I've Already Changed Jobs 4 Times Because of AI: My Story

From research to a VAT number

I started out as a researcher, and back when I was doing research, neural networks were considered a bit of a niche. So much so that research labs actively went looking for researchers in universities, because nobody knew about this stuff and few people wanted to do it. I found myself catapulted into a lab where we did real research, not just on AI with neural networks, but also on artificial life. We ran simulations on clusters. This was well before it was fashionable, before AlexNet in 2012.

And with that I jump straight to 2012. What happened there? A part of the research world took these techniques for simulating nervous tissue and applied them to large quantities of data, knowing they’re models hungry for data, and up to that point all that data simply hadn’t existed. The transition was tied precisely to the fact that finally there was all this data and all this compute to see deep learning bloom. Deep learning is a derivative of these neural networks where you use many layers and lots of data, because the network absorbs and absorbs, models and models, and with a large, deep network made of many layers you learn deep patterns and compress much more complex ones inside far higher-dimensional data. In 2012 the ImageNet challenge came out, which not coincidentally was a computer-vision challenge based on huge amounts of data, because the time was ripe. AlexNet won, and computer vision took off.

Now, since I wasn’t very satisfied with how things were going in the lab I was in, partly because, contrary to what many people think, research, and not just Italian research but European and global research, is constantly hunting for funding. Doing research doesn’t mean exploring. It means publish or perish, publishing as much as possible to be able to enter the calls, to get the money to do more research. Basic research is very hard, and it’s getting harder. You tend to drift toward applied research. It’s no coincidence that AI research labs end up being R&D departments inside multinationals, or startups that are really spin-offs of big tech. See Anthropic and Amazon, see OpenAI and Microsoft.

Anyway, after a lot of crying, I decided to change, both because I wasn’t happy with the situation, not with the people I was with, whom I still talk to and love, but with the prospect of being researcher-plus-one. Okay, I stop doing research, I open a VAT number. Since I was absolutely nobody, I was convinced I could make money off my knowledge, that I’d easily find contacts. I saw the data economy exploding, I knew all this neural network stuff, machine learning, all sorts of algorithms, including bio-inspired ones for working on data. I figured, this stuff has immense economic value. It did have it. But nobody gave me the time of day. Not even close. I took a lot of slaps.

Web development, then the data science wave

Then the first signals started arriving from the United States and people began talking about data science and big data. Italian business owners, convinced they could do big data with four sad Excel sheets held together with tape, started looking for professionals to do their big data. So I changed jobs again. In the meantime I’d gotten into web development, I’d become obsessed with D3.js, there were the wars between the first JavaScript frontend frameworks. So for a while I stopped doing neural networks, but I’d fallen in love with web development, I discovered WordPress and open data, so I actually found some interesting things too. Compared to the skills I had, I was completely overqualified to be doing that stuff, but that was the only place I found a market.

Big data, data science, machine learning were arriving, so finally I found an outlet to play my skills again. I started talking about this stuff at meetups here in Rome. And I came to realize that the big limit to doing data analysis and machine learning, the first old-school forms of data-based AI, wasn’t really “let’s find someone who knows how to do it.” The obstacle was that digitization was in the stone age. It was full of people who talked, who filled their mouths with big data, data science, machine learning, and the moment you mentioned a database or a web microservice, or, God forbid, the semantic web, the moment you proposed anything remotely to do with the technology to curate their data, which was absolutely in their interest for anything they wanted to do, they’d start on about vendor lock-in. “Don’t give all this data to SaaS, to external services, because then it’s not easy to integrate it.” It was the most total ignorance, with exceptions. Luckily those exceptions let me survive a bit longer.

So I moved on to this third stage of my career. From researcher and web developer I became a data scientist. And doing data science, another transition happened. In a few years there was a spike where, from making models on your own computer with data you’d painstakingly collected or been handed, which, I repeat, was the main constraint, the real obstacle was never the algorithms or the AI, it was the digitization. Ridiculous. Anyway, I found some slightly sharper environments and started building custom models, even quite advanced ones: algorithms working on text, on clustering banking transactions. I did a ton of stuff.

When my own laptop stopped being enough

At a certain point computer vision had exploded, it had spiked hard, it had found an outlet in industry, and I remember this transition. From being able to do machine learning, deep learning, both text and images at the same time, building image classifiers and text classifiers, from object detection onward, from models like ResNet, then YOLO, the ones that find where objects are in images, I realized two things. One, I could no longer do everything to do with machine learning. Two, I could no longer keep the models on my own machine. I was used to having my little computer and training everything on it, collecting the data, maybe leaving it running for two weeks, but it managed.

With computer vision, and in particular with object detection, this practice started appearing of downloading the big model that required enormous resources to be trained on millions of images, and using it as a backbone, as the initial part of a model whose downstream layers you’d tailor. You’d take an image classifier trained on millions of images, swap out the tail of the network, and have it do different things on the little data you’d still painstakingly collected. So I changed jobs again, because I could no longer present myself as the guy who did data science in general. I had to specialize, a bit of text, a bit of images, but for a couple of years I actually did only images, only computer vision.

Language changes everything

Except that language arrives. Go to 2018, go to 2019, and in 2020 GPT-3 lands a blow like nothing I’d ever seen in my life. GPT-3 was truly a “wait, everything changes here” moment. In the scientific world people were saying everything changes. It took another couple of years for it to arrive as a product, as ChatGPT, but it really is changing everything.

