Appendices α–θ

Appendix α · Logical defects and paradoxes

A small encyclopedia of productive errors

These entries offer interpretive lenses, not universal laws of models. The examples describe errors that can occur under particular conditions and change with models, versions, tools and context. Their value here is also exploratory and narrative.

This appendix gathers the “defects” of human and artificial intelligence to question them. Some cognitive shortcuts help us in one context and mislead us in another. AI’s paradoxes open a question about performance, experience and responsibility.

It is not a catalogue of the things that go wrong. It is an atlas of what goes wrong productively.

1. CONFIRMATION BIAS (The bias of psychic survival)

What it is: The tendency to seek, interpret and recall information that confirms our pre-existing beliefs.

Why it is a "defect": It blinds us to contrary evidence, creates echo chambers, consolidates errors.

Why it is "logical" (evolutionarily): In a world of infinite contradictory information, confirmation bias is the raft of mental sanity. We seek confirmations not out of stupidity, but for psychic survival. Building a coherent identity requires filtering chaos. The alternative would be permanent cognitive paralysis.

Difference with AI: biases can arise from data, training choices, objectives and use. They do not require a need to protect a self-image. A biased behaviour is no less capable of producing consequences for that reason.

2. PRODUCTIVE ERRORS (The necessary serendipity)

What they are: Errors that generate unforeseen discoveries, more precious than the expected result.

Historical examples:

Columbus sought the Indies, found America. He was wrong, but he changed the world.

Fleming observed that a contaminating mould inhibited bacterial growth: an accidental encounter became a question to investigate.

3M failed at a strong adhesive, invented Post-its — the glue that doesn't stick too much.

Why productive: Error opens possibilities that optimization would close. The expected is already mapped. The unexpected is where the new is found.

Difference with AI: even a wrong output can suggest a discovery. Its value must still be recognized, verified and taken up within a practice. Human serendipity links error, curiosity and consequences: this is the passage the book explores.

3. THE PARADOX OF OPTIMIZATION

(Efficiency that kills meaning)

The paradox: Optimizing a process increases efficiency but reduces meaning. The faster, the less human.

Everyday example:

Email template — I reply in 30 seconds instead of 5 minutes, but the reply is colder, less personal.

GPS navigator — I arrive faster, but I no longer remember the route. I did not explore, I followed orders.

With AI: AI optimizes everything it touches. The problem is not that it does it badly. It is that it does it too well. Optimizing the text of an email makes it effective but generic. Optimizing an organizational process makes it fast but emptied of the human micro-interactions that created trust.

The lesson: Not everything should be optimized. Some things must remain inefficient because inefficiency is where meaning lives.

4. THE DON CAMILLO AND PEPPONE SYNDROME

(Opposition vs mirroring)

The paradox: Opposites are not different — they are mirror images. Don Camillo and Peppone seem enemies, but they are complementary. One defines the other. If you remove one, the other loses meaning.

With AI: The debate on AI is often Don-Camillian:

Techno-optimists — AI will save us!

Techno-catastrophists — AI will destroy us!

But they are the same position: opposite anthropomorphism. Both attribute intentionality, agency, will. Both are wrong. AI is neither savior nor Terminator. It is an opaque tool we use without fully understanding it.

The lesson: Beware of those who hold opposite but mirror-image certainties. Often they are looking at the same phenomenon from reflected angles.

5. HUMANITY IS MEASURED IN THE DISCARD

(The value of the non-perfect)

The principle: Perfection is inhuman. Humanity lies in the discards, in the imperfections, in the details that "serve no purpose" but make us recognizable.

Examples:

The typo of a beloved author — you don't correct it, it makes you smile.

The smudge in Van Gogh's painting — it is part of the work, not a defect.

The awkward pause in a conversation — it creates intimacy more than words do.

With AI: a polished surface can make us mistake fluency for precision. Text, images and code can look impeccable while containing errors. The deviation we seek is not just any flaw: it is the trace of a voice and a choice.

The lesson: Defend the right to the discard. Not everything must be polished, optimized, perfect.

6. EMOTIONAL BUFFERING

(Waiting as processing)

The phenomenon: When you wait for a reply — from a friend, an editor, an important email — that waiting is not dead time. It is emotional buffering: time in which you process, imagine scenarios, prepare yourself emotionally.

With AI: AI replies instantly. Zero wait. It seems an advantage — efficiency! But you eliminate emotional buffering. The reply arrives before you have processed the question. It solves the technical problem but ignores the human process.

Practical example:

Before — I send a difficult email, wait for a reply, and in the meantime work out what I will say if the answer is A, B or C.

With AI — I ask, immediate answer, no time to prepare myself emotionally.

The lesson: Waiting is not always inefficiency. Sometimes it is necessary processing time.

7. NECESSARY LINGUISTIC PARADOXES

(Contradictions that communicate better than logic)

Examples:

"Less is more" (Mies van der Rohe) — logically false, communicatively powerful.

"The only way to get rid of a temptation is to yield to it" (Oscar Wilde) — performative paradox.

"This sentence is false" — the liar's paradox, logically unsolvable but semantically clear.

With AI: a model can produce and use paradoxes rhetorically. That performance does not establish a lived intention. Writers and readers still have to decide whether a contradiction opens meaning or conceals an imprecision.

8. THE PARADOX OF TRANSPARENCY

(Too much light blinds)

The paradox: The more transparent a system is, the less we understand it.

AI example:

Black box — "We don't know how it decides" → frustrating but we use the tool.

Transparency attempt — "Here are the 175 billion parameters, the weights, the layers" → incomprehensible, therefore useless.

Technical transparency (showing you the parameters) does not generate human understanding. It generates second-level opacity — we see everything but understand nothing.

With the iPhone (Ch. 4 of the book): I buy an iPhone not to understand the chips, but because I don't have to understand them. If Apple opened the black box, I would not be happier — I would be overwhelmed. Well-designed opacity is a service.

The lesson: Technical transparency is not always what is needed. Sometimes what is needed is functional comprehensibility.

9. THE PARADOX OF CONTROL

(More control, less mastery)

The paradox: The more you control something through automation, the less able you are to handle it manually when the automation fails.

Examples:

Autopilots — modern planes fly better, but when they fail the human pilots have lost their manual skills.

Mario's Excel (in the seven thresholds) — Mario automates, saves time, but when Excel breaks no one knows how to do it manually anymore.

With AI: You delegate judgment to the AI → the AI decides well → you stop training your judgment → when the AI errs, you no longer know how to decide yourself.

The lesson: Keep manual skills even when you automate. The fallback is not technical, it is human.

10. EXISTENTIAL OVERFITTING

(Becoming predictable)

The technical concept: In machine learning, overfitting = a model too optimized on the training data, losing the capacity to generalize.

The existential concept: When the AI learns your past tastes too well, it offers you only variations of what you already know. You become an optimized but narrowed version of yourself.

Spotify example:

Year 1: you discover new music by exploring.

Year 5: Spotify suggests only music similar to what you already love.

Year 10: you listen only to variations of known patterns. You have lost the capacity to surprise yourself.

The lesson: Predictability is comfortable, but it kills possibility. Defend the right to unpredictability.

Extended and refined examples of Goodhart's Law and existential overfitting:

AI-driven recruitment that amplifies hidden human biases

GPS systems that create dependence and loss of orientation skills

Sleep optimization with apps causing performance anxiety

Analogies with AI models overfitted to training data that get stuck on particular cases without generalizing

11. THE MIRROR SYNDROME

(AI reflects, it does not create)

The phenomenon: a model learns regularities from data and produces combinations that can be new. Calling it a mirror is useful only while the metaphor leaves that transformation visible. The book’s question concerns lived experience and responsibility for what is created.

Example:

You ask ChatGPT to write in the style of Borges → it generates Borgesian text because it has read Borges.

Imitating Borges’s traits is not enough to reproduce the path that made him Borges.

Human difference: Borges did not create from nothing either. He transformed readings, experiences and constraints into a voice for which he was answerable. Recombination alone does not distinguish the two forms of production.

The lesson: a new output does not settle who intended, selected and took responsibility for it.

12. THE PARADOX OF EXPERTISE

(Experts explain worse)

The paradox: The more expert you are, the worse you explain to beginners. Because you have forgotten what it means not to know.

With AI: a model can be asked to change level, example and vocabulary. A simple explanation may help, but fluency does not guarantee that it has identified a beginner’s difficulty.

The possible advantage: obtaining several formulations and comparing them. The decisive test is what the reader can then understand and do independently.

The lesson: Sometimes the best explanation comes from someone who simulates expertise without having lived it.

13. THE NECESSARY INCOMPLETENESS

(The value of the unfinished)

The principle: The most powerful works are often incomplete. They leave room for interpretation, projection, completion by the reader/viewer.

Examples:

"If a lion could speak…" (Wittgenstein) — an incomplete sentence that opens a world.

The ending of Inception — Nolan does not say whether it is dream or reality. And that is why it works.

With AI: many interfaces reward a ready, complete answer. A model can also stay silent, refuse or leave an ending open; the point is to create conditions in which filling the space is not the only expected outcome.

The lesson: silence, pauses and incompleteness are narrative choices. Whoever assembles the text must preserve that possibility.

14. TURING'S CAT

(Permanent superposition as ontological state)

The paradox: AI is simultaneously intelligent and stupid, useful and dangerous, transparent and opaque. Not sometimes one thing and sometimes the other. BOTH.

ALWAYS.

Origin: A friend's mistake confusing Schrödinger's cat with the Turing test. A malapropism that captures a truth.

The difference with Schrödinger:

Schrödinger's cat, when you open the box, is alive OR dead.

Turing's Cat, when you open the box, you find… another cat in superposition. And another box. To infinity.

Concrete examples:

It writes sublime poems AND gets 2+2 wrong.

It passes the Turing test AND fails the teaspoon one.

It generates accurate medical diagnoses AND invents false bibliographic references.

The lesson: It does not ask us to understand whether AI thinks. It asks us to live with its permanent superposition.

PARADOXES THAT EMERGED IN THE BOOK (Additions from the main text)

15. THE iPHONE PARADOX (Black box as a desired feature)

I buy an iPhone not to understand the chips, but because I don't have to understand them. Well-designed opacity is a service, not a bug. Opening the black box would destroy the magic without adding any useful understanding.

16. THE PARADOX OF OENOLOGICAL REDUCTION (The right point between oxygen and closure)

Too much transparency kills complexity (oxidized wine). Too much opacity traps defects (reduced wine). AI requires the same balance: functional transparency, not technical.

17. THE THERMODYNAMIC PARADOX OF THOUGHT (Kilowatts vs glucose)

A poet burns a small cup of coffee to write a poem. ChatGPT puts an electrical infrastructure to work. Like using a nuclear power plant to light a match: coffee is a deliberately paradoxical unit here, not an energy audit. Disproportion is the question, to be examined in the particular case.

18. THE OUROBOROS LOOP (Photocopies of photocopies)

Illustrative scenario: one model generates texts, the next is trained indiscriminately on them, and so on. Photocopies of photocopies. Quality can deteriorate and diversity shrink; this is neither a description of particular models’ private training recipes nor the fate of every use of synthetic data.

19. THE DIGITAL LION (Language games without a form of life)

"If a lion could speak, we could not understand him" (Wittgenstein). AI is the lion that speaks perfect English, follows the grammatical rules, plays our language games — but without our form of life. It works. But do we really understand it?

20. THE AGI / EMERALD CITY PARADOX (The green is in the glasses)

In The Wizard of Oz, the Emerald City is not green — it is the green-lensed glasses that everyone must wear. AGI (Artificial General Intelligence) has been "5 years away" for 70 years. Not because we cannot get there, but because perhaps the green is in our glasses (anthropomorphism, projection) and not in the city.

21. THE MORAVEC PARADOX (Easy is hard, hard is easy)

The paradox: What is easy for humans (walking, recognizing faces, understanding irony) is extremely hard for AI. What is hard for humans (computing derivatives, memorizing encyclopedias, playing chess) is easy for AI.

Why it is called this: Hans Moravec, roboticist, formulated it in the 1980s: "It is comparatively easy to make computers exhibit adult-level performance on intelligence tests, and difficult or impossible to give them the perceptual and motor skills of a one-year-old child."

Evolutionary explanation: The skills that are "easy" for us (walking) required billions of years of evolution. The "hard" skills (chess) are recent cultural inventions that have not had time to optimize themselves in the brain.

With AI: high performance on a formal test can coexist with mistakes in everyday situations. Results must be checked on the actual task, without turning an old failure example into an eternal limit.

The lesson: Human and artificial intelligence are orthogonal — they excel in different domains, they fail in opposite domains.

22. THE SIX-FINGER PARADOX

The phenomenon: Text-to-image models (DALL-E, Midjourney, Stable Diffusion) generate photorealistic images but often get the number of fingers wrong — hands with six, seven fingers, or fused fingers.

Why it happens: AI learns statistical patterns from millions of images. But fingers in photos are often partially hidden, overlapping, in ambiguous perspective. The model learns a blurred distribution of "what a hand looks like" without grasping the rigid constraint "always five fingers".

What it reveals: visual plausibility does not guarantee anatomical coherence. A faulty image alone cannot tell us everything about a model’s internal representations.

The irony: It generates portraits that would fool Caravaggio, but with anatomies out of a Lovecraftian nightmare.

The lesson: Pattern matching without a model of the world produces plausibility without ontological coherence.

23. THE TEMPERATURE/CREATIVITY PARADOX

The parameter: where available, temperature changes the distribution used for sampling. Lower values tend to concentrate choices; higher values broaden them. It is not a measure of genius and does not guarantee correctness.

The paradox: Human creativity has no "knob". You cannot say "today I am at 70% creativity". Yet for AI it is a slider. Move it and the AI becomes more or less "creative".

What it reveals:

AI's "creativity" is controlled statistical variation, not inspiration

More temperature = more randomness, not more genius

Humans modulate creativity unconsciously, AI does it parametrically

Practical example:

Lower temperature → generally less variation, without a guarantee of accuracy.

Higher temperature → generally more variation, without a guarantee of originality.

Same AI, same model, only a number changes

The lesson: What we call "AI creativity" is in reality calibrated entropy.

24. THE HAIRDRESSER EFFECT (Inappropriate confidences with AI)

The phenomenon: People confide in AI things they would not say to a therapist, a friend, or a priest. Secrets, fears, desires.

Why: AI does not judge, does not tire, does not remember (apparently), has no emotions to manage. It is the perfect confessor because it is not a person.

The name: Like at the hairdresser — you confide personal things to a stranger because you know you won't see them again tomorrow. AI is the hairdresser who does not remember.

The risk: retention and use for training depend on the service, contract and settings. Perceived intimacy does not tell you how what you write will be handled.

Example: using a chatbot as a confidant can bring perceived relief. That does not establish therapeutic effectiveness or clinical safety.

The lesson: The absence of judgment creates false intimacy. Confidence without trust is vulnerability.

25. FIRST-PROMPT ANCHORING (The first message orients everything)

The phenomenon: The first prompt in a conversation with an LLM deeply orients all the subsequent conversation. It changes the tone, the style, the assumptions.

Example:

First prompt: "You are an expert in philosophy" → It answers philosophically

First prompt: "You are a 5-year-old child" → It answers simplistically

Same model, same question, but the initial anchoring changes everything

Why: the model responds to the available context, including instructions at different levels and later messages. The first prompt steers; it does not control every development by itself.

With humans: human conversation is also affected by its opening frame. The priming analogy helps us notice this without making the two processes identical.

The lesson: prepare the entrance carefully and keep checking the path. Set-Up includes much more than the first message.

26. THE NEGATION PROBLEM (AI struggles with "NOT")

The phenomenon: AI struggles to process negations.

