How I work with artificial intelligence

This is the introductory post, and I can't think of a better way to introduce myself than by showing the setup exactly as it is, loose cables and all. Machines at home, rented servers elsewhere, models running locally and several AIs in the cloud. I'll tell you why. Why a single tool isn't enough for me, and which parts of all this don't work the way they should.

I started playing with artificial intelligence at the end of the last century. I used a language called Prolog. Then I moved on to playing directly with C, later with neural networks, and in the end I landed on node networks. The most serious thing I managed was a document classification algorithm, and that was already around 2012.

That passed for artificial intelligence back then, which means an SQL database could do the same and more. It looked nothing like what we have now. I got into natural language processing, which is the attempt to make a machine understand what we write, and it turned out to be a colossal effort for laughable results. I came to take it for granted that we would never have machines big enough to process human language. The problem with language was, first, being able to split it into sentences and then recognising each unit. Detecting intent, irony, sarcasm. It definitely looked impossible.

When I saw ChatGPT working in December 2022 I couldn't believe it. It was what I had been looking for for years. Well, no: I had dreamed of something much smaller. And today that GPT-3.5 is a joke compared with current models.

If you're just starting, you know the videos: "improve your life with a single prompt in under ten minutes", "the only tool you'll ever need". I bought it too. In early 2024 I spent an entire weekend trying to get ChatGPT to organise my email, my calendar and my invoices with one very, very long message, and what came out was a beautiful text explaining how I should organise them myself. Not a single email moved.

Since then I've been testing on real systems, my own and clients', and what's left standing is not a magic tool. It's a setup with lots of pieces that works because I've kept patching it by getting things wrong.

If you want to know who I am and what I do, it's on the "About the project" page.

Today I work mainly with Claude, but also with OpenAI, DeepSeek, GLM, Kimi, MiniMax, Ollama, LM Studio… and I could bore you with names. That's not the point. I've ended up in this geeky madness because what's happening is, for me, the most exciting thing in the history of technology, and like so many others I jumped into the pool hoping to learn a lot, and fast.

ChatGPT was my way in, like almost everyone's. GPT-3.5 was a toy: surprising and promising, but a toy. Until one day I asked it to help me set up a Linux server, the first of my life, and it did. With some difficulty, because its knowledge was out of date, but it did, and I, who had never touched Linux, discovered a fascinating world. That server had an RTX 4070 with 12 GB of VRAM (the graphics card's own memory) and 64 GB of RAM. I wasn't very sure what I wanted it for. Today it's a cornerstone of my local AI.

Then new models kept arriving and, as soon as Claude Code came out, I said: "This is exactly what I need." Love at first sight. I tell the whole switch in Why I switched from ChatGPT to Claude.

Until then I had practically never used PowerShell. Today it's a rare day when I don't have 16 sessions open of PowerShell, MobaXterm, Visual Studio Code or anything else with a terminal. I work with SSH (the way to log into another machine remotely), with tunnels, with secure access and with a pile of things I didn't even know existed three years ago.

The cloud gives you two things you don't have at home: protection against physical failures and the fact that it's always online. At home, machines get switched on and off, and Windows makes decisions for you: sometimes it feels as if, as far as Microsoft is concerned, the machine isn't yours but on loan. So they turn it on and restart it whenever it suits them. On top of that, in the cloud you're always visible, and that's where I publish everything; at home is my working environment and I publish nothing. In fact I have to make sure nothing gets published by accident and there are no unwanted intrusions.

I use the cloud to host my blogs and my applications: some in testing and some in production, some multi-user and some private. And suddenly I find it very easy to publish and maintain secure content exposed to the internet.

Why VPS? Because they're cheap, fast, have come down in price, and you don't have to worry about their power consumption, their availability or them breaking because a fan died. You upload a harness, which is the program that wraps the model and lets it do things, like Claude Code or OpenClaw, and you tell it what to do. Coding directly on the server? No. Deploying, installing certificates, linking domains to your applications' code. But forget about local models there. Only APIs and subscriptions to the big models.

Why so many machines? Because you can't run everything on the same one. If you launch a process that takes three days, another that takes six hours, another four, and so on up to a whole bunch, one machine can't keep up. Besides, if you want one agent watching your email, another reading the news and another keeping your accounts, the sensible thing is that they don't share a house. And you have to keep development machines, sandbox test machines (an isolated environment where breaking things has no consequences) and production machines separate.

But don't picture eight ultra-fast, latest-generation PCs. There's one powerful machine, the one with a graphics card, and the rest are mini PCs and old recycled machines. Linux works miracles with old hardware.

You can only run local models on powerful local machines. It sounds like a paradox: the big cloud models don't add a penny to your electricity bill, while the small ones you run at home consume a huge amount. That's because, in the cloud, the provider supplies the infrastructure and pays for its power; locally, you supply it and you pay. And, spoiler, however good your machine is, it will be very limited.

For me, yes, because I'm a convert. I've been evangelised by AI and I believe any future goes through it. But I don't think a sane person needs it. If all you want is to learn what matters, a good laptop with plenty of memory will do. A single AI provider, and here, if you want to get your money's worth, I recommend Claude. I've been evangelised there too, but once you understand Claude Code and what it can do for you, you'll know why I let myself be converted.

In the rest of the blog I'll explain why I use what I use, my toolbox and what kind of machine each use case needs. One day you'll need to design a website, another to read invoices or do the books.

Time to be honest. Some of the eight servers at home are surplus: I set them up out of enthusiasm and today they spend most of their time doing nothing useful. In fairness, though, the first one I set up, the one with the RTX 4070, I bought on a hunch and today it's the one that carries the most work. The other machines become useful as I keep reinventing my life. I automate things that are always running and monitored (in theory), and I even kid myself that I have digital employees. But... I'm evangelised, what can I say.

Maintaining all this eats resources, time and money. I could see it as an expense, but it isn't. It's an investment. My normal work isn't what it used to be. I've upgraded my whole activity by putting AI everywhere I could. And now I can do, and charge for, more than I did before, and I can also try out new businesses. Yes, all this has a cost, and it pays for itself because people pay me to do things I couldn't do before. I give a better service and I can follow better security habits.

To be clear: the blogs and the VPS that host them are in production. OpenClaw, DeepSeek Harness and a good chunk of the local stuff are experiments. I never mix production environments with development or with sandbox experiments. I advise you to follow the same practice.

Experience told with numbers, failures included. Standalone articles on specific cases, such as my local AI or my servers, and series such as the challenge of finding out whether I can build a business that brings in an extra salary, a challenge that is already over and is told in parts, just as it turned out.

Among the principles of this blog are these: I'm not going to sell you anything. I'm learning just like you. When a piece of advice slips out, remember that I've been evangelised and my advice isn't impartial.

And if, after reading all this, you think eight servers, six VPS and fifty models are too much for one person, you're right. My family thinks the same, especially when the electricity bill arrives.