The first few times, it hurt. Now I smile. I smiled this time too. And it occurred to me that there has probably been stuff of mine online for almost 30 years. I did a quick search and... yes, I found traces of my posts on public mailing lists (still indexed) as far back as 1997. I was talking about mobile phones and Linux, and my writing style was similar to the one I have today. But they are fragments, disconnected sentences, often replies to threads. So I decided to dig further, trying to use publicly accessible resources (that is, not my private e-mails).

The "oldest" trace of something sufficiently long (and cited by others, too) dates back to 2002. I still have the LaTeX source, as well as the PDF, but I downloaded it again from the Wayback Machine anyway. It was an analysis of intrusion detection tools and techniques, and that document carries a special date. It was my birthday, and it had been prepared for something very important. But I'll talk about that in another post.

As soon as I started reading, I smiled: partly for the structure, partly for the English (clearly translated quite literally from Italian, something that still happens often today - albeit on a different scale). But the best part came on page 13.

I was explaining the Anomaly Detection Model. First I described the idea: learn the normal behaviour of a user and consider deviations suspicious. Then I summed up the paradox in one sentence, even highlighted:

"Everything which has not been seen previously is dangerous."

And immediately afterwards I describe its main flaw: "The high false alarm rate is generally cited as the main drawback". I also explain why: training cannot necessarily cover all legitimate behaviour, behaviour changes over time, and this can produce more false alarms.

A small comparison.

A human detector says: "He uses this structure." "He makes this kind of list." "He writes therefore, overall, it depends." "This pattern looks like an LLM."

I, in 2002, was reflecting: an anomaly with respect to the model is not necessarily an intrusion. If you treat everything that doesn't match your model as dangerous, you produce false alarms.

The conceptual similarity is almost perfect.

But that didn't answer my question: how did I write, back then? I went back and started reading again, this time to understand some details of a formal nature. And something, as I suspected, was already there. "Too many to be listed, too many to be solved." And the comment: "Nobody said that being a System Administrator is an easy job...that's why I like it!".

Several parts, in fact, match many of the stylistic clichés that today are associated with ChatGPT. Not so much the single words, but the way I tried to organise and explain what I had in mind. Probably my mind is an LLM. I need to ask my parents whether I was vibe-conceived.

For example, I already had a certain tendency to take a problem and break it down neatly. On page 5 I wrote "We can sum up this way:" and split the problem into three steps: Prevent intrusions, Detect the intruder, Response to intrusions. And, of course, I didn't stop at three little headings: for each one I explained what I meant, what could go wrong and what the exceptions were. On page 4 I did even better: "So, to sum up, firewall approach has some pros and some cons", followed by two sections, Pros and Cons. In short: take a topic, split it into pieces, analyse the advantages, analyse the disadvantages and then draw the conclusions. Sound familiar?

There was another thing that, evidently, already belonged to me: the difficulty in giving absolute answers. Is a firewall useful? Yes, but. Is a very restrictive policy more secure? Yes, but. Does a permissive one cause fewer problems for users? Yes, but. On page 3 I talked about the need to find "a compromise between security and usability" and, a few pages later, I turned this idea into a classification: Strict policy, Permissive policy, Balanced policy and even Perfect policy, with false positives, false negatives and practical consequences thrown in. Evidently my love for "it depends" has rather deep roots. Even then I understood that there is no absolute answer to every question. Apart from 42, of course.

The way of explaining was already that one, too. "Let's explain it in more detail". "In other words...". "For example...". First I introduced a concept, then I tried to say it in another way, then I added an example. And, as soon as I risked having oversimplified, I added the clarification. On page 12, after building a table with the "normal" behaviours of System Administrator, Boss, Secretary, Manager and Programmer, I hurried to specify: "This is of course a rather general example. Real profiles deal with more precise categories...". Generalisation, example, caveat.

And I already liked closing a long explanation with a short sentence that tried to sum it up. That is exactly what happens on page 13: "In short, this is the paranoid approach:", followed by "Everything which has not been seen previously is dangerous.". A sort of takeaway, to use a term that at the time I would probably never have used. But, immediately afterwards, here comes the counterweight again: the main problem of that approach is the high number of false alarms. So: explanation, summary, punchline, "yes, but...".

The same scheme reappears when I talk about SYN floods. I take a concrete case, describe how it works, pull out the elements that make it recognisable and then generalise: "In a similar manner indications of other attacks can be figured out." And throughout the document I keep leaving road signs for the reader: "The basic idea is the following", "As you can see", "Have a look at table 2", "We can sum up this way", "In short", "For example". I wasn't just explaining something: I evidently felt the need to explain where I was going while I was explaining it.