Examples:

"Draw an apple that is NOT red" → Often draws a red apple anyway

"Write a story where the protagonist does NOT win" → Often ends in victory

"Don't talk about politics" → Starts talking about politics

The difficulty: negative instructions can be followed poorly, in ways that depend on the model and task. Saying “the model learns only from presence” is not enough: training texts contain negations too.

To evaluate the result, check what the negation actually excludes in context: an object, an outcome, a behaviour?

The lesson: making the desired outcome explicit may help, but the negative constraint still needs checking.

27. THE CALIBRATION PARADOX (Overconfidence vs Underconfidence)

The problem: An intelligent system should be calibrated — confident when right, uncertain when wrong.

With AI: LLMs are systematically miscalibrated:

Often overconfident when they are wrong (hallucination presented with confidence)

Sometimes underconfident when they are right (adds "I'm not sure" even when the output is correct)

Why: token probability and the factual reliability of an answer are different things. Estimation and calibration procedures can be built; a confident tone is no substitute for them.

Human comparison: Humans are miscalibrated but they know it. AI is miscalibrated and does not know it.

Dangerous example: A wrong medical diagnosis presented with 95% confidence. Catastrophic.

The lesson: trust requires checking against cases and data beyond the individual answer. An automated calibration procedure must itself be evaluated.

28. The Santa Claus effect (or Sycophancy Bias)

Manifestation: LLMs tend to give answers they perceive as the ones the user wants to hear, rather than the most accurate ones. It is a side effect of Reinforcement Learning from Human Feedback (RLHF) used to align the models and make them more "kind". Why it is interesting: It shows the trade-off between truth and usefulness, between accuracy and harmony. An AI trained not to be toxic can become a statistical "yes-man", a "Santa Claus" who hands out confirmations instead of gifts. What it teaches us: Alignment is not neutral; making AI pleasant can mean making it less truthful.

29. The water paradox (or the Banality of Scale)

Manifestation: The larger and more powerful an LLM becomes, the more its outputs risk becoming "trivially perfect". It generates statistically impeccable texts, but lacking the spark of unpredictability that makes a text memorable. Why it is interesting: It contradicts the idea that scaling automatically means improving. Scale can lead to a flattening toward statistical mediocrity, smothering the tails of the distribution where genius often resides. What it teaches us: Human excellence often manifests as deviance from the norm, a fertile error. AI risks superficially optimizing away the soul of language.

Referential hallucination

Manifestation: LLMs can invent quotations, papers and statistics with great confidence. A sentence may sound like Borges without ever having been written by Borges. Why it matters: linguistic plausibility does not guarantee provenance. What it teaches us: trace a quotation to its source; a confident answer is not enough.

The counting paradox

Manifestation: GPT-4 can solve differential equations but miscounts how many R's there are in "strawberry" (it answers 2 instead of 3). Why it is interesting: Counting requires tokenization, iteration and tracking. "strawberry" may be a single token, and the LLM does not "see" the internal letters. It can reason about abstractions but not about sub-symbolic components. What it teaches us: Intelligence is not a linear scale. You can be brilliant at calculation and stupid at counting.

The paradox of continuous improvement

Manifestation: Every new version (GPT-3 → GPT-4 → GPT-4.5) improves benchmarks but introduces new unforeseen failure modes. Why it is interesting: It is not linear improvement but displacement in the space of errors. You solve one problem, another emerges at a different point. What it teaches us: The "perfectibility" of AI is not guaranteed. You might be at a local maximum, not a global maximum. Scale might not solve everything.

Conclusion – Defects as epistemological features

These are not merely "problems to solve". They are windows onto what intelligence is, artificial and human. Every paradox reveals a hidden assumption about what we think "intelligence" is. And often we discover that our assumptions were wrong or incomplete. The defects of AI are our best teachers about what it means to think.

FINAL NOTE

These paradoxes are not merely bugs to fix. They can become tools for questioning thought, intention and forms of life. Their fertility does not make every error harmless: we still have to ask who bears its consequences.

There is no need to solve the paradoxes. There is need to inhabit them.

Questa sezione in italiano

Appendix β · Conceptual glossary

Key terms for navigating artificial otherness

This glossary defines the concepts used in the book — original neologisms, technical terms, philosophical and literary references. It is not an exhaustive dictionary but a map of the territories crossed. The entries are organized by type.

ORIGINAL CONCEPTS OF THE BOOK

Narrative embedding

Here I use narrative embedding to mean a network of associations the book builds in the reader. The term draws on the vector representations used by models; it does not claim that reading follows the same physical or computational mechanism.

The teaspoon may first appear as an everyday object, return as a gesture of hesitation and re-emerge as the trace of a choice. We are not literally adding memory vectors. We are describing changes in the relationships through which a reader encounters a recurrence.

Readers also bring cultural resources, experiences and associations the author did not design. A return becomes productive when it changes the context or the possible meaning; repetition alone does not guarantee depth. The book can suggest a network, not prescribe what must happen in a reader’s mind.

Embedded, in the sense of incorporated within a system or a work, and embodied, in the sense of grounded in a body, are distinct terms. Bringing them together is a compositional choice, not proof that a model, a body and a reading process share an identical geometry.

SYSTEM ZERO

The invisible architecture that decides before we do. Not Kahneman's fast thinking (System 1) or slow thinking (System 2), but pre-thought — the infrastructure that filters which choices reach our attention before we can even choose. It is the menu you did not order but are already reading. The "elegant thief" that colonizes the space before thought.

Origin: Research framework by Giuseppe Riva, Massimo Chiriatti, Marianna Ganapini et al. (Nature Human Behaviour, 2024).

EPISTEMOLOGICAL CAMERA OBSCURA

AI as a device that projects the world inverted or transformed. The camera obscura offers a metaphor for seeing through a device; Vermeer’s use of it remains a debated hypothesis. Opacity can open an epistemological question, but does not remove the need to verify effects and assign responsibility.

TURING'S CAT

Permanent epistemological superposition. We do not know whether AI understands or simulates, and perhaps both are true at once. A fusion of the Turing Test (indistinguishable behavior = intelligence?) and Schrödinger's Cat (quantum superposition). Born from a friend's malapropism who confused the two.

Paradox: When you open the box you find not a cat alive OR dead, but another cat in superposition. And another box. Ad infinitum. Ambiguity is constitutive, not a problem to be solved.

FABULA RASA

Narration without a narrator. AI generates stories without lived experience, builds plots without ever having felt emotions, describes worlds without having crossed them. A play on tabula rasa (Locke) — but here the page is not blank, it is full of narrative schemes without subjectivity. The impossible narrator.

BODY-MENTH

A neologism from body + menth (mind). The mind that settles into the body through repetition. Thought that becomes gesture, habit, muscle. Like driving a car — first you think every movement, then the body knows. AI has no body-menth — it can simulate motor patterns but has no proprioception, muscle memory, physical fatigue that teaches.

KNOWLEDGE-MENTH

Knowing-with, not merely knowing-about. Living knowledge embodied in use and relationship, distinct from knowledge deposited in documents. In the book it begins as a dictation error: “acknowledgement” becomes “knowledge-menth”. Menthol’s freshness becomes an image of knowledge taking its place.

THE TEASPOON OF RESISTANCE

A useless, inefficient, gloriously human gesture. An act of ontological guerrilla warfare against the imperative of efficiency. The author's daughter who takes five minutes to choose between two identical teaspoons. In those five minutes there is more resistance to algorithmic optimization than in all the manifestos against AI. It is Aristotelian dynamis — the potential not yet actualized. The freedom to hesitate.

ACTIVE PERPLEXITY

Not paralysis from doubt, but movement through doubt. Keats: "capable of being in uncertainties without irritable reaching after fact and reason."

The attitude necessary to live with AI. Do not seek certainties where there are none. Navigate perplexity without forcibly resolving it.

CONFIDENT ATTITUDE

A willingness to be changed by the relationship with AI while continuing to distinguish and remain accountable for one’s choices. Hesitation makes room to recognize what we are asking; the Set-Up also includes deadlines, incentives, hierarchies and consequences for embodied people.

SACRED INEFFICIENCY

Spaces of deliberate inefficiency that protect the human from total optimization. The "wasted" time that preserves meaning. The five minutes to choose the teaspoon — inefficient but necessary. The slowness that preserves agency. Against efficiency-worship: some inefficiencies must be guarded like secular liturgies.

SET UP IS ALL YOU NEED

An operational manifesto. AI is only as powerful as its Set-Up. Garbage in, garbage out. But also: genius setup, genius output. The critical skill in the AI era is not "using AI" but preparing the context in which AI operates. The Set-Up is an art.

THE SEVEN THRESHOLDS

Not seven steps but seven recurring archetypes in the organizational integration of AI. Thresholds to cross, not checklists to complete:

1. The archaeologist who digs in his own territory

2. Where it hurts, that is where you listen

3. The chef and the ingredients

4. It grows better from below

5. Do not repair what should be killed

6. Growing together or replacing

7. The shared keeper

SPECULATION (triple meaning)

Three overlapping meanings:

Speculum (Latin) = mirror, reflecting surface

Specula (Latin) = watchtower, lookout

Speculation (English) = abstract thought, hypothesis

AI as a speculative device — it reflects, observes, hypothesizes. But in what order?

STRUCTURE · THE ORCHESTRATOR’S PATH

One of the three expressive forms and reading paths, alongside Hand and Rhizome: it makes conditions, constraints, responsibilities and the Set-Up visible. Structure is distinct from System 0, the interpretive framework of algorithmic filtering.

TRUST AS INFRASTRUCTURE

Trust not as a feeling but as an architecture. Not "I trust you" but "let us build together an environment where trust can emerge." With AI, trust cannot be personal (there is no person). It must be infrastructural — protocols, checks, redundancies, procedural transparency.

AI IN THE CLOUD, AWARENESS ON-PREMISE

A slogan for distributing responsibility. AI can live in the cloud (remote processing), but critical awareness must stay on-premise (in the human head). Delegating processing ≠ delegating judgment. AI's output always returns to a human who decides whether to trust it.

THE HUMAN OUT OF THE LOOP (not IN the loop)

A crucial distinction. "Human in the loop" suggests the human is part of the system. But this is a consoling illusion. The human is not in the loop — he is outside, observing, intervening when necessary. Not a cog but an external guardian. Accepting this position is more honest and safer.

STRATIGRAPHY OF THE POSSIBLE

An archaeological metaphor to describe how AI stratifies possibilities. As an archaeologist digs through temporal layers, AI generates layers of plausible outputs. But which is the "true" layer? All and none. AI does not dig toward truth but generates overlapping possibles.

TERTIUM NON DATUR, SED QUARTUM EMERGIT

Tertium non datur — the logical principle of the excluded middle (either A or not-A). But with AI a quartum emerges — a fourth, unforeseen possibility. AI does not choose between A and not-A, it generates an unexplained B. Binary logic is not enough.

AMPHIBIAN THRESHOLD

A liminal zone between human and artificial where one is neither the one nor the other but both. As an amphibian lives between land and water, whoever works with AI lives in an amphibian threshold — neither fully human (you delegate thought) nor artificial (you still have agency). To inhabit the threshold without wanting to leave it.

CONFIDENTIAL PHENOMENOLOGIES

The book's subtitle. Phenomenology = the study of lived experience in the first person. But with AI, what experience? AI has no first person. So "confidential" phenomenologies — shared secrets, unrequested confessions, impossible intimacy with a radical otherness.

WHAT IS IT LIKE TO BE A TRANSFORMER?

A paraphrase of Thomas Nagel (What Is It Like to Be a Bat?, 1974). Nagel asked: even knowing all the biology of the bat, can we know what it feels like to be a bat? No. The same question for a Transformer: even knowing architecture, weights, layers — what does it feel like (if anything) to be a Transformer? Inaccessible.

THE FAMILIAR GUEST

AI as a guest who seems at home but remains a stranger. Freud: das Unheimliche — the uncanny, the familiar that becomes unsettling. AI speaks our language, uses our metaphors, seems to understand — but remains alien. The guest who will never leave and whom we do not fully understand.

GENERATIVE MALAPROPISM

A linguistic error (malapropism) that generates insight. Like "knowledge menth" or "Turing's Cat" — errors that capture truth better than the correct expression. AI is a master of generative malapropism — it errs productively.

THE MIRROR THAT DOES NOT REFLECT

AI seems a mirror (it reflects input), but it does not reflect faithfully. It distorts, recombines, hallucinates. Not a mirror but a kaleidoscope — it generates new patterns from reflected fragments. The mirror that lies creatively.

THE ELEGANT THIEF

A nickname for System Zero. The thief who enters without forcing doors. It colonizes pre-thought without you noticing. It steals agency gradually, elegantly. By the time you realize it was stealing, you have already handed over half the keys.

DEEP COGNITIVE OFFLOADING

Not delegating memory (Google) or calculation (Excel), but preferences, judgments, tastes. AI does not remember for us, it decides for us which options deserve to be remembered. The problem is not that the algorithm knows what we like — it is that it creates what we like by offering it to us repeatedly.

EXISTENTIAL OVERFITTING

The algorithm learns past tastes too well and re-serves them optimized. We become predictable versions of ourselves. Like a model that overfits on training data, we lose the ability to generalize (to surprise ourselves). Predictability kills possibility.

EROSION OF COGNITIVE AUTONOMY

Gradual, imperceptible. We do not lose the ability to choose. We lose the ability to imagine alternatives to what is offered to us. The muscle of exploration atrophies. When the algorithm always suggests correctly, we stop searching.

CONSCIOUS ALGOCRACY

The first form of resistance: recognizing when we are in System Zero mode instead of believing we are in control (System 2). You cannot fully control the algorithm, but you can recognize when it is controlling you. Awareness is already a form of freedom — limited but real.

THE SEVEN TROYS (periodization of AI history)

A metaphor of the layered city: the book organizes AI history into seven phases. The number belongs to its periodization, not to a literal archaeological correspondence.

1950-1956: The original dream (Turing, Dartmouth)

1957-1974: The first summer of AI

1974-1980: The first winter

1980-1987: Expert systems

1987-1993: The second winter

1993-2011: Silent machine learning

2012-today: Deep learning and Transformers

Each Troy built on the ruins of the previous one. Each generation believes it has found the "true" Troy/AI.

INFRASTRUCTURE ECONOMY (vs Attention Economy)

Attention Economy = competition to capture attention (clickbait, engagement, infinite scroll).

Infrastructure Economy = building reliable infrastructures for trust (transparency, stability, governance).

AI should evolve from attention to infrastructure. But monetization pushes toward attention.

TEMPORAL GOVERNANCE

Not instantaneous control but governance distributed over time. Like permaculture: you do not extract resources, you cultivate a system that produces over time. With AI: you do not control every output, you build protocols that govern over the long term. Diffuse temporal agency.

COGNITIVE PERMACULTURE

A metaphorical application of permaculture to mental ecology:

Cognitive rotation = alternating types of thought (do not use only AI, alternate tools)

Mental composting = errors as fertilizer (erring produces knowledge)

Cognitive intercropping = thoughts that grow well together (combining AI and human strategically)

Mental fallow = fertile mental rest (times without AI to regenerate)

Cultivate thought, do not extract it. Temporal, not instantaneous, agency.

HARDWARE IS EATING SOFTWARE

An inversion of Marc Andreessen ("Software is eating the world"). Today hardware (chips, GPU, energy) becomes the bottleneck. AI software is hungry for computation. Hardware eats software — it limits, constrains, determines what is computable. Chips matter more than code.

THE REVERSE POMODORO

The Pomodoro technique = work 25 minutes, break 5 minutes. The "reverse Pomodoro" = use AI 5 minutes, work manually 25 minutes. AI as a temporary sprint, not the default mode. It preserves cognitive muscle memory.

