The horse is not enough
One and the same “horse” before worlds far apart — an archive, a temple in the mountains, a medieval village. The animal counts too — but above all where you take it, and how you rig it.
This summer I had Buddhist sermons translated from Thai and Pali, I untangled two temples that even Google Maps confuses, I read medieval documents about the history of a village in Tuscany. With the same “artificial intelligence” that, asked in a chat window, would have handed me mostly plausible inventions, delivered with a straight face.
Same name, opposite journeys. The whole difference lies in the horse underneath, in how you rig it, and in who holds the reins.
And yet we almost always talk about “artificial intelligence” in the singular — “AI,” the way we say “the sun” or “the rain”: a single entity that, depending on who’s telling the story, will either save us or ruin us. It’s an optical illusion. “Artificial intelligence” doesn’t exist as a single technology: it exists as a phenomenon — an umbrella under which very different technologies coexist, used in different ways with wildly different results, and along with them the relationship we humans are building with them as they enter our lives. And the difference between magic and disappointment doesn’t depend on the technology alone: it depends on who uses it, and on the system built around it.
Then there’s a gap that strikes anyone who uses these tools every day: the volume of opinions is far greater than the volume of experience. If everything you know about AI comes from articles written by others — maybe written with AI itself — your opinion rests, in large part, on secondhand impressions. It’s not the first time this has happened: it went this way with the personal computer, with the web, with the smartphone. Every time, the enthusiasts promised paradise, the skeptics cried hype, and whoever learned to actually use the tool gained an edge over both. The point was never whether the technology was good or bad, but what you could do with it.
With AI we’re at the same point, with one aggravating factor: this time the technology talks. It answers you in correct English, it seems to understand, it seems to think — and it’s all too easy to project our hopes and our fears onto it, instead of seeing it for what it is. To do that you need real examples: in a moment I’ll tell you about two things I built this summer that I could never have done on my own — and the lesson I drew from them, which isn’t the one I expected. But first it’s worth opening up those two words — “artificial intelligence” — and looking at what’s inside.
What’s really under the word “AI”
Here’s the point I care about most. When someone says “artificial intelligence,” they’re packing very different things into two words. Let’s try to unpack them.
AI covers many families of technology — from computer vision to robotics. Here I’m dealing with the one that dominates the public conversation today: generative models, above all language models.
What an LLM is (and what it isn’t)
The heart of what everyone is talking about today is called an LLM, Large Language Model: a large-scale language model. Put bluntly but honestly: it’s a system trained on an enormous amount of text to predict the next fragment.
That’s all? Not quite. Predicting the next fragment is the underlying mechanism, not the measure of what it can do: from an apparently simple principle, amplified by training and by the techniques that follow, far more complex capabilities emerge. But it doesn’t reason the way we do and it doesn’t “know” things the way you know them. The model, on its own, doesn’t even keep a personal memory of what it told you yesterday: it can, however, receive that memory from the product or the system around it.
From this its two faces follow directly:
- Why it sometimes seems magical. It has “seen” so many texts that it can hand you the right form for almost any request: an email, a poem, a piece of code, the explanation of a concept.
- Why it sometimes “hallucinates.” When it’s missing the information, it can still generate a plausible continuation: sometimes right, sometimes made up, but said with the same confidence. It’s not a moral bug. Left alone, it completes, it doesn’t verify; to verify it needs sources, tools and a process that forces it to use them.
Understanding this is already half the work. An LLM is not an oracle and not a database: at bottom it’s an extraordinarily powerful text completer, and everything else follows from there — the best and the worst. Treating it for what it is changes everything.
What a harness is: the horse and the tack
Here’s the word missing from almost every barroom conversation about AI: harness. And it’s the perfect word, because it already holds two meanings at once — in tech it’s the scaffolding around a model, but its older meaning is the one that matters here: a harness is a horse’s tack, the gear that rigs it to pull and be steered.