Then there is the general structure: parts, subparts, numbers, lists, tables, diagrams, classifications. But here I would be unfair to poor 2002 Stefano: it was a technical document, written by a Computer Science student who, in that period, read far more manuals, papers and technical books than novels. It was fairly inevitable that it would be structured that way. More bits and less narrative, in my mind. But also more pizzas, and this, perhaps, was the beauty of youth.

The interesting thing, though, is that this way of reasoning didn't stay confined to a university paper. I find it again later, when I no longer had to respect any academic structure and was simply writing on my blog. And that's where it starts getting even funnier, because I decided to carry on. I was enjoying myself. A search born out of curiosity that became, along the way, a journey into my past.

So I thought of my Italian blog, very active in its first years (2006-2009), with other bursts of publishing afterwards. Mainly, until 2018. Well before LLMs arrived and, I'll add, with posts that can also be verified on the Wayback Machine. While I was at it, I migrated the blog from Nikola to BSSG, with the idea of not ruling out writing there again. Maybe. If I find the time. Who knows...

So I read some of the more representative posts and, there too, I smiled, enjoying the journey. They are all in Italian, so I'll quote the originals and translate as I go.

The first is from 2006: Gli informatici, gli "informatici" e quelli che "ci capiscono" (IT people, "IT people" and those who "know about computers"). It's a scene that anyone doing this job has lived through: the neighbour's computer, three antivirus programs piled on top of each other, twelve minutes to boot and friend X, the one who "knows about computers". I build it as a back-and-forth: he says, I answer, he objects. "A quel punto..." ("At that point..."), and off we go again. Three times. In between there are the parentheses, the typos, the smileys and a demonstration that includes "due cicli di Krebs (che non c'entrano niente ma ci stanno bene)" - "two Krebs cycles (which have nothing to do with it but fit nicely)". Those, unmistakably, are mine. But the best part is the ending: I get summoned before a sort of improvised court, I explain everything in full detail and I get convicted anyway. The trial was already over. Even in 2006.

2008 was a prolific year. And, rereading it today, a textbook one.

In Cos'è il VOIP? (What is VOIP?) I proceed by dry statements, in bold: "VOIP means...", "Skype is not THE VOIP...", "We all use VOIP...", "VOIP only needs...". A short sentence, then the explanation, then the clarification (complete with a grandmother being called from the old rotary phone, to show that we all use VOIP without knowing it). Up to the inevitable "Il VOIP può far risparmiare molto, ma dipende..." - "VOIP can save you a lot, but it depends...". There it is. I knew it was coming. At the end I define it myself as "una sorta di mini FAQ" ("a sort of mini FAQ"). No detector needed: I had already turned myself in.

In Il popolo dei blogger e dei loro lettori (The people of bloggers and their readers) I start classifying. First the bloggers. Then the readers (or rather, the "LETTTORI", with three Ts: the typos, as I was saying). Those who pass by and leave, those who take part, those who agree, those who disagree. And then a category that, reread today, made me laugh out loud: those looking for a personal vendetta, who never engage with what you actually wrote and simply call you incompetent. One of them wrote to me that I was ugly and should remove my photo from the blog. Eighteen years ago I had already put them in a taxonomy. Only the word "slop" was missing. At the bottom, two twin questions: "Chi sono, dunque, i blogger?" and "Chi sono, dunque, i lettori?" - "Who, then, are the bloggers?" and "Who, then, are the readers?". Each with its answer, of course.

And then there is BlackBerry: i suoi pro e i suoi contro (BlackBerry: its pros and its cons). Here there is nothing to interpret, it's all in the title. PRO. CONTRO. Both openly in no particular order. But not even the single points manage to be clear-cut: "dipende" ("it depends"), "nel complesso" ("overall"), "personalmente" ("personally"). Even among the cons I find a way to add that, after all, the competition doesn't do much better. At a certain point I ask the question myself: "A questo punto la domanda che uno potrebbe porsi e': si, ok, ma quindi?" - "At this point the question one might ask is: yes, ok, but so what?". And I answer: I, for now, am happy with it. "Tu...fai quello che vuoi! :-)" - "You...do whatever you want! :-)". The most honest recommendation I have ever given.

In 2009, Identità reale o pseudoanonimato virtuale? (Real identity or virtual pseudo-anonymity?). I start from Fidonet, when people showed up online with name and surname, and I arrive at two categories of people who use a nickname: those who want to protect their privacy and those who want to feel free to behave badly. The first I understood perfectly, then as now. "Nel primo caso...". "Nel secondo caso...". In the first case, in the second case. Which choice is better? Both have their pros and their cons. I don't know why, but I saw that coming. I close by saying that I tend to sign with my name almost always. "Ho detto quasi :D" - "I said almost :D".