I change jobs again, because in the meantime I’d specialized in PyTorch, TensorFlow, all computer vision, and this language stuff arrives. Computer vision had pushed even further, the models were more and more general, needing less and less fine-tuning, so I figured, I’ll go do these language things too, language models. And I realized that there too I had to completely change strategy, because with language models you need a cluster, and a GPU cluster at that. There was no way around it. And you didn’t even need to do training anymore. The whole reason those things exist is precisely that fine-tuning isn’t needed, despite so many now saying “small language model, fine-tuning.” You’ve gone back to the year 2018. Small language models make no sense. The whole reason GPT-3 and everything after it exists is that you don’t have to do fine-tuning, unless it’s strictly necessary.

So I change jobs yet again, and this time I tell myself: now I stop training models, I can no longer afford to have them on my own computers. This was a bit of a trauma. But since a long time ago I’d learned web development, I go back to it and start doing web development around these language models, because from my point of view it was obvious that things like agents would arrive, which people were already talking about even before GPT. I said, there’s a whole world of software written around this language model, because you use it as a language engine, but it produces and responds in language, so there’ll be a need for a whole set of accessory things: how do I connect it to databases, how do I connect it to the internet, how do I connect it to your grandmother?

The Cheshire Cat, and what is actually scarce

And here I said, I’ll go do this. When ChatGPT came out I launched the Cheshire Cat. The Cheshire Cat drew a lot of attention, especially here in Italy, and it was a way, well ahead of the wave, to show that the conversation would above all be about software around the language model. I was genuinely inclined to release it, and to do it open, because that’s how I wanted to do it. I won’t hide that many of the people who then came near the project put me off it, but luckily there are also many good contributors who helped me and who understood the essence of the project. For me it was technological socialism. I’m not ashamed of it, I don’t give a damn, it was a gift and it still is a gift.

But there too I changed jobs again, because I’d spent my entire life studying algorithms, training models, putting pieces together, doing statistical analysis, sometimes visualization. Just being adjacent to AI had already destroyed and revolutionized my career at least four times. And I came to this realization not long ago, thinking: now look, I worry because all these developers, myself included, who build applications with agents now find themselves lost, and out of nowhere they don’t know what to do, their craft has been completely changed. But the truth is it’s already happened to me at least four times, and I didn’t do all this whining. What I did each time was understand where I had to move, learn as fast as possible what there was to learn, and put myself out there again.

From a certain point of view, and I’ll include myself even though it’s not quite true, we’re spoiled. In a lot of the content I’m making I try to lead with empathy for all these devs who find themselves shaken up. But then what do we do? The agents? Some people are in pure denial, they don’t see this thing. And look, it’s not that you’re wrong to feel it, the point is you’ve been spoiled. What were you going to do if you had to change jobs four times in ten years?

The empathy stays, and I’m telling you we’re not done, things have only just begun. Because what I’m seeing lately is that very probably writing code itself makes less and less sense. What makes more and more sense is to learn continuously, to continuously decide where to position yourself, and to put yourself out there with credibility, with the right values. Because say what you want, it’s true that code has been a scarce resource until now, but in my view an even scarcer resource has been credibility. All these people who want to change the world with their venture-capital seed money. In my book, if you want to change the world, go volunteer at a soup kitchen, or go plant trees. You’ll change the world far better that way than by accumulating stuff for your own sake. That’s not changing the world. If you openly say “I want to make money,” I respect you a lot more. So what’s missing, in my view, is credibility, values. What’s missing is men and women with guts. That’s the scarcest resource.

And folks, I’m waiting for you with real affection, because I’m telling you the thing is traumatic, but having been through it several times, I still love AI, reinventing yourself continuously isn’t such a big problem if you find your nerve, if you’ve got any left, and invent whatever you need to invent. Where am I going with all this? I don’t entirely know, I just wanted to open up a bit. I’m making this video mainly for YouTube, where several people told me to keep making long-form content where I can go deep. This is the deep dive I feel like doing right now, because the deep dive on the technical and scientific side, I’m less motivated to make, since honestly it gets little traction. What I’ve found over time is that the content that circulates most is the dumbest, the loudest, the most emotionally polarized, so I got fed up. Not forever, maybe the urge to do smart things will come back, but the smartest thing I wanted to do, dedicated to everyone thinking hard about these themes right now and facing important choices, is exactly what I said. You can do it, don’t worry, find your nerve and go, change, learn, reinvent yourself, and then we’ll see.

On the technical and scientific content, I’ll close here. One thing I’ve heard little about, and wanted to make some content on, is why this jump to agentic AI happened across 2025 and 2026. It’s been discussed very little, in fact nobody has really discussed it. If it’s become hard to build bigger models that don’t scale, what actually changed to get to Claude Code, and especially to the latest versions of Claude Code, Codex, and company? That I studied. There really are extremely interesting things that happened at the foundation, all with interesting implications, work-related as well as scientific and commercial. I’ll probably talk about it alongside the Cheshire Cat, because version 2 is coming out soon and it contains all the elements you need to reason about agents, what the essential building blocks are. So with Cheshire Cat version 2 we’ll do a bit of training again, some experiments, and have fun. Unfortunately, a whole set of things, from now on, I’ll be keeping to myself. For now.