Picture the model as a wild horse: powerful and responsive, but not trained for your journey. Without gear you can jump on its back, but you can’t hope it’ll go where you want: it takes you where it wants, or it throws you.
To actually get somewhere you need the tack:
- The saddle and stirrups, to mount and stay on. This is context: the horse finally carrying your weight, your project, your rules, instead of running on empty.
- The reins, to give it direction. These are the tools and the instructions you steer it with: read here, do this, don’t touch that.
- The blinkers, for long journeys. They remove distractions: a model left free wanders; well rigged, it stays on point.
This is, in a single image, the difference between chatting and orchestrating. Chatting is jumping almost bareback onto a wild horse: one question, one answer, little context and little continuity — the most common use, the one that breeds both easy enthusiasm and disappointment. Orchestrating is saddling, bridling, putting on the blinkers and steering: you don’t ask it a question, you hand it a piece of work inside your world, and you stay the one holding the reins. To be fair it’s not a sharp line but a scale: even a modern chat already has a bit of tack — some context, some memory, a few tools. But as you climb, a difference of degree becomes a difference of kind.
Outside the metaphor, a harness is the system that surrounds the model: it gives it context, memory, tools, permissions, procedures and controls.
It depends what you need to do with it
How much tack you need depends on what you need to do with that horse.
- Do you need a walk around the block? Not much: hop on bareback, a couple of questions, off you go. That’s the everyday chat — I use it that way too, when I ask for the carbonara recipe and what I can use instead of guanciale. And that’s perfectly fine: you don’t rig a thoroughbred to go buy bread.
- Do you need it to pull a stagecoach? Everything changes: you need serious tack and someone up on the box driving. That’s ordinary agentic use — having it write files, build a small app. A job, not an outing.
- Do the horses have to go into battle? Then it’s a whole other world: training, armor, tactics, discipline. That’s the level of the two projects I’m about to tell you about — a knowledge base of documented facts with citable sources — or, in my line of work, a piece of software with three clients, two APIs, Redis and MongoDB on a Kubernetes cluster.
There’s no right horse in the absolute: what counts is which journey it has to make, and with what tack.
Same starting animal, three incomparable feats. Judging AI by the walk around the block — as many have done — and concluding that “it can’t go to war” makes no sense: you asked a pony at the park to go into battle.
And there’s one last detail, the most important: today not all horses are the same. Not only because they’re born different — small models and enormous ones — but because they’re trained for different things: some write code, some reason at length over a problem, some converse, some live inside a phone. The common name, on its own, says almost nothing: no rider would choose “a horse” without asking which one — while we judge the merits and limits of “artificial intelligence” as if the name were enough to tell us what it can do.
Two things I did myself
Theory is worth little without examples. So I’ll tell you two true things, that I built myself this summer using these tools. Not two party tricks to impress: two projects close to my heart, that without these tools I could never have done on my own. They’ll also serve to show the flip side: what AI didn’t do, and couldn’t have done.
1. Keeping a monk’s memory
In northern Thailand there’s a Buddhist temple among the mountains — Wat Tham Chetawan — and the figure of a monk, Kruba Noi, whose work risked staying confined to a few Facebook pages of devotees in Thai. In English there’s almost nothing. In Italian, forget about it.
Keeping a local story also means knowing how to recognize its place, its language and its people.
I decided to build a site to preserve this story and make it accessible in Thai, English and Italian. On my own. With no budget, as an act of devotion.
What the AI did. The most precious thing isn’t the site, but what lies beneath it: a knowledge base. Put simply, it’s a network of text files (in Markdown format) and folders in which you build, piece by piece, the meaning of places and people — who Kruba Noi was, where the temple is, what the sources say, what’s certain and what’s still to be verified — with every piece of information hooked to its origin. It’s from this ordered base that the site, the translations, everything else is then born.