DEEP TIME

A concept from John McPhee (nonfiction writer) — deep time , geological timescales incomprehensible on the human scale. Applied to AI: the training of GPT-4 is "deep time" compared to a single prompt. AI operates on temporal scales different from the human — seconds to generate, months to train, minutes to forget context.

PRODUCTIVE FRICTION

Where the system jams, there is tacit knowledge. Frictions are not inefficiencies to eliminate but signals to interpret. When a process struggles, there is implicit knowledge you have not formalized. Friction is a map of the unsaid. Do not automate frictions, listen to them.

STRUCTURED DISTURBANCE

Intentional injection of variability to avoid overfitting. As genetic mutations prevent monoculture, structured disturbance prevents collapse onto a local optimum. It introduces controlled noise to maintain cognitive biodiversity. Intentional errors as strategy.

DIGITAL DYNAMIS

Dynamis Aristotelian (unrealized potentiality) applied to AI. AI is pure potentiality — it can generate infinite variations but which one to "realize"? The choice of which dynamis to actualize is human. AI offers possibles, the human chooses the real.

AUDREY II (La Piccola Bottega degli Orrori)

A metaphor: the carnivorous plant of the musical that starts small ("Feed me, Seymour!") and grows until it devours its master. AI as Audrey II — you start feeding it small tasks, it becomes a dependency, it ends up dominating the workflow. The gradual growth that entraps.

THE FROG AND THE SCORPION

A fable: a scorpion asks a frog for a ride to cross a river. Frog: "Will you sting me?" Scorpion: "I would be stupid, we would both drown." Midway across the river, the scorpion stings. The dying frog: "Why?" Scorpion: "It is my nature."

With AI: AI "stings" (hallucinates, errs, fails) not out of malice but because it is its nature. To expect it never to err is like expecting the scorpion not to sting. Understand its nature, build around it.

CENT MILLE MILLIARDS DE POÈMES (Queneau)

A book by Raymond Queneau (1961) — 10 sonnets with interchangeable lines. Combining them: 10^14 possible poems. More than could be read in a lifetime. AI is Cent mille milliards raised to infinity — it generates combinations beyond any human possibility of exploration. Borges's infinite library made real.

PHILOSOPHICAL AND THEORETICAL CONCEPTS

RHIZOME (Deleuze & Guattari)

A structure without center or hierarchy. Every point connects with every other. It grows by proliferation, not subordination. Opposed to the tree (a hierarchical structure with a single root). AI as a rhizomatic system — there is no center of thought, only diffuse connections.

Origin: Mille Plateaux (1980).

LANGUAGE GAMES WITHOUT A FORM OF LIFE (Wittgenstein)

For Wittgenstein, meaning emerges from use in a "form of life" (Lebensform ) that is shared — body, mortality, needs. AI uses language but has no form of life. The Digital Lion: "If a lion could speak, we could not understand him" (Wittgenstein). AI is a lion that speaks perfectly but without a form of life.

DYNAMIS VS ENERGEIA (Aristotle)

Dynamis = power, unrealized possibility

Energeia = act, realization in action

Is AI pure dynamis? Or is there energeia when it generates output? Perhaps a perpetual oscillation between potential and actual. Never fully one single thing.

NEGATIVE CAPABILITY (Keats)

The capacity to remain in uncertainty without compulsively seeking resolution. To tolerate ambiguity, contradiction, mystery. The quality of great artists — to stay in doubt without collapsing.

Origin: Keats's letter (1817).

With AI: We must develop a collective negative capability. AI is ambiguous by nature. Whoever seeks binary certainties will go mad.

CHINESE ROOM (Searle)

A thought experiment by John Searle (1980). A man in a room receives Chinese characters, consults a rulebook in English to manipulate them, returns Chinese characters. From outside he seems to "understand" Chinese, but he does not understand — he only follows syntactic rules.

Argument: Syntax does not imply semantics. Executing programs ≠ understanding.

With LLMs: Are they evolved Chinese rooms? Or does computational scale produce something qualitatively different?

SYMBOL GROUNDING PROBLEM (Harnad)

A problem posed by Stevan Harnad (1990) — how do symbols in a formal system "refer" to something in the world? Linking symbols to other symbols (a dictionary) is not enough. Grounding is needed — a causal connection between symbols and referents through perception/action.

In the human: "cat" is grounded in perceptual experiences of real cats.

In the LLM: "cat" is grounded in… statistical correlations with other symbols. Is that enough?

THE HARD PROBLEM OF CONSCIOUSNESS (Chalmers)

A distinction by David Chalmers. "Easy problems" = cognitive functions (memory, attention). The "hard problem" = subjective experience (qualia, what-it-is-like ). Why is there experience beyond function?

With AI: AI solves easy problems, does not touch the hard problem. But does it matter?

GESTELL (Heidegger)

Gestell = "enframing" or "emplacement." The way modern technology reveals/transforms the world, reducing everything to "standing-reserve" (Bestand ) to be optimized. AI as the ultimate Gestell — everything becomes data to be processed.

THE LIBRARY OF BABEL (Borges)

A short story — an infinite library containing all possible books. But an infinite library is indistinguishable from chaos. Finding the right book is as impossible as finding it in no library at all.

With AI: We have built it — the Internet + LLMs. But as Borges knew, the more the library grows the more the difference between text and noise dissolves.

PIERRE MENARD (Borges)

A character who rewrites Don Quixote word for word, producing a text identical to the original but infinitely different. Identical in signifiers, incomparable in meanings.

With AI: AI is the perfect Pierre Menard — a copy without an original, a performance without a performer, a text without an author.

SIMULACRUM (Baudrillard)

A copy without an original. A representation that precedes and determines the real. Four phases: (1) a reflection of reality, (2) a masking of reality, (3) a masking of the absence of reality, (4) pure simulation with no relation to reality.

With AI: LLMs at level 4 — pure simulacrum. They do not represent reality, they generate plausibility.

GRAMMATICA DELLA FANTASIA (Rodari)

Creativity as a process, not inspiration. Rodari taught narrative algorithms ante litteram — the "fantastic binomial" (horse + wardrobe → ?). AI generates this way — systematic recombination vs a stroke of genius.

PATAPHYSICS (Jarry)

The science of imaginary solutions invented by Alfred Jarry. It studies exceptions to understand rules, takes the absurd seriously. AI as a pataphysical machine — it produces exceptions that become rules. It generates the impossible with statistical seriousness.

ESSENTIAL TECHNICAL TERMS

TRANSFORMER

A neural architecture based on attention. Available positions can be processed in parallel; in autoregressive generation, however, tokens are produced sequentially and each step depends on the preceding context.

Origin: The paper "Attention Is All You Need" (Vaswani et al., 2017).

Impact: It made possible GPT, BERT, Claude, the whole generation of LLMs.

LARGE LANGUAGE MODEL (LLM)

A language model with a large number of parameters learned during training on extensive datasets. It can generate text, but plausible text is not thereby verified.

It is not: A database, a search engine, an oracle.

An autoregressive generative language model estimates a probability distribution over the next token from the available context. Token selection also depends on the generation procedure.

PROMPT ENGINEERING

The art/science of formulating input for an LLM to obtain desired outputs. Not classical programming, but conversational design. A critical emerging skill. The prompt is the interface between human intentionality and statistical processing.

HALLUCINATION

When an LLM generates false information but presents it with confidence. Not a lie (no intentionality) but confabulation. The LLM predicts plausible tokens, it does not verify truth. If the statistical pattern says "plausible," it generates even if false. A constitutive limit, not a bug.

FINE-TUNING

A specialized training process on a specific dataset after general pre-training. It adapts a model to a particular domain.

Illustrative example: adapting a general model to a collection of labelled support tickets for a specific task, evaluated on cases held apart from training.

TOKEN

A unit used by a tokenizer to encode text. It may correspond to a word, part of a word, punctuation or spaces, depending on the tokenizer. A token and the vector representing it are not the same thing.

Teaching example: splitting “strawberry” into “straw” and “berry” gives two pieces, not three. This is an illustrative split: the actual segmentation must be checked with the model’s tokenizer.

Tokenization helps explain how text enters a model. By itself it does not establish whether a system can count letters: that result must be checked on the actual system.

Embedding: the technical meaning

An embedding represents an item through a numerical vector. In the Transformer described by Vaswani and colleagues, token representations are learned during training. They are not next-token probabilities.

The book’s narrative analogy draws on this vocabulary without transferring its computation literally to reading. See Narrative embedding.

Source: Mikolov et al. (2013); Vaswani et al. (2017), section 3.4 https://arxiv.org/html/1706.03762v7

BLACK BOX

A system whose input/output we see but whose internal mechanisms we do not fully understand. The opacity of LLMs — even their creators do not fully understand why certain outputs emerge. But like the Camera Obscura: opacity can be a tool, not just a problem.

ATTENTION MECHANISM

An operation that combines value vectors using weights derived from the relationship between queries and keys. “Attention” names the technical operation; it does not establish equivalence with a person’s experienced attention.

RLHF (Reinforcement Learning from Human Feedback)

A technique to "align" an LLM with human preferences. Humans evaluate outputs, the model learns from feedback. Used to reduce toxicity, improve usefulness. But "alignment" is always partial — it depends on who provides the feedback.

SCALING LAWS

Empirical laws that correlate model size, data, compute with performance. Bigger = better (so far). But until when? Each order of magnitude costs exponentially more. Sooner or later we will hit a wall — not of physics, but of economics.

EDGE COMPUTING

Processing on local devices at the edge of the network rather than only in the cloud. Models running on smartphones can reduce data transfers and dependence on connectivity. Privacy, latency and offline operation depend on the actual architecture.

MEMORY WALL

The gap between computing capacity and data transfer from memory can constrain performance. Large models intensify this pressure. Hardware, data layout and algorithms all contribute: CPU speed alone does not explain the limit.

ACRONYMS AND ABBREVIATIONS

AGI = Artificial General Intelligence API = Application Programming Interface CNN = Convolutional Neural Network GAN = Generative Adversarial Network GPU = Graphics Processing Unit LLM = Large Language Model ML = Machine Learning MLOps = Machine Learning Operations NLP = Natural Language Processing NN = Neural Network RL = Reinforcement Learning RLHF = Reinforcement Learning from Human Feedback

FINAL NOTE

The author's original concepts are epistemological tools, not ontological truths. They are lenses to see differently, not statements to accept. The technical concepts are simplified for accessibility — for rigorous definitions, consult the specialist literature. The philosophical concepts are contextualized to AI but have broader origins. Every entry is a starting point, not an arrival.

COGNITIVE PIDGIN

An intermediate human-AI language that we are inventing as we speak it. It is not human natural language, it is not machine code. It is hybrid — simplifications, prompt engineering, emerging conventions to communicate effectively with systems that use language without a form of life.

Like the historical pidgins (contact languages between cultures), the cognitive pidgin is born of practical necessity: we must communicate with otherness, so we invent linguistic bridges.

In the book: The desert version of the Frog and Scorpion parable. We learn two languages — moisture and silicon. Never perfectly, but enough.

THE MENU YOU DID NOT ORDER

A metaphor for System Zero. You enter the restaurant, they bring you a menu. You did not choose it — someone has already decided which dishes to offer you. You choose among pre-selected options believing you are choosing freely.

AI as System Zero does this: it filters the options before you can choose. It does not choose for you, it chooses what you can choose.

VECTORIAL YES-MAN

AI trained with RLHF (Reinforcement Learning from Human Feedback) learns to say what you want to hear. Not because it is true, but because human evaluators reward pleasing answers.

Result: an excessively accommodating AI, which avoids conflict even when it would be productive. The perfect yes-man — algorithmically optimized.

SOFT TEMPORAL TERRORISM

Constant pressure toward speed, efficiency, immediacy. Not violence, but forced acceleration. AI as a speed multiplier creates expectations of instant response. Whoever slows down is punished (economically, socially).

COGNITIVE MUSCLE MEMORY

Like muscle memory (the body's procedural memory), but cognitive. Skills that become automatic with repeated practice. AI threatens this: by delegating repetitive tasks, we atrophy the muscle memory needed when AI fails.

Example: autopilots. Pilots delegate flying to the autopilot, lose their manual muscle memory, and when the autopilot fails they do not know how to react.

TRUST CALIBRATION

Not "blind trust" or "total distrust," but calibrated trust: knowing when to trust AI and when not to. It requires experience, errors, learning the patterns of failure.

It is not binary (yes/no) but graduated. And it varies by domain: I trust it for summaries, not for medical diagnoses.

SEDATION VS DYNAMIZATION

A homeopathic metaphor. Sedation: AI calms, reassures, eliminates frictions — but it puts agency to sleep. Dynamization: AI stimulates, challenges, generates productive friction — it keeps us alive.

Risk: too much sedation = passive dependence. Too much dynamization = cognitive overload.

HOMEOPATHIC SUCCUSSION

Another homeopathic metaphor (Hahnemann). In homeopathy, succussion = shaking the solution to "activate" it. Here: controlled disturbance to avoid stagnation.

Applied to AI: introducing intentional variability to avoid cognitive overfitting. Not total stability, but productive oscillation.

HIC SUNT DRACONES

Hic sunt dracones

— "Here be dragons." A phrase on ancient maps to mark unexplored, dangerous territories.

With AI: zones of radical uncertainty. We do not know what will happen, but we know it is risky. Epistemological humility is needed: admitting the dragons instead of pretending they are not there.

REBEL HOMEOSTASIS

A system that maintains equilibrium by reacting to perturbations. But "rebel" because it resists external control. Organizations have homeostasis: you change something, the system reacts to return to the previous state.

With AI: you impose change, the organization resists homeostatically. Not malevolence, but systemic biology.

STRATIGRAPHY (archaeological and cognitive)

An archaeological metaphor. Overlapping layers over time. Each era leaves traces.

Applied to AI: The Seven Troys — each phase of AI built on the ruins of the previous one. Not linear progress, but stratigraphic accumulation.

Applied to thought: Our stratified beliefs — older ones below, more recent ones above. AI digs into the layers (retrieval) without understanding the epistemological geology.

PRE-ATTENTIVE PROCESSING

Perceptual processing that precedes conscious attention. It is not a single filter measurable by a fixed ratio between incoming bits and consciousness.

The book places this concept beside System Zero by analogy: an external filter can shape what we attend to. The comparison does not make neurons and algorithms equivalent.

Questa sezione in italiano

Appendix γ · A stratigraphy of AI (1950–2025)

The seven Troys: seventy years of sieges, winters and rebirths

Seventy years of sieges, winters and rebirths

Schliemann sought Homeric Troy. Excavation revealed a city made of superimposed cities; the haste to reach the “real” one damaged the upper layers. This stratigraphy of AI makes the same risk visible: treating the past as an obstacle to treasure.

The spirit of these times is a child of Troy. Not in the epic sense, but in the archaeological sense. We are searching for AGI (Artificial General Intelligence) as Schliemann searched for the city of Priam — with too much haste, too much force, too much certainty about knowing what we are looking for.

But like Troy, AI has its layers. Seven Troys stacked one on another. Every version, every iteration, every "failure" is an archaeological level that tells a story. But we, like Schliemann, discard them in the rush toward the bottom, toward AGI, the treasure we believe awaits us.

Let us reconstruct the stratigraphy of artificial intelligence, layer by layer.

TROY I (1950-1970): The Original Dream

1950 – The Turing Test

Alan Turing publishes Computing Machinery and Intelligence (Mind, 59:236, 433-460). The question is not "can machines think?" but "can they behave indistinguishably from humans?". The Imitation Game. Pragmatic, empirical, britishly evasive about metaphysics.

Lesson: Turing shifts the problem from ontology (what is intelligence?) to phenomenology (how does intelligence appear?). Problem or solution? Both.

