation away. The local model sat with me at 1am while I tried to fix my 3D printer. I'd ask it to explain a Linux error, then what one of the words in its explanation meant, then why that mattered, until it actually clicked.
It wasn't always right, either. A local model only knows what it learned up to the day it was trained, so it'll happily walk you through a setting that got renamed six months ago. I ended up hooking it up to web search and a few other tools so it can check current information before it answers, and I learned to treat it like a smart friend who's sometimes overconfident.
Within weeks I was writing Python and finding my way around the Linux command line, and not long after that I was running my own models and setting up secure network tunnels. I'm still a strategy person at heart. The distance between having an idea and knowing how to build it just got a lot shorter.
The bigger change took me longer to notice. I'd always thought of AI as a website I visit. Once it was running on my own network, it became something I could build with, a part that anything I made could talk to without an account or an API key.
So when I wanted one of those little digital clocks you set on a shelf to help structure the workday, I decided to try making one instead. While looking at small ESP32-P4 touchscreens about the size of a light switch, I picked one that also had a speaker and two microphones in case I wanted to experiment later. It showed up as an unprogrammed board with a blank screen and a USB port.
I started by asking how to put a clock on the screen. The model explained what the chip could do and how its firmware gets written in C, and every answer led to another question. (That's where the speaker pop came from.)
A few weeks later, the bare board had become a voice assistant. It shows an animated orb that reacts when I talk, streams my voice over Wi-Fi to the box on my desk, and answers back in real time. None of it leaves my house.
Here it is in action:
It's not a polished product, and I'm still working through some quirks with the touch interface. But a few months ago I didn't know what a microcontroller was, and now there's one on my desk running custom firmware I put together and actually understand.
There's a real debate about whether AI will take more from people than it gives back, and I'd be lying if I said I had it figured out. What I can tell you is what it looks like from one desk: a 150-watt box that answers my 1am questions without counting them, and a little orb I built because I finally knew how.
My guess is that a lot of the reorganizing this time will happen quietly at individual desks, one person at a time, wiring AI into their own tools and asking the beginner questions they used to be too embarrassed to ask.
If you'd like to see where this started, you can still talk to the original AI assistant on my homepage. It's running on the same 35-watt mini PC that kicked off all of this.