In 2010 the BlackBerry comes back. Or rather, it leaves: Da BlackBerry Bold ad Android (HTC Dream G1) - quasi 3 mesi dopo (From BlackBerry Bold to Android - almost 3 months later). Two sections, facing each other, in capital letters: "QUANDO HO APPREZZATO L'ANDROID" and "QUANDO MI E' MANCATO IL BLACKBERRY". When I appreciated the Android, when I missed the BlackBerry. Nine points on one side, seven on the other. Then: "Tirando dunque le somme..." ("So, summing up..."). The sums, naturally, say that both have positive and negative sides and that neither has everything. Sixteen years later, the BlackBerry is gone. The "dunque", my beloved "therefore", is still here.

2013, Il SEO, questo sconosciuto. O forse no? (SEO, this unknown. Or maybe not?). The doubt is already in the title and the answer arrives in the third paragraph: "O forse no? Dipende" - "Or maybe not? It depends". Two "Sia chiaro" ("Let me be clear") follow, then "in some cases", "in many other cases", "in my opinion". I recognise the value of those who do that job seriously, I take issue with those who judge other people's work without having the skills (and, as it happens, with an aggressiveness proportional to their incompetence), then I try to distinguish when a consultant is needed and when not. In the middle, a line I stand by: "il mondo è bello perché è vario. O avariato." The world is beautiful because it's varied. Or gone off. (It works better in Italian: vario, avariato.) And at the end, after all that caution, I offer my two cents.

2014, La mia esperienza (positiva) con BTRFS (My (positive) experience with BTRFS). With the parenthesis already in the title. I describe the test bench (the Atom server at home, 4 GB of RAM, three 1 TB disks) and then: "Il risultato? Eccellente" - "The result? Excellent". Question, answer. Right after, though, come the details to keep in mind: copy-on-write, recent kernels, a stability that is not yet complete. And at the end, the question of all questions: "Mi sentirei dunque di consigliare BTRFS al grande pubblico? Dipende." - "Would I therefore recommend BTRFS to the general public? It depends." One word. Full stop included. Eight years after the first post, the gist was still the same.

And then the small scattered pearls. In the 2008 review of Nav4All, a navigator that promised to stay free until 2010: "Promessa rispettata? In larga parte, si'." - "Promise kept? For the most part, yes." Further on: "Come va? Bene." - "How does it work? Well." And in the one from a few months later: "Come si comporta il Curve con il Nav4All? Benissimo" - "How does the Curve behave with Nav4All? Very well". Dry question, dry answer. Today a construction like that is enough to be convicted without appeal. (For the record, in 2010 I wrote Nav4all non c'è più - Nav4all is no more. Punctual.)

I could go on for a long time: there are many posts, and each of them has something. Every now and then I start from an image and, at the end, I go back to it, to close the circle. I still do that today and, for some, it will only be yet another confirmation. I'll probably do it in this post too. Or maybe not. And somewhere, even back then, I wrote that open source doesn't get along with closed minds. I still think so. But by now the point is made, and there's no need to list everything.

Eight years, wildly different topics: phones, navigators, filesystems, bloggers. And always the same habits. I ask a question and I answer it. I anticipate the objection. I split into cases. I put the pros on one side and the cons on the other. And when the moment of the verdict comes, "it depends". None of this arrived with LLMs. It was already my way of reasoning. And of writing. With one difference: inside those structures there are the Krebs cycles, the three Ts of "LETTTORI" and an embarrassing amount of smileys. The form, perhaps, looks LLM. The rest, I'm afraid, is all mine.

The me of today is a product of the me of back then. What I liked, I kept doing. Other things, instead, I stopped writing. Some I would write differently today. Others I wouldn't write at all. But I enjoyed writing then, just as I enjoy it today. Actually, today I have more things to say. More experience and, perhaps, even more certainty that there are rarely simple answers to difficult problems.

Does any of this prove that I wrote this post without an LLM? No. Nor was that the point. The old posts show that these patterns were there long before LLMs, not how this post was written. And when someone demands "prove it isn't", the burden of proof has already been reversed.

But 2002 Stefano had already written the answer, on page 13. A detector that treats everything outside its model of "how a human writes" as dangerous will produce false alarms. I'm one of them, and probably not the only one. And the high false alarm rate, I'm told, is generally cited as the main drawback.

So what does all this prove, then? Basically two things: that "LLM patterns" have always been part of the way I write and that, for those who have been reading me for twenty years, many of those patterns shouldn't be news at all. A pattern is not a verdict.

Or that LLMs learnt from my texts - or perhaps, getting serious again, that people wrote like this and LLMs "learnt" from those who, like me, were already writing long before they arrived. Not because we wrote particularly well, but because we produced a lot of text, on the Internet, in that form.

Or... or I really am an LLM. Stefano, the human LLM of the 2000s. And instead of electricity, I run on pizza, tiramisù and (a bit of) coff -OUT OF TOKENS-.