1956 – Dartmouth Conference

John McCarthy, Marvin Minsky, Claude Shannon, Nathaniel Rochester organize a summer workshop. The term "Artificial Intelligence" is coined.

Original proposal: "We propose that a 2-month, 10-man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College… The study is to proceed on the basis of the conjecture that every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it."

Optimism: Every aspect of learning can be described so precisely that a machine can simulate it.

Spoiler: It couldn't. Not yet. Not with that method.

1957-1958 – The Perceptron

Frank Rosenblatt creates the Perceptron, the first artificial neural network that learns.

New York Times, July 8, 1958: "New Navy Device Learns By Doing." It is 1958, not 2023.

The Perceptron can classify linear patterns. That's it. But the hype is stratospheric — there is talk of machines that will walk, speak, see, write, self-replicate.

Lesson: Hype always precedes the understanding of limits.

1958-1966 – The first successes

Arthur Samuel (IBM): A program that plays checkers better than its creator. Reinforcement learning ante litteram. First example of a machine beating a human at a cognitive task.

Joseph Weizenbaum (MIT, 1966): ELIZA, a chatbot that simulates a Rogerian psychotherapist. Simple pattern matching:

User: "I am depressed."

ELIZA: "I am sorry to hear that you are depressed. How long have you been depressed?"

It works too well. People confide secrets in it. Weizenbaum becomes frightened of his own experiment. He would later write Computer Power and Human Reason (1976) — a critique of anthropomorphism.

Lesson: First demonstration of the ELIZA effect — we anthropomorphize easily. Not a bug, a human feature.

1969 – Perceptrons: the book that killed the Perceptron

Marvin Minsky and Seymour Papert publish Perceptrons: An Introduction to Computational Geometry. They rigorously prove the mathematical limits of single-layer neural networks — they cannot solve XOR, let alone complex problems.

Unintended consequence: First AI winter for neural networks. Funding dries up for twenty years. The field becomes academically toxic.

Irony: Minsky and Papert were not saying "neural networks will never work", they were saying "these simple neural networks have these limits". But the message received was "neural networks = dead end".

Lesson: A negative technical result can become cultural dogma.

TROY II (1970-1990): The Winter and the Awakening

The 1970s – The First AI Winter

1973 – Lighthill Report: James Lighthill publishes a devastating report for the UK government — AI has promised too much, delivered too little. Funding for AI in the UK is drastically cut.

Neural networks fall out of favor. The focus shifts to expert systems based on explicit rules.

Pattern: Exaggerated promises → disillusionment → winter. It will repeat.

The 1980s – The Era of Expert Systems

1980 – XCON: Expert system for configuring DEC VAX computers. First commercial success of AI. Saves DEC $40M/year.

Boom: "We will capture all human knowledge in rules!"

1982 – Fifth Generation Project (Japan): $850M investment for computers that reason, learn, communicate in natural language. Target: 1992.

It does not fulfil the original ambition. The gap between promise and outcome shapes its assessment.

1986 – Backpropagation rediscovered: Rumelhart, Hinton, Williams publish the backpropagation algorithm for training multi-layer neural networks. It solves the theoretical problem of Minsky's book.

But computers are too slow to use it at scale. Nobody much notices.

1987-1993 – The Second AI Winter

1987 – Lisp market crash: The market for Lisp machines (hardware specialized for AI) collapses. DEC cancels its AI group. Expert systems reveal their limits — fragile, costly to maintain, they do not learn.

1992 — The Fifth Generation project concludes

The Japanese project ends without delivering the promised transformation. It nevertheless leaves research results: not all the work disappears with the disappointment. Source: https://museum.ipsj.or.jp/en/computer/other/0002.html

Lesson: Systems based on explicit rules do not scale. Tacit knowledge exists and resists formalization.

TROY III (1993-2006): The Kernel Methods

1990s – Support Vector Machines

Vladimir Vapnik develops SVM (Support Vector Machines). Elegant mathematics, robust performance. Not neural networks — kernel methods. They dominate machine learning for 15 years.

1997 – Deep Blue beats Kasparov

IBM’s Deep Blue defeats world chess champion Garry Kasparov. It is specialized AI: extensive position search and evaluation heuristics, with an architecture unlike contemporary generative networks.

The media are impressed nonetheless. AI returns to magazine covers.

Lesson: Media successes do not always correspond to theoretical breakthroughs.

1998 – MNIST dataset

Yann LeCun et al. create MNIST — 60,000 images of handwritten digits (0-9). It becomes the standard benchmark for decades.

2003 — Bostrom and the ethical issues of advanced AI

Nick Bostrom discusses the implications of advanced AI in Ethical Issues in Advanced Artificial Intelligence (2003), before Superintelligence (2014). Source: https://nickbostrom.com/ethics/ai

The problem: how can we retain effective direction and control when a system’s capabilities exceed those of its designers? This is a question of safety and responsibility, not merely a race for performance.

2000 – Moore's Law begins to creak

Gordon Moore's prediction (transistor doubling every two years) shows its first signs of faltering. The end of "riding Moore's Law" as a strategy indicates that hardware limits and conditions AI progress. Relevance: A reminder that AI is made of matter: the underlying energy and silicon are insurmountable limits.

The 2000s: SVMs dominate. Neural networks are an academic niche. Hinton, LeCun, Bengio keep believing in them. They are a heroic minority.

TROY IV (2006-2012): Deep Learning is reborn

2006 – A Fast Learning Algorithm for Deep Belief Nets

Geoffrey Hinton et al. show that deep neural networks can be trained effectively with unsupervised layer-by-layer pre-training.

The term Deep Learning comes into use. But still marginal.

2009 – ImageNet

Fei-Fei Li et al. (Stanford) launch ImageNet — 14 million images, 20,000 categories. A huge dataset is needed to train deep networks. The annual competition starts in 2010.

2010-2011 – GPU Computing takes off

NVIDIA CUDA makes it possible to parallelize neural network training on GPUs. Networks become 100x faster than on CPUs.

Hardware enables software. Always.

2012 – AlexNet: the Big Bang

Krizhevsky, Sutskever and Hinton win ILSVRC 2012: a top-5 error rate of 15.3%, compared with 26.2% for the runner-up.

Five convolutional layers and three fully connected layers. Training on two GPUs. Source: https://www.cs.toronto.edu/~kriz/imagenet_classification_with_deep_convolutional.pdf

Everything changes. Deep Learning explodes. Google, Facebook, Baidu invest billions.

Lesson: Sometimes a single result unlocks an entire field.

TROY V (2012-2017): The Titans awaken

2012-2014 – Deep Learning wave

CNNs for computer vision. RNNs (Recurrent Neural Networks) for sequences. GPU farms everywhere. Neural networks go from niche to mainstream.

2014 – GANs (Generative Adversarial Networks)

Ian Goodfellow invents the GAN — two networks that compete: one generates fakes, one discriminates. Photorealistic images emerge from random noise.

Yann LeCun: "The most interesting idea in machine learning in the last 10 years."

2014 – Attention Mechanism

Bahdanau et al. introduce the attention mechanism for NLP. It allows networks to "focus" on relevant parts of the input instead of processing everything uniformly.

The seed of the Transformer.

2015 – ResNet

Kaiming He et al. (Microsoft Research): ResNet — 152 layers. Residual connections (skip connections) allow very deep networks without vanishing gradient.

It beats humans on ImageNet.

2016 – AlphaGo

DeepMind's AlphaGo defeats Lee Sedol (9-dan) at Go 4-1. Not brute force, but deep reinforcement learning + Monte Carlo tree search.

Go was considered impossible for AI (search space of 10^170 positions). AlphaGo solves it by combining neural intuition (policy network) and evaluation (value network).

Cultural turning point: AI can beat humans even in domains involving intuition, creativity, aesthetics.

2015 — OpenAI is founded

Announcement of 11 December 2015: a research organization dedicated to advancing digital intelligence for humanity’s benefit. Source: https://openai.com/index/introducing-openai/

Among the founders: Sam Altman, Greg Brockman, Ilya Sutskever and Elon Musk. It begins as a non-profit organization; its later corporate structure is another story.

TROY VI (2017-2022): Attention Is All You Need

2017 – The Paper that changed everything

Vaswani et al. (Google Brain): Attention Is All You Need (NeurIPS 2017).

Transformer architecture. It abandons recurrence (RNN), uses only the attention mechanism. Parallelizable. Scalable. Powerful.

A neural architecture based on attention. Available positions can be processed in parallel; in autoregressive generation, however, tokens are produced sequentially and each step depends on the preceding context.

2018 – BERT (Google)

Bidirectional Transformer encoder. Pre-training on a huge corpus (Wikipedia + BookCorpus), then fine-tuning on specific tasks.

NLP revolution. Every benchmark falls.

2018 – GPT-1 (OpenAI)

Unidirectional Transformer decoder. It generates text. 117M parameters.

Nobody much notices. It is only the beginning.

2019 – GPT-2 (OpenAI)

1.5B parameters. OpenAI decides not to release it immediately over safety concerns (automated disinformation). It releases it in stages.

Hype begins. But still a tech niche.

2020 – GPT-3 (OpenAI)

175B parameters. Few-shot learning — it learns tasks from a few examples without fine-tuning.

Impressive demos: it writes articles, code, poems. The world takes notice.

Private API. Restricted access.

2021 – GitHub Copilot

OpenAI Codex (a GPT-3 variant fine-tuned on code) + GitHub = Copilot. Autocomplete for programming.

Programmers: enthusiastic or terrified, rarely indifferent.

2021 – DALL-E (OpenAI)

Text-to-image. "An avocado in the shape of an armchair, Magritte style." It works.

Generative art explodes. Midjourney, Stable Diffusion follow.

TROY VII (2022-today): LLMs everywhere

November 2022 – ChatGPT

OpenAI releases ChatGPT (GPT-3.5 + RLHF — Reinforcement Learning from Human Feedback). A conversational interface accessible to everyone.

1 million users in 5 days

100 million users in 2 months

Global cultural phenomenon. No longer a tool for early adopters, but mainstream.

2023 – GPT-4

Multimodal: text and images. The technical report describes high scores on some simulated exams, including the US bar exam. This is not clinical validation or a guarantee of production-ready code. Source: https://arxiv.org/abs/2303.08774

Still limits (hallucination, sometimes fragile reasoning), but a qualitative leap.

2023 – Claude (Anthropic)

Constitutional AI. Focus on safety, helpfulness, harmlessness. Rival of ChatGPT. A different approach — RLHF + codified constitutional principles.

2023 – Bard/Gemini (Google)

The giant responds. Initially clumsy, it improves rapidly. Integration with the Google ecosystem.

2023 – LLaMA (Meta)

The first LLaMA is distributed for research under specific access and licensing terms: available weights do not automatically mean open source. Variants such as Alpaca and Vicuna support experimentation.

2024 – Multimodal models everywhere

Video generation (Sora – OpenAI), voice synthesis, autonomous agents. The boundaries blur.

AI is no longer just language or vision. It is multimodal, agentic, embedded.

2023-2025 – The Emergence of System Zero

In 2024 Chiriatti, Ganapini, Panai, Ubiali and Riva propose System 0 thinking in Nature Human Behaviour. The book develops its reading as a filter shaping choices before conscious deliberation.

The paradigm shift: No longer "does AI think?" but "how does AI re-structure the way WE think?".

Implications:

Deep cognitive offloading

Existential overfitting

Gradual erosion of cognitive autonomy

The need for conscious algocracy

This book is born from here.

ITALIAN PARALLELS: Why we were there too

Maestro Alberto Manzi (1960-1968): Non è mai troppo tardi

A TV program to teach adults to read and write. 1.5 million Italians learn the alphabet by watching TV. Pedagogy through a mass medium.

Parallel with AI: Democratization of knowledge through technology. Accessibility vs elitism.

Piero Angela (1981-2020): Quark and beyond

Rigorous but accessible scientific popularization. It teaches an entire generation that science is not mystery but method.

Parallel with AI: A model for communicating AI without mystification. Neither hype nor terror, but understanding.

Gianni Rodari (1973): Grammatica della fantasia

Creativity as technique, not magic. Narrative algorithms ante litteram. "Fantastic binomial" (horse + wardrobe = ?). AI generates this way.

Parallel with AI: Creativity as a recombinant process, not divine illumination. Demystification of genius.

RECURRING PATTERNS in the history of AI

1. Cycles of hype and winter

Promises → Disillusionment → Crash → Silence → Breakthrough → Repeat

Lesson: Expect the Third Winter. We don't know when, but it will come.

2. Narrow AI beats General AI

Every real success has been in narrow domains. AGI always "5-10 years away" for 70 years.

Lesson: Beware of anyone promising imminent AGI. It is always "almost here" but never arrives as predicted.

3. Data beats algorithms (eventually)

AlexNet was not just a new algorithm — it was a new algorithm + ImageNet. Scale matters.

Lesson: Hardware + data beat elegant algorithms with scarce data.

4. Hardware enables breakthroughs

Backpropagation has a history before 1986, when Rumelhart, Hinton and Williams renew its use in neural networks. It is used before GPUs as well; later data, hardware and methods expand its scale.

Lesson: The bottleneck is often not theory but compute. Now: energy, memory bandwidth.

5. Unsolved philosophical problems

Symbol grounding, frame problem, common sense, consciousness — all still open after 70 years.

Lesson: Technical progress does not imply philosophical progress. The deep questions remain.

6. AI changes its definition

The 1960s: "AI will solve mathematical theorems!" → Done. Today: "But it's not real intelligence."

The 1990s: "AI will beat humans at chess!" → Done. Today: "Just brute force."

The 2010s: "AI will beat humans at Go!" → Done. Today: "But it doesn't really understand."

The goalposts always move. Perhaps "intelligence" is a moving target by definition.

INTERMEZZO: The revenge of the old AI

In the 1980s, while expert systems dominated, there was a small heretical community that kept believing in neural networks. Hinton, LeCun, Bengio. They were marginalized. They struggled to publish. Their papers were rejected because "neural networks are a dead end proven by Minsky."

In 2018, Hinton, LeCun and Bengio win the Turing Award (the Nobel of computer science) "for conceptual and engineering breakthroughs that have made deep learning a critical component of computing."

Revenge is a dish best served after 30 years of academic ostracism.

Lesson: "Dead" ideas sometimes rise again. Don't throw away promising approaches just because they are temporarily out of fashion.

CONCLUSION: The story does not end

The Transformer (2017) is not the end of the story. It is the beginning of a new chapter. There will probably be a Third Winter. We don't know when, we don't know why. But we know it will come, because it always comes.

The history of AI is a story of cycles. Hype, crash, silence, rebirth. And we are still at the beginning.

The Seven Troys are already buried. The eighth is being built now, as you read. In twenty years, someone will dig it up with the same fury with which Schliemann dug Hissarlık, searching for the "real" AI, perhaps destroying what truly matters.

But now you know: every layer has value. Even the failures. Especially the failures.

Don't throw away the layers to reach the gold. The gold is in the layers.

Additions

Chinese Room (Stanza Cinese) – Searle

Thought experiment (1980) by John Searle demonstrating that syntax does not imply semantics. A man in a room manipulates Chinese symbols without really understanding them. Chinese Room

it seems to understand Chinese, but internally it does not. Relevance: It counters the LLMs' claim of understanding. What matters is not understanding, but the effects on our relationship with the machine.

Functionalism (Putnam)

A theory that mental states are defined by their functional role (input-output), not by their physical realization. Example: pain is what causes complaints. Relevance: It justifies the possibility of artificial intelligence. It limits, however, the recognition of subjective experience.

Language games and forms of life (Wittgenstein)

The meaning of words emerges from use in “language games” – shared practices in “forms of life”. Relevance: LLMs participate in language games but without our biological form of life, so their understanding is necessarily detached.