The source material is the most delicate part. Many sources come from social networks, in a language far removed from ours: Thai — and, in the sacred texts, Pali, the liturgical language of Theravada Buddhism. It’s not enough to translate “well” in the schoolbook sense: when it comes to monks there’s a devotional register that a literal translation flattens. An example: from the texts and the sermons the Pali term kamma emerges; in English as in Italian, though, we’re used to the Sanskrit form karma. In a Theravada text, using the wrong form is like mangling a proper name.
That’s why the method wasn’t “give me the translation.” I downloaded some sources, both media ones (also used as a yardstick for the right register) and open academic ones, all the way to some videos of the monk’s sermons on YouTube, transcribed and translated by a model specialized in Thai, run locally; only after rereading and understanding them did the knowledge base catalog people, geolocated places, events and sources. Many days of work even with AI — but which, on my own, without it, would simply have been impossible.
And then the AI built the actual site — trilingual, static, fast — while I directed the architecture.
But control never left my hands. I reread all the translated Italian, word by word, and from that review I derived a glossary of the terms to use in future translations — with a precise rule for the model: when it meets an ambiguous term that isn’t in the glossary, it must stop and ask me, instead of choosing on its own. I’ll come back to this rule later, because it’s worth far more than the glossary it protects.
From this base an English Wikipedia entry was also born. And that’s no small thing, because there the rules are strict: it’s not enough to “know” something, you have to be able to cite it from authoritative sources — academic, journalistic, institutional. Gathering and vetting them is half the work; AI is precious for finding, reading and comparing them, but deciding “this source holds, this one doesn’t” stays with whoever signs.
What it didn’t do — and couldn’t. It didn’t decide what was important to preserve: that’s mine. And it didn’t verify the facts for me: on the contrary, here the risk of hallucination is especially insidious, because on a local, poorly documented story the model, if you let it, “fills the gaps” with invented plausibilities. There’s even a double trap that no model could untangle on its own. The first I hinted at at the start: in the same village there are two temples a kilometer apart with nearly identical names, that sources, addresses and even the tourists confuse. The second is subtler: “Kruba Noi” isn’t a proper name, it’s an honorific — kruba is the title of revered monks in northern Thailand, noi means “small” — and there have been many Kruba Noi, so a naive search fishes through all their lives. Untangling our monk from his namesakes and our temple from its neighbor was work I did. The AI was the tireless laborer; the keeper of the truth I had to remain, myself.
Online: krubanoi.org — Wikipedia entry (EN)
2. Bringing a village’s history back to light
Castevoli is a hamlet of a handful of souls in Lunigiana, where I have a family house. A medieval village, whose existence is attested as far back as the year 1000, and about which, however, very little is written in an accessible way. The Wikipedia entry, truth be told, I created myself in 2014 — by hand, one piece at a time, with no help other than Google’s search bar. Twelve years later I wanted a site that would tell the village’s story, and to bring that entry up to standard: expanded, precise, based on solid sources. And this is where AI changes the pace.
Arriving at a village is also arriving at a story: the landscape preserves the questions, the sources let you reconstruct the answers.
What the AI did. The same scaffolding as the Thai project — a knowledge base where every fact stays hooked to its source — but with a challenge of the opposite nature. There the problem was a distant language; here it’s time. The traces of a thousand years lie scattered across the archives of small municipalities, each with a different website and a different way of cataloging. The work was combing through them one by one: some sources are only cited by others, some have the manuscript available. For each one you take the PDF (or the image), you OCR it to extract the text, you read it, and you fill the knowledge base — always noting where it comes from.
And here a precise rule governs: on Wikipedia primary sources can be used only within strict limits, without original interpretation. I can know a fact is true because I’ve read the original document — but for an encyclopedic reconstruction I have to find a secondary source that reports it. So half the work is indexing: separating the primary from the secondary, and for each true thing finding who has already told it in a citable way.
And it’s a process that never runs in a straight line: every new source cites others, to be found and read in turn, and it happens that two contradict each other — there a targeted search opens up to beat new paths. The AI is the engine that bears this toil, and keeps in order a thread that would otherwise be lost.
