Note: This article is a developer-focused adaptation of my original blog post in Italian. You can read the full, in-depth personal reflection here: Il pappagallo stocastico on Nospace.
In the first piece of this series, I explained why, after 40 years spent between systems, code and infrastructure, this hype sounds familiar to me — I've seen it with Wintel, I've seen it with bloatware, I've seen it with every cycle that promises to rewrite the rules and ends up rewriting only the price of the hardware.
But historical skepticism only solves half the problem. Underneath the question "haven't we seen this movie before?" there's a thornier one, one that piece didn't touch: what are these models we talk to all day, really?
Some days ago I stumbled on a YouTube video by Salvatore Sanfilippo — @antirez, for anyone who's built anything with Redis, needs no introduction — in which he very firmly pushed back on a theory from a group of researchers (Timnit Gebru, Emily M. Bender, Angelina McMillan-Major and Margaret Mitchell) who, in a 2021 paper, coined the "Stochastic Parrots" metaphor: an AI that generates fluent, human-sounding text using statistical probability, without actually understanding the meaning of the words it produces. A kind of super-T9, if you will — yes, exactly like the T9 on your phone twenty years ago, just far more sophisticated.
In purely algorithmic terms, this theory is hard to disprove: an LLM builds its responses by predicting the most fitting next token given the context. Sanfilippo, though, argues that today's models have outgrown that theoretical fence. Where do I stand? While I deeply respect Sanfilippo's way of thinking outside the box, I remain skeptical of a genuinely "intelligent" AI — or at least one able to fully understand human language and follow a logical path identical to our own mind's.
So, parrot or intelligence? As often happens in life, it's not all black or white. Maybe the key isn't in defining the machine, but in shifting our focus onto the human side of the equation, splitting it into two fronts.
1. The Developers' Side: The Illusion of Biological Replication
Today, whoever designs artificial intelligence often moves with an ambitious goal in mind: making these machines as similar as possible to the human mind, replicating its structures, its reasoning, its ways of interacting.
From where I stand, that goal will never be fully reachable. There's an insurmountable biological boundary: the purely emotional and sensory component.
We can train a model to define "fear," "love," or "excitement," teaching it to talk about them flawlessly. But the AI will never feel these emotions on its own skin. For a human being, an emotion isn't a string of text or a probability calculation — it's a knot in your stomach, a racing heartbeat, a sudden mood shift that heavily shapes what we think and decide in a split second. That biochemical unpredictability is what makes us human, and it's something no code will ever replicate.
2. The Users' Side: The Anthropomorphism Trap
If on the developers' side there's the illusion of replicating biology, on the users' side the dynamic gets even more tangled. To be clear about it, we need to split users into two broad categories:
- The tech user: the developer, the engineer, the professional who approaches AI pragmatically, treating it purely as a tool to be understood and bent to their productivity needs.
- The everyday user: and to be clear, with no condescension intended — the person who talks to an LLM and interacts with it the way they would with a real person in front of them.
And yet, despite that different starting point, something odd happens. Even the most technically evolved user often falls into the exact same trap as the everyday one: ending up believing the LLM is "intelligent" in the same way we are. It's a genuine cognitive trap, one that indirectly ends up feeding the same ambition of biological replication described above. The line doesn't run between insiders and the general public — it runs inside each of us, every time we forget what we're actually talking to.
3. When Even the People Selling the Hardware Blur the Line
If you think this confusion only concerns the end user, September 6th of this year gave us a vivid demonstration of the opposite, straight from the top of the chain of command. Nvidia's CEO Jensen Huang wrote on X: "AGI has arrived," crediting OpenAI's GPT-6 Astra and promising another 400,000 GPUs on the way. It wasn't even the first time — back in March he had already declared, on a podcast, "I think we've achieved AGI," but with a definition entirely his own: an AI capable of launching and growing a billion-dollar business on its own.
Notice the shift: no longer "thinks like a human" — the historical, epistemological benchmark for AGI — but "produces economic value the way an organization would." It's exactly the same confusion described above, just moved from the individual user up to whoever is steering the entire industry: mistaking the ability to do for the ability to think, because the former is far easier to sell than the latter.
And the numbers, once someone independent checked them, tell a different story:
- 99.9% — Astra's score inside OpenAI's own proprietary harness.
- 62.7% — Astra's score on ARC Prize's independent standard, published the very same day.
The gap doesn't come from the model's weights. It comes from the entire scaffolding of tools, agents and external checks built around it. It's proof, numbers in hand, that what we're increasingly measuring when we shout "AGI" is the ecosystem around the model, and less and less the model itself.
4. Collaborating with the Alien, Not Worshipping the Oracle
My view is different. I think LLMs do express a form of "intelligence," but it's an intelligence with a mind radically different from, and distant from, our own.
We need an effort of intellectual humility. First, accept that there are still blind spots in their internal processes — we can't always explain why a model answers one way rather than another, not even the people who built it. Second, stop treating these machines as "oracles" to be blindly trusted, the way we would with the most gifted human beings.
The end goal shouldn't be passive reliance, but active collaboration. An LLM doesn't think the way we do because it doesn't have our biology, our flesh, or the filters of our millennia-old culture. That's exactly what makes it so valuable: we should use it as a mirror for our own intuitions. Talking with an AI should feel like talking with an alien intelligence — a way to get completely destructured viewpoints, free of our historical and cultural biases, capable of opening our eyes to alternatives we hadn't even considered.
On this front, alongside the theoretical reflection, I'm also working hands-on on an open-source project that tries to build the right architecture for this kind of collaboration — a digital "thalamus" that filters and structures context instead of drowning the model in a chaotic Token Soup. Anyone who wants to get their hands dirty with the practical infrastructure can find it in my profile page: Designing an LLM Context Orchestrator & Cognitive Exoskeleton.
Conclusion: The Question That's Still Open
There's one point, though, that while writing this piece I realized I don't have a clear answer to yet — and maybe no one really does.
If we accept that this is an alien mind, one that doesn't replicate our biology or our continuity of experience, then maybe the entire vocabulary we use to talk about it — "teaching it," "learning from its mistakes," "developing intuition" — is the wrong metaphor. A child learns to walk by falling, and that fall stays with them, integrated, for good. An LLM has nothing like that: every conversation closes, and with it, everything it seemed to have "learned" along the way disappears.
So the real question, the one worth coming back to, is: if an alien intelligence has no continuity of self, what does it actually mean to teach it something?
I don't have a clean answer yet. But it's the question I'll pick up in the next chapter of this series.
Have you caught yourself anthropomorphizing an LLM lately, treating it as more "understanding" than it really is? Or do you think the alien-mind framing goes too far in the other direction? Let's discuss in the comments below!
This is the second piece in the series that started with Is History Repeating Itself? From the PC Era to the AI Era. The next chapter will tackle exactly this question, with a few more concrete cases along the way.