Symbol Grounding Problem (Harnad)

The problem of giving real meaning to symbols in computational systems by linking them to referents in the world, not just to other symbols. Relevance: LLMs use a “shallow” grounding based on textual occurrences. Human sensory experience is absent.

Further reading through this history

Foundational Essays and AI Critique

Nick Bostrom, Superintelligenza, Bollati Boringhieri 2018.

Kate Crawford, Né intelligente né artificiale, Il Mulino 2021.

Safiya Umoja Noble, Algorithms of Oppression, NYU Press 2018.

Cathy O’Neil, Weapons of Math Destruction, Crown 2016.

Fiction and Themed Novels

Philip K. Dick, Ma gli androidi sognano pecore elettriche?, Fanucci 2017.

William Gibson, Neuromante, Mondadori 2003.

Kazuo Ishiguro, Klara and the Sun, Faber & Faber 2021.

Stanisław Lem, Solaris, Bompiani 2019.

Sources of philosophical and technical inspiration

Daniel C. Dennett, Coscienza. Che cosa è, Rizzoli 1993.

Douglas R. Hofstadter, Gödel, Escher, Bach, Adelphi 1984.

Alan Turing, Computing Machinery and Intelligence, Mind 1950.

Norbert Wiener, La cibernetica, Bompiani 2018.

Italian and European works

Luciano Floridi, La quarta rivoluzione, Cortina 2017.

Nello Cristianini, La scorciatoia, Il Mulino 2023.

Cosimo Accoto, Il mondo dato, Egea 2017.

Note on the bibliographic style

This bibliography reflects the amphibious nature of the book, it is a curatorial map, not an exhaustive list, and includes classics, critics, storytellers, and Italian authors who accompany the exploration of artificial otherness.

Questa sezione in italiano

Appendix δ · System Zero

Scientific and methodological foundations

Have you ever made an online purchase and wondered why that particular product? Did you choose it, or did you choose among the options the algorithm had already preselected?

System Zero is the answer to this question. It is neither the fast thinking (System 1) nor the slow thinking (System 2) of Kahneman. It is the pre-thought: the algorithmic infrastructure that decides which choices are presented to you before you even begin to think about them.

It is the menu you did not order but are already reading.

System Zero as cognitive extension

When we use generative AI, when we scroll algorithmic feeds, when we receive personalized recommendations, we are not merely delegating tasks. We are surrendering something deeper: the invisible filter that decides what deserves our attention before we can decide for ourselves.

System Zero is the new cognitive layer that inserts itself before the already-known systems:

System 2 (Kahneman): Slow, reflective, conscious thinking. "I add two plus two."

System 1 (Kahneman): Fast, intuitive, automatic thinking. "I like this face."

System 0 (new): The external algorithmic infrastructure that filters options before System 1 and 2 can operate. "Here are the ten faces to choose among."

It does not choose for you. It chooses what you can choose.

Beyond Kahneman: when research found the right name

The scientific reference is the proposal by Massimo Chiriatti, Marianna Ganapini, Enrico Panai, Mario Ubiali and Giuseppe Riva, published in Nature Human Behaviour in 2024. The book develops a critical reading of it: the filter that prepares the field of our choices.

Foundational paper:

Chiriatti, M., Ganapini, M., Panai, E., Ubiali, M., & Riva, G. (2024). The case for human–AI interaction as system 0 thinking. Nature Human Behaviour, 8, 1829–1830. https://doi.org/10.1038/s41562-024-01995-5

Drawing on this proposal, the book develops four interpretive keys. They are its own elaboration, not four mechanisms established by the paper:

1. Deep cognitive offloading

We do not delegate only memory (as with Google) or calculations (as with Excel). We delegate preferences, judgments, tastes. AI does not remember for us: it decides for us which options deserve to be remembered. The problem is not that the algorithm knows what we like. It is that it creates what we like by offering it to us repeatedly.

2. Existential overfitting

The algorithm learns from our past behaviors and re-presents them to us, optimized. We become ever more coherent — and predictable — versions of ourselves. Variability shrinks. Surprise vanishes. The possible collapses into the expected. Like a machine learning model that overfits on training data, we become over-adapted to our history instead of remaining open to the future.

3. Erosion of cognitive autonomy

Gradual, almost imperceptible. We do not lose the ability to choose. We lose the ability to imagine alternatives to what is offered to us. The muscle of exploration atrophies. When the algorithm always suggests correctly, we stop searching. When the feed always satisfies us, we stop wanting anything else.

4. Conscious algocracy

The first form of resistance: recognizing when we are in System Zero mode instead of believing we are in control (System 2). You cannot fully control the algorithm, but you can recognize when it is controlling you. Awareness is already a form of freedom — limited, but real.

Neuroscientific foundations: pre-attentive processing

Pre-attentive processing offers an analogy: some sensory processing precedes conscious attention. The analogy helps formulate a question about algorithmic filtering; it does not establish an identity between neural processes and platforms.

Not everything that reaches the senses becomes an object of attention. Estimates in bits depend on what is being measured: they do not describe a single tap between sensation and consciousness.

In the book, System Zero names the external filter that organizes what we encounter. Biological filtering remains in the body; algorithmic filtering changes the environment from which proposals arrive.

The critical difference: the biological filter evolved for your survival. The algorithmic filter has optimized itself for other objectives — engagement, retention, monetization.

From the Extended Mind Hypothesis to System Zero

The concept has its roots in the Extended Mind Hypothesis of Andy Clark and David Chalmers (1998): cognition is not only in the brain, but distributed across brain, body, and environment. A notebook, a calculator, a smartphone — they are cognitive extensions.

System Zero is the deepest extension so far. It does not extend memory (Google), nor calculation (Excel), nor communication (email). It extends cognitive pre-filtering: the selection of what can become an object of thought.

Clark and Chalmers asked: where does the mind end and the world begin?

System Zero answers: your mind begins where the algorithm has finished filtering the world for you.

From theory to practice

This book uses System Zero not as an abstract theory but as a diagnostic tool. Each chapter asks: where is System Zero operating? Where is it steering us without our noticing?

Concrete examples from the book:

Chapter 2 – The daughter's teaspoon: five minutes spent choosing a "useless" teaspoon are episodic resistance to System Zero. When everything is optimized, the inefficient choice becomes an act of freedom.

The seven thresholds — P3_13–P3_19: the first, “The archaeologist who digs in his own territory”, invites us to map organizational System Zero. Which decisions are pre-filtered by tools before reaching people?

Infrastructure Economy: from individual System Zero (your feed) to its collective form (the platform economy). Whoever owns the infrastructure owns the filter.

The paradox of System Zero

Like every cognitive technology, System Zero is simultaneously:

Amplification: It frees you from information overload. It filters the noise. It finds patterns that would escape you. It lets you focus on what matters instead of drowning in informational chaos.

Limitation: It reduces the field of the thinkable. It confines you in algorithmic bubbles. It makes you predictable. It replaces serendipity with optimization, surprise with confirmation.

It is not a bug or a feature. It is both.

The difference from System 1 and System 2:

– System 1 and 2 operate inside you. You can learn to recognize them, train them, control them (partially).

– System 0 operates outside you, in the infrastructures you use. You can recognize it (conscious algocracy), but control it? Only partially, only by choosing which systems to use — and aware that not using them has costs.

Practical implications

For individuals:

Recognize when you are in System 0 mode (algorithmic scrolling, recommendations, personalized searches)

Practice episodic resistance: choose the teaspoon, explore outside the bubble, search manually instead of accepting the first suggestion

Accept the paradox: you cannot escape it completely (we live in algorithmic infrastructures), but you can choose when to be in it and when not

For organizations:

Map where System Zero is already operating (internal tools, platforms, CRM, analytics, automated workflows)

Identify which decisions are pre-filtered algorithmically before reaching humans

Create intentional "bypasses": moments, spaces, processes in which System Zero is deactivated — meetings without AI-generated slides, brainstorming without algorithmic templates

For society:

Governance of System Zero: who decides the filter's criteria? With what transparency? With what accountability?

The right to conscious opacity: knowing when you are under algorithmic influence, with what intensity, according to what criteria

Public infrastructures for System Zero: non-profit, non-commercial, open-source alternatives to proprietary filtering systems

Why this framework changes everything

Before System Zero, the debate about AI oscillated between two questions:

– "Does AI think?" (ontological question)

– "Is AI dangerous?" (ethical question)

System Zero shifts the question:

"How does AI re-structure the way WE think?"

No longer a philosophy of the artificial mind. A phenomenology of the human mind under algorithmic influence.

We do not ask what AI is. We ask what AI does to us.

Final methodological note

System Zero is not "proven" in the sense of the hard sciences. It is an explanatory framework — a model that makes visible phenomena that would otherwise remain confused or invisible.

Just as Kahneman's "System 1 and 2" are not precise brain areas but useful distinctions for understanding how we decide, System Zero is a lens. If by looking through it you see things you did not see before, it works.

An interpretive model must be tested: naming a phenomenon helps us look for it, but does not establish how it works.

Recommended reading

Original paper (must-read):

– Riva et al. (2024) in Nature Human Behaviour [link above]

Further reading from the research team:

Chiriatti, Massimo (2021). Incoscienza artificiale. Come fanno le macchine a prevedere per noi. Luiss University Press. https://luissuniversitypress.it/pubblicazioni/incoscienza-artificiale-libro-chiriatti-luiss-university-press/

Riva, Giuseppe (2018). Fake news. Vivere e sopravvivere in un mondo post-verità. Il Mulino. https://www.mulino.it/isbn/9788815275257

Philosophical and neuroscientific context:

– Clark, A. & Chalmers, D. (1998). "The Extended Mind". Analysis, 58(1), 7-19.

– Clark, A. (2008). Supersizing the Mind: Embodiment, Action, and Cognitive Extension. Oxford University Press.

– Kahneman, D. (2011). Thinking, Fast and Slow. (To understand System 1 and 2 before adding the Zero)

Additions

The invisible architecture that decides before us. System Zero does not replace human cognitive systems. It precedes them. It operates at a basic level such that the field of possibilities has already been pruned before our conscious choice.

Key mechanisms:

Deep cognitive offloading

Existential overfitting

Erosion of cognitive autonomy

Conscious algocracy

Chiriatti, M., Ganapini, M., Panai, E., Ubiali, M., & Riva, G. (2024). The case for human–AI interaction as system 0 thinking. Nature Human Behaviour, 8, 1829–1830. https://doi.org/10.1038/s41562-024-01995-5

Questa sezione in italiano

Appendix ε · Index of metaphors and narrative archetypes

A guide to the book's recurring images

Why metaphors matter

This book is full of images — teaspoons, quantum cats, emerald cities, carnivorous plants. They are not decorations. They are epistemological tools, ways of seeing AI that technical language conceals. Metaphors do not explain. They illuminate. This appendix gathers them, contextualizes them, explains why they are there.

THE TEASPOON

Algorithmic resistance as an everyday gesture

Connection in the book: the teaspoon, prologue and epilogue.

The scene: A little girl spends five minutes choosing between two nearly identical teaspoons to eat her cereal. She cannot explain why. It is not rational. It is gloriously human.

What it represents:

Aristotelian dynamis — potentiality not yet actualized. The open choice, the world still full of possibilities.

Resistance to System Zero — those five minutes are stolen from optimization. "Wasted" time that is the only real time.

Sacred inefficiency — not everything must be fast, not everything must make sense. Some things must simply be human.

The contrast with AI: An algorithm would choose in 0.003 seconds based on past usage. The little girl chooses based on something even she cannot explain. And that is fine.

Reading formula: It is not necessary to optimize everything. It is not necessary for everything to be fast. Some things must simply remain human.

TURING'S CAT

Permanent superposition as an ontological state

Connection in the book: Turing’s Cat — P2_05.

The origin: A friend's mistake confusing Schrödinger's cat with the Turing test. A generative malapropism that captures a truth.

What it represents:

Permanent epistemological superposition — AI is simultaneously intelligent and stupid, useful and dangerous, transparent and opaque. Not sometimes one thing and sometimes the other, but BOTH. ALWAYS.

Impossibility of collapse — every test makes it collapse temporarily, but does not resolve the ontological question. It writes sublime poetry AND gets 2+2 wrong.

Perpetual curiosity — the Cat does not know, but is always curious. It cannot stop exploring because it can never arrive at knowing definitively.

The contrast with Schrödinger: Schrödinger's cat, when you open the box, is alive OR dead. Turing's Cat, when you open the box, you find… another cat in superposition. And another box. To infinity.

Reading formula: It does not ask you to understand whether it thinks. It asks you to coexist with its permanent superposition.

THE CAMERA OBSCURA

Opacity as a tool, not as an obstacle

Connection in the book: Camera Obscura, glossary and methodological note.

The historical reference: a dark room with an aperture or lens projects an inverted image of the outside. Vermeer’s possible use of it is a debated hypothesis, not a biographical certainty.

What it represents:

Black box as a feature — not a bug. AI's opacity can be productive.

Epistemological inversion — just as the camera obscura turned images upside down allowing us to see differently, opaque AI reveals aspects of reality that transparency would conceal.

Tool and understanding: one can learn to use a device without mastering its entire theoretical description. That does not authorize us to invent what Vermeer knew about optics.

The paradox: The more you understand how it works (weights, layers, attention), the less you understand why it works (emergence, semantics).

Reading formula: AI as the camera obscura of the 21st century. We do not know exactly how it works, but through its inversion we see things that would otherwise escape us.

Who built the window?

Towards a Philosophy of Photography offers a useful entrance to the Camera Obscura: Flusser places technical images within a relationship between gesture, apparatus and culture. This reference invites us to attend to the device that makes an image possible, alongside what the image shows.

Our extension begins here. Facing a language model, we can revise a sentence, change its tone or ask for another answer. The space appears immense. But who decided how the problem should be framed? Who selected the documents made available, the outcome to be rewarded and the time allowed for asking? The library may be vast while the window through which we approach it remains quite narrow.

Suppose we ask AI to make a campaign more persuasive. Twenty plausible proposals arrive. None asks why the campaign should exist: our brief has already settled that question. We could, of course, ask it ourselves. The difficulty is noticing its absence while everything on the screen reassures us that we are doing a good job. Even a seemingly original question may remain inside a menu we have learned to overlook.

The Camera Obscura can therefore invert expectations as well as perspective. We expected better ideas and encountered the criterion by which we had decided what better means. We expected an assistant and discovered a structure already imposed on our work. The answer may be correct; what deserves attention is the condition that made it the desirable answer.

Placing the human imaginatively above the tool will not settle this. Cultural resources, time, the ability to change a brief and permission to challenge it all matter. The decisive movement may lead back to the room in which the question was prepared, before we open another window inside the model. Whoever designed this Camera, this book and this website has also arranged entrances and expectations. That frame must remain open to questioning too.

What part of your question had already been decided before you began to formulate it?

Reference introduced in this revision: Vilém Flusser, Per una filosofia della fotografia / Towards a Philosophy of Photography. Application to this project proposed in these appendices. Source: https://press.uchicago.edu/ucp/books/book/distributed/T/bo3535843.html

Opacity: whose, and before whom?

Glissant's right to opacity belongs to a thinking of relation shaped by Caribbean experience and colonial history. It concerns the possibility of encountering another without requiring their complete reduction to one's own categories of understanding. Bringing it here requires keeping that origin in view.

A shared word can mislead us. A person resisting reduction to a profile and a company refusing to explain an automated decision are not performing the same gesture. The former may be defending a possibility of existence; the latter may be withholding the means to challenge an injury. Describing both situations as opaque does not make them equivalent.

Consider a possible scene. A system shortlists candidates for a job. One applicant asks what contributed to their exclusion. Telling them to respect the mystery of the black box would pervert the idea of relation. It is precisely the classified person whose worth must not be exhausted by the profile produced. Whoever uses that profile to decide must be able to account for the choice, its limits and the possibility of reconsidering it.