What it didn’t do. The rule I set myself is clear-cut: you don’t write about the village’s history from memory or by inventing. First you read the source, then you write. This because I know all too well what the model’s flaw is: if you ask it “tell me the history of Castevoli,” it tells you — confident, fluent, and largely invented. The value isn’t in its confidence, but in the constraint I imposed on it: speak only of what is documented, and cite the source. AI speeds up the research; it doesn’t replace it, and above all it doesn’t decide what is true.
Online: castevoli.it — Wikipedia entry (IT)
The most important rein: knowing how to stop
Look back at the three key moments I described. The Thai glossary: faced with an ambiguous term it doesn’t know, the model stops and asks me. The two temples and the namesake monks: before any search, a rule forces it to disambiguate, instead of trusting the first match. The village’s history: nothing gets written without having read the source — until the proof is there, you stop.
Three rules born in different contexts, but they’re the same lesson: the rules that count don’t tell the model what to do. They tell it when to stop.
And it’s worth understanding all the way, because it overturns the paradigm we’re used to. In common use it’s always the human who asks the AI questions, and the AI that answers — always, no matter what, even when it doesn’t know: we’ve seen it, it completes, it doesn’t verify. With these rules the opposite happens: it’s the model that stops and asks me a question. The direction of the question reverses, and with it the roles: no longer an oracle that won’t admit it doesn’t know, but a collaborator that knows when it has to ask.
Careful, though: it’s not a spontaneous virtue of the horse. It’s a rein, a rule of the harness — and in my case not even written up front: derived from the errors found in the first review, rereading everything word by word. First the error, then the rule that bars its way. That’s how the harness grows: every corrected error can leave behind a permanent rule — and the most precious rules are the ones that teach the system when to stop.
A necessary caveat: this preference holds for this kind of work. The industry pushes in the opposite direction — every generation of models is trained to be more autonomous, to work longer without human intervention — and for a vertical task, a single long prompt to let run for hours, a horse that stops at every ambiguity would be a defect, not a virtue. It’s always the same rule: it depends on the journey. On a knowledge base that guards true facts, I want a horse that stops. And not only because I want to be the one who decides: every question it asks is a point where I understand the subject better myself. Stopping to ask doesn’t only serve control. It serves learning — mine.
In all these months of work, the moments when the models were most useful to me aren’t the ones where they answered. They’re the ones where they stopped.
Conclusion
Look at what these two projects have in common. They’re not “AI doing something wonderful on its own.” They’re me who, with AI, does on my own things that used to require a team — a translator, an archivist, a researcher, a developer, an editor. But the vision, the constraints, the truth, the reason it’s worth doing: all of that stayed mine.
And there’s a gain you can only see up close. Delegating everything to an autonomous entity is convenient, at times even satisfying — but when the work is done you risk not having learned anything. Holding the reins, as I did here, the opposite happened: it taught me a great deal. About content — a village, a liturgical language, a monk’s life. And about method, which counts even more: not only vetting a source or stopping at the right moment, but something bigger — I learned to rig the horse. To build the harness, and to do it better with every new project, putting this power to work in the most varied fields. It’s a competence that isn’t in the books, you learn it only by doing it — like staying in the saddle: until you mount, and the first few times you end up on the ground, no one can teach it to you.
So I come back to where I started. When someone tells you what “artificial intelligence will do,” stop for a moment and ask yourself two things.
The first: which AI are you talking about? The text completer in the chat, an agent orchestrated inside real work, or one of the many rungs in between? Because they’re different worlds, and confusing them is the root of almost every misunderstanding.
The second, even more important: have you actually used it? Not a couple of questions tossed off, but a real project, with context, constraints, corrections. Because the most revealing line isn’t between the enthusiastic and the skeptical: it runs between secondhand opinions and those earned in the field.