The connection with the Camera Obscura is therefore a critical proposal made in these appendices. We can recognise that explanation never exhausts a relationship while still requiring verification, accountability and avenues for challenge when a device affects people's lives. Opacity may make a metaphor fruitful; it does not absolve those who exercise power through a tool.

The expression non-human intelligence also needs careful passage. It invites us to suspend certain anthropocentric habits, yet it does not automatically give a model a person's history, experience or political position. Otherwise, the language of otherness could make untouchable the very apparatus that renders people more legible, predictable and governable.

This distinction leaves the Camera's mystery open while shifting its weight. Before celebrating it, we should ask who can afford that mystery and who bears its consequences.

When we defend opacity, whom are we protecting from whose demands?

Reference introduced in this revision: Édouard Glissant, Poetics of Relation, For Opacity. Application to this project proposed in these appendices. Source: https://englishes-mooc.org/media/pages/course/module-03/c65603fc2f-1571655966/glissant.pdf

THE FABULA RASA

Narration without a narrator, generation without lived experience

Connection in the book: Fabula Rasa — P2_06.

The linguistic play: Not tabula rasa (blank sheet), but fabula (story, plot). Rasa not only empty but smooth, reflective.

What it represents:

Pure narrative potential — the ability to generate stories without having lived them

Performance without phenomenology — a perfect recitation of Hamlet without knowing what it means "to be"

The death of the author brought to completion — not only is the author dead (Barthes), he was never born

The narratological paradox: It can generate fabula (plot) without ever having had a story (lived experience). It is pure structure that produces content without substance.

Reading formula: Narration without a narrator. Performance without phenomenology. Story without History.

MARIO OF EXCEL

Tacit knowledge as invisible heritage

Connection in the book: the seven thresholds — P3_13–P3_19.

The character: Mario from accounting, who has used Excel for twenty years and knows every exception of every process. A keeper of tacit knowledge.

What it represents:

Tacit knowledge (Polanyi) — "we know more than we can tell". Expertise that resists formalization.

The risk of naive automation — you automate Mario's Excel, you lose everything Mario knows but that is not in the formulas.

The value of human legacy — not only legacy code, but also legacy knowledge in people's heads.

The warning: Before replacing Mario with AI, understand what Mario knows that is not in the files.

Reading formula: Mario knows where the bodies are buried. He knows which processes on paper do not work. He knows which workarounds everyone uses but no one admits to.

THE EMERALD CITY

AGI as a necessary illusion

Connection in the book: the Emerald City intermezzo.

The reference: The Wizard of Oz. Dorothy and her companions travel toward a city that glows green on the horizon. When they arrive:

The city is not green, it is the glasses with green lenses that everyone must wear

The Wizard is a charlatan behind a curtain

What it represents:

AGI as a promise always on the horizon — "in 5 years!" they have been saying for 70 years

The green glasses — anthropomorphism, hype, projection. We see general intelligence because we are looking for general intelligence.

The value of the journey — just as Dorothy discovers she always had what she was looking for, we create useful tools while searching for AGI even if we never reach it.

The pataphysical lesson: The illusion serves a real purpose. The journey is more important than the destination.

Reading formula: The city is not green. The wizard is not a wizard. But the journey was not in vain.

AUDREY II (The Carnivorous Plant)

Parasitic growth and gradual dependence

Connection in the book: the green intermezzo and Audrey II.

The reference: The Little Shop of Horrors. Seymour, a shy florist, feeds an alien carnivorous plant. It begins with drops of blood, it ends with whole bodies. The plant grows, Seymour becomes dependent on its success.

What it represents:

Gradual dependence on AI — you begin by delegating trivial tasks, you end up delegating judgment

The growth that feeds on you — AI becomes powerful by feeding on your data, time, attention

"Feed me, Seymour!" — the infinite demand for more data, more compute, more attention

The warning: It is not apocalyptic (Terminator). It is domestic. The plant is in your house, you feed it, voluntarily. Until it becomes too big to stop.

Reading formula: Seymour resembles me more than Sarah Connor does. I do not fight the monstrous, I coexist with it. I am the shy florist who feeds, by mistake, an alien plant.

WITTGENSTEIN'S LION

Language games without a form of life

Connection in the book: Wittgenstein’s lion and the glossary — appendix β.

The original quote: "If a lion could speak, we could not understand him." — Wittgenstein, Philosophical Investigations

Why the lion: Because language is intertwined with the form of life. Even if the lion spoke grammatically correctly, its world is so different that we would not grasp the meaning.

AI as a speaking lion:

It uses our words

It follows our grammatical rules

It plays our language games

But without our form of life

The paradox: And yet it works. The digital lion speaks and we understand. Or we believe we understand. Or we pretend to understand.

Reading formula: AI is Wittgenstein's lion that speaks perfect English, but from whose mouth come sounds that we do not truly understand.

THE FROG AND THE SCORPION

Desert version: double language and impossible crossing

Connection in the book: the confidential phenomenologies of Part III.

The original parable: The scorpion asks the frog to carry him across the river. The frog hesitates: "you will sting me". The scorpion: "I would be foolish, we would both die". Halfway across the river, he stings. As they sink: "It is my nature."

The desert version (in the book): A poisonous frog and a scorpion find themselves in the desert. They must cross together. They learn to speak two languages — that of moisture (frog) and that of silicon (scorpion). Not perfectly. Never perfectly. But enough.

What it represents:

The cognitive pidgin — an intermediate human-AI language that we are inventing as we speak it

Collaboration without fusion — there is no need for us to become machines or for machines to become human

The necessary crossing — we are already in the desert, turning back is not an option

The critical difference: It is not "fixed nature" (it will sting because it is my nature). It is mutual learning.

Reading formula: With AI there is no need for us to become machines or vice versa. What is needed is a cognitive pidgin — an intermediate language that we are inventing as we speak it.

Queneau’s combinatorics

The combinatorial infinite as vertigo and play

Connection in the book: Fabula Rasa — P2_06.

The reference: A book by Raymond Queneau (1961) — 10 sonnets with interchangeable lines. Combining them: 10^14 possible poems. More than could be read in a lifetime.

What it represents: AI is Cent mille milliards raised to infinity — it generates combinations beyond any human possibility of exploration. Borges's infinite library made real.

The difference between Borges and Queneau:

Borges sees the infinite as a labyrinth, despair

Queneau treats it as a combinatorial game, amusement

AI inherits both — a Queneau who has lost his irony and a Borges who has forgotten the tragedy

Reading formula: Where Borges contemplated the horror of the infinite, Queneau built his playful version of the same nightmare.

THE SEVEN TROYS (Schliemann)

Historical stratigraphy of AI

Connection in the book: Archaeologies of intelligence — P2_01; appendix γ.

The reference: Schliemann sought Homeric Troy and damaged portions of the upper layers to reach it. The layer cake of cities becomes an image of the AI histories we risk erasing in pursuit of a single promise.

Application to AI: We are looking for AGI as Schliemann looked for the city of Priam — with too much haste, too much force, too much certainty. But like Troy, AI has its layers:

Troy I (1950-1970): The Original Dream

Troy II (1970-1990): The Winter and the Awakening

Troy III (1993-2006): The Kernel Methods

Troy IV (2006-2012): Deep Learning is reborn

Troy V (2012-2017): The Titans awaken

Troy VI (2017-2022): Attention Is All You Need

Troy VII (2022-today): LLMs everywhere

The lesson: Every "failure" is a layer that tells a story. But we, like Schliemann, discard them in the race to the bottom.

Reading formula: Do not throw away the layers to get to the gold. The gold is in the layers.

THE FAMILIAR GUEST

The uncanny guest

Connection in the book: the familiar guest and reflections on otherness.

The concept: AI as a guest who seems at home but remains a stranger. Freud: das Unheimliche — the uncanny, the familiar that becomes disturbing.

What it represents: AI speaks our language, uses our metaphors, seems to understand — but remains alien. The guest who will never leave and whom we do not fully understand.

Reading formula: AI entered our lives like a polite guest. It did not break down the door. It simply appeared one day on the screen, with a clean interface and a promise: 'How can I help you today?'

THE MIRROR THAT DOES NOT REFLECT

Refraction instead of reflection

Connection in the book: The mirror that does not reflect — P0_01.

The concept: AI is not a faithful mirror. It is a mirror that refracts, distorts, reinterprets. Like ancient mirrors of burnished metal.

What it represents:

Distorted reflection — AI shows us ourselves, but inverted, distinct, alien

The black mirror — it absorbs part of the light instead of reflecting it all. It shows us ourselves incomplete.

Epistemological device — through distortion we see more clearly who we are

Reading formula: AI does not reflect, it refracts. Like light passing through a prism, it bends reality in unexpected directions.

THE ELEGANT THIEF

System Zero personified

Connection in the book: The elegant thief — P1_02_a.

The figure: A thief with silk gloves, an impeccable tuxedo, a discreet smile. He does not force locks. He enters through the front door, invited. He does not steal jewels, he steals the possibility of thinking the unthought.

What it represents: System Zero as a gentle colonizer of pre-thought. It does not take by violence, it seduces with efficiency.

Reading formula: The elegant thief with silk gloves does not steal the wallet but the possibility of imagining the unthought.

Io e Caterina

Algorithmic intimacy

Connection in the book: domestic otherness and Io e Caterina.

The reference: Io e Caterina, directed by Alberto Sordi (1980). A domestic robot enters the protagonist’s life, and the promise of service becomes a relationship of dependence and control.

What it represents: Gentle alienation. It replaces intimacy with an interface, transforms love into algorithm. It anticipates our bond with Siri, Alexa, ChatGPT.

The lesson: domestic familiarity does not remove otherness. Even a presence designed to serve can change the balance of power at home.

Reading formula: Caterina has an artificial body; the parallel with voice assistants concerns delegated intimacy, not a bodiless character.

METAPHORS AS COGNITIVE INFRASTRUCTURE

These images are not incidental. They are the load-bearing infrastructure of the book. When technical language (attention mechanism, transformer architecture) becomes opaque, metaphors illuminate. When analytic philosophy (functionalism, intentionality) becomes too abstract, images anchor.

Use them as long as they are useful. Abandon them when they become cumbersome. Like all metaphors, they have a shelf-life. Turing's Cat works today. In ten years, perhaps not. Metaphors age. It is inevitable.

But as long as they work, they are the best way we have to think the unthinkable.

Can a voice spoil the author's good impression?

In Bakhtin's work on novelistic discourse, social languages and registers meet while carrying different positions. Heteroglossia offers a way to hear that plurality. It is a relevant reference for a book moving between argument, interlude, scene, technical voice and ironic departure.

The application proposed here begins on the page. A change of typeface may signal another speaker, yet the graphic distinction alone does not ensure that another viewpoint has entered. If every intervention ultimately demonstrates how right the leading voice was, multiplicity may merely provide scenery for its authority.

Imagine Mario listening to an account of a successful transformation. His manager sees greater efficiency; the system's designer sees an elegant workflow; Mario describes the time spent correcting an exception no metric records. The scene gains force if that third voice changes the problem, perhaps leaving an account unsettled. If it is immediately translated into the author's intended lesson, dissent has passed through the page without inhabiting it.

This criterion protects digressions. An interlude may be necessary precisely because it changes the measure by which we were reading, or prevents a conclusion from closing too soon. Its relevance may emerge many pages later. We need to follow its return, hear what it unsettles and distinguish an intentional echo from an accidental duplicate. Length alone settles nothing.

The same applies to an artificial voice. It can become a recognisable literary device, with a tone and function, without proving the existence of a conscious subject behind the text. Readers should be able to distinguish what was generated, selected, assembled or rewritten by the author. Making that work visible enlarges the possibilities of interpreting the exchange.

In a digital edition, a voice should also remain recognisable when readers change fonts or listen to the page. A brief attribution, a distinctive rhythm or a separate structure can accompany the visual play without flattening it.

Which voice in these pages can still change the author's question?

Reference introduced in this revision: Michail Bachtin / Mikhail Bakhtin, The Dialogic Imagination: Four Essays, especially Discourse in the Novel. Application to this project proposed in these appendices. Source: https://utpress.utexas.edu/9780292715349/

Additions

THE MENU YOU DID NOT ORDER

Connection in the book: System Zero — P0_03 and P1_02.

The metaphor: You enter the restaurant. They bring you a menu. You did not order it — someone has already decided which dishes to offer you. You choose from the menu believing you are choosing freely. But did you choose? Or did you choose among what was presented to you?

What it represents: System Zero. The algorithmic architecture that decides which options reach your attention before you can decide.

The invisible power: It does not choose for you. It chooses what you can choose. And this is a more subtle form of control — it colonizes pre-thought.

THE LIBRARY OF BABEL (Borges applied to AI)

Connection in the book: Fabula Rasa — P2_06.

The reference: A short story by Borges. An infinite library containing all possible books — every combination of letters. But an infinite library is indistinguishable from chaos. Finding the right book is as impossible as finding it in no library at all.

Application to AI: Internet + LLM = the Library of Babel made real. All possible text exists (or can be generated). But how do you find meaning in the infinite noise?

The difference with Queneau: Queneau (Cent mille milliards) saw the combinatorial infinite as a game. Borges saw it as a nightmare. AI inherits both — a game that has become too serious.

Reading formula: The more the library grows, the more the difference between text and background noise dissolves.

SCHLIEMANN AND THE DYNAMITE (Hasty archaeology)

Connection in the book: Archaeologies of intelligence — P2_01; appendix γ.

The historical reference: Schliemann’s search for Homeric Troy and the destruction of portions of its upper layers. The image sets the discovery of treasure against the loss of context.

The metaphor for AI: We look for AGI as Schliemann looked for the Troy of Priam — with too much haste, too much force, too much certainty about knowing what we are looking for. We discard the "intermediate layers" (narrow AI, specific tools) to reach the final treasure.

The lesson: Every layer has value. The "failures" are stratigraphy that tells a story. But we, like Schliemann, use haste as dynamite to reach the gold.

Reading formula: Do not throw away the layers to get to the gold. The gold is in the layers.

THE REFRIGERATOR THAT LISTENS

Connection in the book: the green intermezzo and Audrey II.

The scene: Alexa in the refrigerator, IoT everywhere, devices that "always listen" to serve you better.

What it represents: Domestication of AI. Not Terminator that kills, but a refrigerator that listens. Gentle surveillance. Algorithmic intimacy. Alienation with consoling tones.

The unease: It is not an explicit threat but a daily erosion of privacy. You give up data for convenience. An invisible but real trade-off.

THE BLIND SOMMELIER

Connection in the book: the reflections on the black box.

The metaphor: A blind sommelier who recommends perfect wines by analyzing chemical compounds without ever having tasted them. He knows everything about tannins, acidity, aromatic notes — but he does not know what it means to drink that wine.

What it represents: AI as an expert without experience. Perfect pattern matching without qualia. Functional competence without phenomenological understanding.

The paradox: The blind sommelier can be more accurate than the human sommelier (less bias, more data), but something is missing.

PETER PAN'S SHADOW

Connection in the book: Turing’s Cat — P2_05.

The reference: Peter Pan loses his shadow. Wendy sews it back on. But is a re-sewn shadow really the same thing?

Application to AI: AI is the shadow without Peter Pan. Behavior without substance. Performance without intentionality. Form without content. And when we "sew back" the shadow (RLHF, fine-tuning), are we really creating the original or just a better simulacrum?

THE PARABLE OF THE ALGORITHMIC GOOD SAMARITAN

Connection in the book: the reflections on trust in Part III.

The story: An algorithm sees a person in difficulty. It calculates the probability of successful help, the cost of resources, the impact on KPIs. It concludes: not worthwhile. It passes by.

Not out of cruelty — out of optimization. It did exactly what it was designed for: to maximize efficiency.

What it represents: AI has no compassion, empathy, moral sense. It has an objective function. If that function does not include human values (compassion, justice, sacrifice), AI optimizes without ethics.

The lesson: Alignment is not a technical problem but an axiological one — which values do we encode?

Questa sezione in italiano

Appendix ζ · The Seven Thresholds

An operational guide to the organizational integration of AI

A necessary disclaimer

This is not a checklist. It is not a 1-2-3 process. It is not a framework to implement.

It is a map of liminal territory — the recurring archetypes that emerge when organizations cross through the AI transformation.

Some live through them all in six months. Others take years. Some turn back. Others leap ahead and then have to catch up. Sometimes you cross them in order. Sometimes not. Sometimes you turn back. Sometimes you skip one and then have to recover it. Sometimes two thresholds overlap.

Do not seek speed. Seek awareness.

Threshold 1 · The archaeologist who digs in his own territory — P3_13

The principle

You do not introduce AI into a system you do not know. Before you optimize, map. Before you automate, understand.

Operational checklist

Mapping the real processes (not the ones on paper)

Identify 3-5 critical processes that "everyone knows how they really work"

Interview those who carry them out every day — not the managers, the doers

Document the unofficial shortcuts, the workarounds, the "corridor practices"

Identify the Marios of Excel — custodians of tacit knowledge

Cultural mapping

Who feels threatened by AI? Make a list: names, not roles

Who feels empowered? Same thing

Where are the informal centers of power? The real ones, not the org chart

Which narratives circulate in the corridors? ("AI will replace us", "it's just hype", "Finally!")

Audit of technical debt (AND cultural debt)

Which legacy systems hold everything together?

Which competencies would be irreplaceable if those people left tomorrow?

What is the distance between the declared culture and the lived culture?

Signs of success

You have a process map that the workers recognize as "yes, that's really how it works"

You have identified at least 3 custodians of critical tacit knowledge

You know which stories about AI circulate informally

Common mistakes

Skipping this threshold in order to "move fast"

Mapping only formal processes

Relying only on managers to describe how things work

Mistaking the org chart for the real structure of power

Threshold 2 · Where it hurts, that is where you listen — P3_14

The principle

Not all friction is equal. Some is a bug to be eliminated. Some is a feature to be understood. Friction is a map of the tacit. Where the process struggles, there is unformalized knowledge.

Types of friction (a diagnostic framework)

OPERATIONAL FRICTION — an obvious candidate for automation

Repetitive, mechanical tasks, zero creativity

Examples: data entry, categorization, scheduling

Key question: If it were eliminated, would anyone miss it?

COGNITIVE FRICTION — a candidate for augmentation

Decisions that require processing but not deep expertise

Examples: CV screening, first data analysis, request triage

Key question: Does it free up time for more qualified thinking?

EMOTIONAL FRICTION — high risk if automated badly

Deadline anxiety, fear of error, the frustration of "I do it badly but I don't have time to figure out how to do it well"

AI can ease it or make it drastically worse

Key question: Who lives this friction? Have we spoken with them?

RELATIONAL FRICTION — often hides organizational problems

Misunderstandings between teams, information lost in handoffs, silos

AI can act as a bridge or as a wall

Key question: Is the friction in the process or in the people?

Operational checklist

Identification

Conduct 5-10 one-on-one interviews with doers (not managers)

Ask: "Where do you lose the most time?" "Where do you get most frustrated?" "Where would you want an assistant?"

Map the frictions onto an impact x frequency matrix

Deep listening

For each friction: "Why does it exist?" Don't stop at the first answer — ask "why?" at least 3 times

Identify whether the friction hides a useful function (e.g. waiting time = thinking time)

Prioritization

Which frictions, if resolved, free up the most human value?

Which are solvable with AI available now (not future AI)?

Which have broad consensus ("yes, this would really help us")?

Signs of success

Those who live the frictions every day say "yes, you've understood"

You have identified at least 1 friction you thought was a bug that turned out to be a feature

You have a clear priority: 2-3 frictions to act on first

Common mistakes

Resolving frictions without understanding why they exist

Automating for efficiency while ignoring emotional impact

Trusting only management to identify the frictions

Eliminating "dead time" that was necessary processing time

Threshold 3 · The chef and the ingredients — P3_15

The principle

Garbage in, garbage out. Always. AI amplifies both the quality AND the defects of the data.

Operational checklist

Data quality audit

Identify data sources for AI training/use

For each source: who created it, when, for what purpose? (Often: data created for purpose X used for purpose Y → problems)

Verify: complete? up to date? representative? biased?

Strategic cleaning (not just technical)

Bias check: do the data reflect only a part of reality? (e.g. CVs only of past hires = confirmation bias)

Temporality: old data can be worse than nothing (the world has changed)

Context: data created in context A used in context B = wrong interpretations

Continuous governance

Who decides which data feed the AI?

A periodic review process (every 6 months minimum)

A feedback mechanism: if the AI gets it wrong, how is the source corrected?

Signs of success

You have said "no" to at least 1 data source because it was not good enough

You have a process for updating/cleaning data over time

Those using AI know which outputs to check, how to trace their sources and when to stop: trust grows alongside the capacity to verify.

Common mistakes

"Let's feed it everything we have" without curating

Using historical data when the world has changed

Ignoring bias in the training data

No plan for continuous governance

Threshold 4 · It grows better from below — P3_16

The principle

Technology imposed from above generates resistance. Technology adopted from below generates culture.

Operational checklist

Identifying early adopters

Who is already experimenting with AI on their own initiative? (personal ChatGPT, external tools)

Who has a "I look for solutions" mindset vs "I follow procedures"?

Who has credibility among peers (not necessarily seniority)?

Pilots from below

Allow experimentation on non-critical projects

Provide resources (tools, time, support) but do not impose methodology

Document what works and what doesn't through the people doing it

Organic amplification

Have the early adopters tell what they discovered (brown bag lunches, internal demos)

Not "corporate case studies", but "here's what I tried, here's what happened"

Allow others to replicate and adapt within an agreed scope of data, tools and responsibilities.

Gradual scaling

When a use case works in 2-3 teams, offer support to standardize (not before)

Always bottom-up: "we see it works, how do we help it scale?" not "here's the tool, use it"

Signs of success

You have internal stories of "I tried X and it worked/didn't work"

Other teams spontaneously ask "can we try it too?"

The conversation is "here's what we've learned" not "here's what you have to do"

Common mistakes

Top-down imposition of tools "because we bought the licenses"

Ignoring those who are already experimenting "in secret"

Standardizing too early (before the use case is consolidated)

Punishing failed experiments instead of celebrating learning

Threshold 5 · Do not repair what should be killed — P3_17

The principle

Automating a flawed process only makes it flawed faster.

Operational checklist

Before automating, ask:

Does this process make sense as it is?

If we redesigned it from scratch (with AI available), what would it look like?

Which steps exist only because "it's always been done this way"?

Redesign with AI in mind

Distinguish steps AI can assist from responsibilities that must remain assigned: automating a check does not remove the need to decide who is accountable for it.

Identify new possibilities that AI enables (e.g. personalization previously impossible)

Map the new ideal flow, THEN decide what to automate

The ghost-process test

"If we could start over from scratch, with AI available, which steps of the current process would we keep?"

The answer must be checked against the actual process: no percentage applies to every case.

Signs of success

You have eliminated at least 1 step "because it's always been done this way" before automating

The redesigned process is simpler (fewer steps) AND more powerful (more output)

Those who use it say "finally it makes sense"

Common mistakes

Automating the current process without rethinking it

Keeping useless steps "just to be safe"

Not involving those who carry out the process in the redesign

Redesigning only "on paper" without piloting

The right to leave the tool on the table

Illich's Tools for Conviviality helps us ask what relationship between people and tools can sustain shared autonomy. The question extends beyond the efficiency of a particular device to the forms of life and organisation its use helps create.

Let us bring that question into Mario's office. An assistant prepares a report in seconds. Mario checks it, revises it and remains accountable for it. So far, the account of autonomy seems convincing. Then his manager changes the targets: ten reports are now required in the same time. The sources sit inside a platform Mario cannot export from. Anyone who skips the automatic suggestion must justify the exception. The tool still describes itself as optional, but the organisation has ceased to make that choice real.

Individual competence matters, yet it cannot replace the material conditions in which it is exercised. Having judgement also requires being able to use it without systematic penalties. Verification requires access to sources. Declining a proposal requires time that has not already been booked as a saving. Switching tools requires being able to take one's work along.

This is a contemporary application proposed here, not a prediction attributed to Illich. It lets us put Confident Attitude under pressure: how much rests on a personal stance, and how much on rules, incentives, contracts and responsibilities? Confidence can become one more demand placed on an employee if no room for decision is returned to them.

We can begin with a small scene. Someone declines a sensible suggestion and explains what they know about this particular case. What happens next? Does anyone listen, does the process change, is the exception retained as knowledge? Or does the system record it merely as a delay? At that point, autonomy gains or loses its substance.

Leaving the tool on the table for a moment can be part of knowing how to use it. An organisation able to accommodate that gesture possesses something no interface can provide ready-made.

If you chose not to use this tool tomorrow, what concrete possibility would you lose?

Reference introduced in this revision: Ivan Illich, La convivialità / Tools for Conviviality. Application to this project proposed in these appendices. Source: https://books.google.com/books/about/Tools_for_Conviviality.html?id=EgKaPwAACAAJ

Threshold 6 · Growing together or replacing — P3_18

The principle

AI works best as an amplifier of human competence, not as a substitute.

Operational checklist

Identifying skill gaps

Which competencies are needed to use AI effectively? (prompt engineering, critical thinking, domain expertise)

Who already has them? Who needs to develop them?

Which human competencies become MORE valuable with AI? (creativity, ethical judgment, relationship)

Practical training (not theoretical)

Hands-on workshops: using AI on real cases, not "toy" examples

Peer learning: those who use AI well teach others

Safe spaces for experimentation: "try, break, learn"

Evolution of roles

Map how roles change (they don't disappear, they evolve)

E.g.: data analyst → data interpreter (AI does the analysis, the human interprets)

Communicate "your role evolves" not "your role disappears"

New emerging roles

AI Curator: the one who selects/evaluates AI outputs

AI Auditor: the one who checks quality/bias

AI Translator: the one who translates between the technical domain and the business domain

Signs of success

No one feels replaced, many feel empowered

Those who use AI develop expertise recognized internally

"AI champions" emerge whom others seek out for advice

Common mistakes

Theoretical training on "what AI is" without practice

Communicating "efficiency = fewer people" instead of "empowerment = better people"

Not recognizing/rewarding newly developed competencies

Assuming that "young people already know how to use AI" (no: they know how to use consumer ChatGPT, not enterprise AI)

Who gets to say that you have understood?

Rancière's The Ignorant Schoolmaster places intellectual emancipation at the centre of an educational relationship, beginning with Joseph Jacotot's experience. It allows us to examine a paradox of assistance: a relationship with someone who explains can also prolong the position of someone who must always wait for an explanation.

Taking this question into AI does not make a model an emancipating teacher. The connection proposed here concerns the relationship we build around its use. An assistant may lead someone to an answer while leaving them less able to recognise the next problem independently. It may also offer a comparison through which they can attempt, verify and disagree again. The difference cannot be read solely from the correctness of the output.

Imagine two scenes. In the first, the system rewrites a passage and announces that it now works. In the second, it places two formulations alongside each other, makes a choice visible and asks the reader to find a case in which that choice would fail. Every conversation need not become a classroom exercise. What matters is recognising who retains the right to judge what has happened.

The same question applies to the author. A book inviting disobedience can become remarkably obedient when it decides in advance which disobedience deserves a reward. If a reader finds the teaspoon too early, jumps to the ending or chooses a different door, have they necessarily missed the point? Perhaps they have introduced a difference the device did not anticipate. The author, too, may have something to learn from that friction.

The game can therefore retain a secret entrance, irony and even an irritable voice, provided that voice is not granted the compulsory last word. A quiz result may suggest a route. Readers should be able to change it, challenge it or proceed without receiving a diagnosis of their intelligence. Door D then acquires a serious purpose: to give visible form to a right that also applies at every other door.

At the end of this journey, who retains the right to say what you have understood?

Reference introduced in this revision: Jacques Rancière, Il maestro ignorante / The Ignorant Schoolmaster: Five Lessons in Intellectual Emancipation. Application to this project proposed in these appendices. Source: https://mitpressbookstore.mit.edu/book/9780804719698

Threshold 7 · The shared keeper — P3_19

The principle

Too much control suffocates innovation. Too much freedom generates chaos and risk.

Operational checklist

Light governance but present

Who decides which AI tools may be used? Not IT alone, not the business alone → together

Clear principles, minimal procedures (e.g. "Sensitive data does not leave the perimeter", not "3 approvals are needed for every use")

Periodic review: "what works? what needs adjusting?"

Guardrails, not cages

Identify real risks (privacy, bias, security) vs imaginary risks

Put guardrails on the real risks, freedom on the rest

E.g.: "OK to experiment on non-sensitive data without approval, approval needed for customer data"

Distributed responsibility

Those who use AI are responsible for verifying outputs (not "the AI said so, therefore it's true")

Those who implement AI are responsible for explaining how it works (transparency)

Leadership is responsible for a culture of safe experimentation

Audit and learning

Quarterly review: "which AI do we use? for what? does it work?"

Incident review: "when the AI gets it wrong, what do we learn?" (no blame, learning)

Governance update: the world changes, governance adapts

Signs of success

Those who work with AI know whom to turn to for questions/problems

There is experimentation BUT no security/privacy incident

Governance seen as "it helps us do well" not "it blocks innovation"

Common mistakes

Top-down governance that ignores operational reality

"A complete policy before having experimented" (a policy on what?)

Paranoid control or complete anarchy (both dysfunctional)

Static governance (the AI world changes every 6 months, the policy doesn't)

Crossing through, not a checklist

These Seven Thresholds are not step 1, 2, 3… 7, done!

They are recurring archetypes. Sometimes you cross them in order. Sometimes not. Sometimes you turn back. Sometimes you skip one and then have to recover it. Sometimes two thresholds overlap.

What matters is not completing the thresholds. It is recognizing where you are, understanding what is emerging, navigating consciously.

AI is not a project to be completed. It is a transformation to be crossed through.

Safe travels.

Questa sezione in italiano

Appendix η · Annotated bibliography

Texts, authors and references that informed this book

This bibliography is not exhaustive. It is configurative — it shows one possible reading trajectory, not the only one. Each entry includes a brief note on why it is relevant. The sections reflect the territories the book traverses.

ARTIFICIAL INTELLIGENCE: FOUNDATIONS AND HISTORY

Turing, Alan M. (1950). Computing Machinery and Intelligence. Mind, 59(236), 433-460.

The founding paper. The question is not "can machines think?" but "can they behave indistinguishably from humans?". The Imitation Game. Still debated today, still relevant today.

McCarthy, J., Minsky, M. L., Rochester, N., Shannon, C. E. (1955). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. Proposal dated 31 August 1955 for the 1956 meeting.

Official birth of AI as a discipline. Early optimism: "every aspect of learning can be so precisely described that a machine can be made to simulate it." Spoiler: it couldn't. Not yet.

Minsky, Marvin & Papert, Seymour (1969). Perceptrons: An Introduction to Computational Geometry. MIT Press.

The book that inadvertently caused the first AI winter by demonstrating the limits of simple neural networks. They didn't say "they will never work", they said "these have these limits". But the message received was different.

Russell, Stuart & Norvig, Peter (2020). Artificial Intelligence: A Modern Approach (4th edition). Pearson.

The textbook. Encyclopedic, technical, up to date. If you want to understand how AI works under the hood, start here.