“Artificial intelligence” as a single object, with a single destiny — whether it saves us or ruins us — doesn’t exist. There are different horses, rigged in different ways, for different journeys — and those who steer them. The right question was never “what will it do to us.” It’s always been the same one as thirty years ago, in front of the computer, and then the web, and then the smartphone: what do you do with it.
P.S. — If something stuck with you, the way to really understand isn’t to read more articles like this one. It’s to take something close to your heart — a project, a piece of research, a passion — and try to do it with these tools. Not to see what they can do. To discover what you can do, when something amplifies you.
Appendix — Not one horse, a stable
I come to the most technical part, and to the choice that made the difference for me. I don’t use a single model (I never root for one). I keep several in the stable and put them to work in two opposite ways: sometimes together, in collaboration; sometimes one against the other, in competition — the same task to two or three, and I compare what they bring back.
Not the best model in the absolute: a stable of different tools, to choose and compare for the work at hand.
That’s how I discovered something that whoever talks about “AI” in the singular can’t even see: each is better in different areas. And the differences aren’t only between different companies: they’re also between different versions from the same company. An example from my daily experience: I work with Opus 4.8 with great satisfaction since it came out and I limit myself to checking the result at the end of the session; when instead I use Opus 5 — same house, later and more powerful version — I read everything it does, and I often interrupt it while it agonizes over a useless detail: it’s its greater autonomy — a virtue for vertical tasks, a friction for work like mine.
Then there’s the harness layer, which mustn’t be confused with the model. In my daily use:
- Claude Code is the one I’ve gotten along with best — but the model you slot into it completely changes the outlook.
- Codex I’ve relegated to elementary reading and writing, not complex reviews: in my setup, every so often it makes changes I didn’t ask for.
- Hermes is the Swiss Army knife: you can connect almost any model to it and it’s customizable to the bone. In exchange it has to be configured by hand — a tool for nerds, not for everyone.
And from the competition, over time, real characters emerge — at least in my experience and on the tasks I face:
- GPT 5.6 sol — very good at composing text: concise, precise, proactive, but without overdoing it.
- Opus 4.8 — cautious and diligent, an excellent collaborator: it asks before changing anything.
- Opus 5 — ambitious and stubborn, it works for minutes and hours without stopping; for writing an app from scratch it’s fabulous, for working on a knowledge base it becomes a more difficult collaborator: it takes initiative, sometimes it agonizes and writes verbosely.
- Fable 5 — the orchestrator I’ve gotten along with best, capable and precise; also the most expensive, so to be used sparingly.
- DeepSeek 4 — pragmatic and industrious; on simple tasks, among those I’ve tried, it’s one of the fastest and cheapest.
- Gemini 3.6 — a bloodhound for tracking down sources, where it leverages Google’s research heritage; less reliable on synthesis and conclusions, where in my experience it’s more exposed to hallucinations. For this reason, in my setup, it was forbidden to write or modify files.
And at the opposite extreme of the stable there’s a case almost no one talks about: sometimes not even the generalist online services are enough, and you need a vertical model — specialized in one thing only, so small it doesn’t even know how to use tools, but light enough to run locally on my machine. That’s the case of Typhoon (developed by SCB 10X, the innovation arm of the Thai banking group SCBX), with which I translated the Thai: on its own it would get nowhere, but piloted by an orchestrator it becomes a precious cog in the chain.
Others — GLM, MiniMax, Qwen — I haven’t tried enough yet to weigh in: I’d rather stay quiet than give a judgment I haven’t earned in the field.
One clarification, which is really the heart of the whole article: I’m talking about the models of today, and in six months these names will be old. It’s not a ranking, it’s a method. The point isn’t “which is the best” — there is no best, there’s the most suited to that task, at that moment, inside that harness. And putting them in competition is the most honest way to find out.
P.P.S. — Yes, these lines too I wrote with the aid of AI: same method as the rest — me on the reins, it at a trot, and everything reread word by word a dozen times over. No horse was mistreated during the writing — only, every so often, stopped in time.