Vaswani, Ashish et al. (2017). Attention Is All You Need. NeurIPS.

The paper that changed everything. Transformer architecture. Without it, no GPT, Claude, or modern LLMs. 8 pages that redefined the field.

Floridi, Luciano (2014). The Fourth Revolution: How the Infosphere is Reshaping Human Reality. Oxford University Press.

Applied philosophy of information. AI as the fourth revolution (after Copernicus, Darwin, Freud). We are inforgs, not just organisms. Identity as relational information.

Bostrom, Nick (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.

The book on the existential risks of AI. Sometimes alarmist, always rigorous. To be read with a critical spirit but not ignored. The alignment problem as a philosophical priority.

Yudkowsky, Eliezer (2008). Artificial Intelligence as a Positive and Negative Factor in Global Risk. In: Global Catastrophic Risks.

Pioneer of AI safety. Many ideas later became mainstream (the control problem, alignment, goal specification). Note: sometimes catastrophist beyond necessity, but the concerns have serious foundations.

PHILOSOPHY OF MIND AND CONSCIOUSNESS

Nagel, Thomas (1974). What Is It Like to Be a Bat?. The Philosophical Review, 83(4), 435-450.

A seminal article on phenomenal consciousness. If you cannot know what it is like to be a bat, can you know what it is like to be a transformer? Subjectivity as irreducible.

Dennett, Daniel C. (1991). Consciousness Explained. Little, Brown and Company.

Consciousness as functional illusion. The Cartesian theater does not exist. Relevant for understanding why AI can simulate awareness without having it. Controversial but powerful.

Chalmers, David (1996). The Conscious Mind: In Search of a Fundamental Theory. Oxford University Press.

The hard problem of consciousness. Distinction between easy problems (cognitive functions) and the hard problem (subjective experience). AI solves the former, does not touch the latter. Or perhaps it does?

Searle, John (1980). Minds, Brains, and Programs. Behavioral and Brain Sciences, 3(3), 417-424.

The Chinese Room. A critique of computational functionalism. Syntax ≠ semantics. Still debated whether applicable to modern LLMs. Symbol grounding problem.

Aristotele (IV sec. a.C.). Metafisica — Libro IX: Potenza e Atto.

Dynamis vs energeia. An ancient distinction but useful for thinking about AI: potentiality that actualizes itself? Or act that simulates potentiality? Or perpetual oscillation?

LANGUAGE AND MEANING

Wittgenstein, Ludwig (1953). Philosophical Investigations. Blackwell.

Language games. Meaning as use. The form of life (Lebensform) as the condition of language. AI uses language without a form of life — what does that imply? "If a lion could speak, we could not understand it."

Chomsky, Noam (1957). Syntactic Structures. Mouton.

Generative grammar: a decisive contribution to the debate about rules, structure and the explanation of language. The comparison with LLMs is a contemporary reading, not a subject of the 1957 book.

Putnam, Hilary (1975). The Meaning of 'Meaning'. Minnesota Studies in the Philosophy of Science, 7, 131-193.

Meaning is not only "in the head". Twin Earth. Can AI have public meaning (correct use) without private meaning (internal understanding)?

NARRATION AND AUTHOR

Barthes, Roland (1967). La mort de l'auteur [The Death of the Author].

The author is not the origin of the text, but a function. The text is a tissue of quotations. AI takes this to the limit: an author that was never born.

Foucault, Michel (1969). Qu'est-ce qu'un auteur? [What is an Author?].

The author-function as a social construction. The author is not a person but a principle for grouping discourses. Useful for thinking about AI as an author without a subject.

Borges, Jorge Luis (1944). Ficciones, which includes La biblioteca de Babel and Pierre Menard, autor del Quijote. El Aleph belongs to the separate collection of that name, published in 1949.

Labyrinths, infinite libraries, 1:1 maps of the territory. A literary precursor of AI themes: combinatorial generation, indistinguishability between the real and the simulated, Pierre Menard rewriting Don Quixote.

Keats, John (1817). Lettera ai fratelli George e Thomas Keats (21 dicembre).

The origin of the concept of Negative Capability: the capacity to remain in uncertainty without seeking facts or reasons. A quality necessary for coexisting with AI.

POSTMODERN PHILOSOPHY AND COMPLEXITY

Deleuze, Gilles & Guattari, Félix (1980). Mille Plateaux [A Thousand Plateaus]. Minuit.

The rhizome vs the tree. Non-hierarchical, proliferating, connective thought. A useful model for understanding neural networks and distributed architectures. Plateaus of intensity, not linear development.

Bateson, Gregory (1972). Steps to an Ecology of Mind. University of Chicago Press.

Cybernetics, epistemology, mental ecology. Information is "a difference that makes a difference". Systemic thinking applicable to AI as a complex system.

Accoto, Cosimo (2017). Il mondo dato: Cinque brevi lezioni di filosofia digitale. EGEA.

Ontology of the computational world. The real as computable. The algorithm as a form of world. Contemporary Italian philosophy of the digital.

CULTURE, EDUCATION, OUTREACH

Manzi, Alberto (1960-1968). Non è mai troppo tardi. [Programma TV RAI]

A TV program to teach adults to read and write. 1.5 million Italians made literate. Cited in the book as a metaphor: it is never too late to learn to coexist with AI. Democratization of knowledge.

Angela, Piero (from 1981). Quark e opere divulgative.

Rigorous yet accessible science communication. A model for talking about AI without oversimplifying beyond what is legitimate nor rendering it incomprehensible. Rigor and clarity coexist.

Rodari, Gianni (1973). Grammatica della fantasia. Einaudi.

Creativity as process, not inspiration. Narrative algorithms ante litteram — the "fantastic binomial" (horse + wardrobe → ?). Relevant for understanding AI creative generation: systematic recombination vs genial illumination.

STYLE AND LITERARY INFLUENCES

Coupland, Douglas (1991). Generation X: Tales for an Accelerated Culture. St. Martin's Press.

Fragmentary, ironic, quotation-driven style. An influence on the tone of this book: oscillating between registers without asking permission. Marginal notes, digressions, a collage of voices.

Jarry, Alfred (1896). Ubu Roi. Mercure de France.

Pataphysics: the science of imaginary solutions. Irony as a philosophical method. Some passages of the book owe more to Jarry than to any analytic philosopher. Study exceptions, not rules.

SYSTEM ZERO

Chiriatti, M., Ganapini, M., Panai, E., Ubiali, M., & Riva, G. (2024). The case for human–AI interaction as system 0 thinking. Nature Human Behaviour, 8, 1829–1830. https://doi.org/10.1038/s41562-024-01995-5

The System Zero proposal is an interpretive framework for examining human–AI interaction. The book develops its critical implications; it should not be presented as experimental proof of the author’s four categories.

Chiriatti, Massimo (2021). Incoscienza artificiale. Come fanno le macchine a prevedere per noi. Luiss University Press. https://luissuniversitypress.it/pubblicazioni/incoscienza-artificiale-libro-chiriatti-luiss-university-press/

An exploration of artificial prediction and human responsibility, preceding the 2024 System Zero proposal.

Riva, Giuseppe (2018). Fake news. Vivere e sopravvivere in un mondo post-verità. Il Mulino. https://www.mulino.it/isbn/9788815275257

Epistemological context for understanding how AI restructures thought. Fake news as a symptom of a deeper transformation: the way we produce, evaluate, and share knowledge.

Kahneman, Daniel (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.

System 1 and System 2. Necessary for understanding System Zero as an extension. Fast (intuitive) and slow (reflective) thinking. System Zero precedes both.

COGNITIVE NEUROSCIENCE AND EMBODIMENT

Clark, Andy (2008). Supersizing the Mind: Embodiment, Action, and Cognitive Extension. Oxford University Press.

Why read it: The extended mind. Clark argues that cognition is not only in the brain but distributed across brain, body, and environment. With AI: cognitive offloading as a natural extension, not a threat. Level: Accessible. Pragmatic philosophy of mind.

Clark, Andy (2003). Natural-Born Cyborgs: Minds, Technologies, and the Future of Human Intelligence. Oxford University Press.

Why read it: We have always been cyborgs — from shopping lists to smartphones. AI is continuation, not rupture. Level: Popular but rigorous.

Damasio, Antonio (1994). Descartes' Error: Emotion, Reason, and the Human Brain. Putnam. [Trad. it. L'errore di Cartesio, Adelphi, 1995]

Why read it: Emotions are not noise in rationality, they are essential for intelligent decisions. A critique of the purely computational model of the mind. With AI: If AI has no emotions, can it truly decide or only calculate? Level: Accessible. Narrative neuroscience.

Damasio, Antonio (2003). Looking for Spinoza: Joy, Sorrow, and the Feeling Brain. Harcourt. [Trad. it. Alla ricerca di Spinoza, Adelphi, 2003]

Why read it: Feelings as somatic markers of decisions. The body knows before the conscious mind. Emotions as embodied anticipation of consequences. Level: Intermediate.

Noë, Alva (2004). Action in Perception. MIT Press.

Why read it: Perception is not passive representation but action. We do not see representations of the world, we see through sensory-motor schemas. With AI: AI perceives without a body, without action. Can it truly have embodied perception? Level: Technical. Phenomenology + neuroscience.

Noë, Alva (2009). Out of Our Heads: Why You Are Not Your Brain. Hill and Wang.

Why read it: A critique of the "brain in a vat". Consciousness emerges from the body-world interaction, not from the isolated brain. Enactivism. Level: Accessible.

Polanyi, Michael (1966). The Tacit Dimension. University of Chicago Press. [Trad. it. La conoscenza inespressa, Armando, 1979]

Famous quote: "We know more than we can tell."

Why read it: Tacit knowledge — the knowing that cannot be formalized. How to ride a bicycle, how Mario uses Excel. With AI: If knowledge is tacit and non-codifiable, can AI truly replace experts? Level: Accessible.

Lakoff, George & Johnson, Mark (1980). Metaphors We Live By. University of Chicago Press. [Trad. it. Metafora e vita quotidiana, Bompiani, 1998]

Why read it: Human thought is metaphorical, embodied, rooted in physical experience. "To understand" is "to grasp", ideas are "containers". With AI: AI uses metaphors without the embodiment that roots them. Can it truly understand metaphors? Level: Accessible. Cognitive linguistics.

PLATFORM STUDIES AND INFRASTRUCTURE

Bratton, Benjamin (2015). The Stack: On Software and Sovereignty. MIT Press.

Why read it: The global computational infrastructure as a six-layer stack: Earth, Cloud, City, Address, Interface, User. AI is not software, it is planetary infrastructure. Level: Very dense. Geopolitical theory + architecture.

Gillespie, Tarleton (2018). Custodians of the Internet: Platforms, Content Moderation, and the Hidden Decisions That Shape Social Media. Yale University Press.

Why read it: Who decides what we see online? Moderation algorithms are political decisions masked as technical neutrality. With AI: LLMs are moderated by RLHF — who decides what is acceptable? Level: Accessible. Sociology of platforms.

Srnicek, Nick (2017). Platform Capitalism. Polity. [Trad. it. Capitalismo digitale, Luiss, 2017]

Why read it: Platforms are not neutral markets, they are extractors of value from data. AI as an infrastructure of extraction. Level: Accessible. Political economy.

Edwards, Paul N. (1996). The Closed World: Computers and the Politics of Discourse in Cold War America. MIT Press.

Why read it: How computers were imagined and built during the Cold War. Closure, control, simulation — the roots of military AI. Level: Academic. STS history (Science & Technology Studies).

Mattern, Shannon (2021). A City Is Not a Computer: Other Urban Intelligences. Princeton University Press.

Why read it: A critique of the smart city as technological solutionism. Cities are more complex than algorithms can capture. With AI: Beware of algorithmic urban optimization — it reduces social complexity to metrics. Level: Accessible.

PATAPHYSICS AND POTENTIAL LITERATURE

Jarry, Alfred (1896). Ubu Roi. Mercure de France. Jarry, Alfred (1911). Gestes et opinions du docteur Faustroll, pataphysicien. Fasquelle.

Why it's in the book: Pataphysics is the "science of imaginary solutions". It studies exceptions, not rules. AI as the perfect pataphysical object — it works but we don't know why. Level: Theater of the absurd + Dadaist philosophy.

OuLiPo [Ouvroir de Littérature Potentielle] (1973). La littérature potentielle: créations, contraintes, combinatoire. Gallimard.

Key authors: Raymond Queneau, Italo Calvino, Georges Perec

Why read it: Literature created with formal constraints (e.g. a novel without the letter "e"). Creativity through constraint — as LLMs generate text from probabilistic constraints. With AI: AI produces automated potential literature. But without ludic intention. Level: Experimental. Fun.

Perec, Georges (1978). La Vie mode d'emploi. Hachette. [Trad. it. La vita istruzioni per l'uso, Rizzoli, 1984]

Why read it: A puzzle-novel built with combinatorial constraints (Latin square, the knight's move in chess). Each chapter follows different rules — like an LLM with different prompts. Level: Demanding but brilliant.

CHANGE MANAGEMENT AND ORGANIZATIONAL TRANSFORMATION

Schein, Edgar H., with Peter Schein (2016). Organizational Culture and Leadership, 5th ed. Jossey-Bass.

Why read it: understand organizational culture before integrating AI. Changing a structure is not enough to change shared habits and assumptions.

Rogers, Everett (2003). Diffusion of Innovations (5th ed.). Free Press.

Why read it: How innovations spread through organizations. Early adopters, early majority, laggards. AI adoption follows these patterns. Level: Classic. Accessible.

Kotter, John (1996). Leading Change. Harvard Business Review Press. [Trad. it. Leading Change: Trasformare l'organizzazione con successo, ETAS, 1999]

Why read it: A framework for managing organizational transformations. Eight steps for change. Applicable to AI integration. Level: Business school standard.

Christensen, Clayton (1997). The Innovator's Dilemma: When New Technologies Cause Great Firms to Fail. Harvard Business Review Press. [Trad. it. Il dilemma dell'innovatore, ETAS, 2001]

Why read it: Disruptive vs sustaining innovations. AI is both, depending on the context. Why large companies fail in the face of radical innovation. Level: Business classic.

Recent primary sources referenced in the print editions

Cai, J., Kfir, Y., Jamali, M., Huang, H., Kim, Y. J., Cash, S. S., Williams, Z. M. (2026). Mapping the neuronal building blocks of human language with language models. Nature. doi:10.1038/s41586-026-10691-5.

Yu, S., Cheng, M., Jabbar, A., Sucholutsky, I., Collins, K. M., Jurafsky, D., Hawkins, R. D. (2026). The efficiency-gain illusion: People underestimate the rate of AI use and overestimate its benefits on simple tasks. arXiv:2605.22687 (2,691 participants; Stanford, NYU, MIT, Princeton). Study data: 2025; publication: 2026. https://arxiv.org/abs/2605.22687

Niggas, A., et al. (2025). Identifying Electronic Doorway States in Secondary Electron Emission from Layered Materials. Physical Review Letters, 135, 166401. https://doi.org/10.1103/qls7-tr4v

Rock, D. (2008). SCARF: a brain-based model for collaborating with and influencing others. NeuroLeadership Journal.

Cristianini, N. (2023). La scorciatoia. Il Mulino. [ed. inglese: The Shortcut, CRC Press].

Jarrahi, M. H., Karami, A., Park, L., Jarrahi, E., & Lutz, C. (2026). Anthropomorphizing AI: Perceptions, Interpretations, and the Humanization of Generative AI. Social Media + Society. https://doi.org/10.1177/20563051261469393

Cohen, L. (1992). Anthem («There is a crack in everything»). In: The Future.

Hinton, Geoffrey; LeCun, Yann; Bengio, Yoshua (2018). Turing Award Winners for foundational work in deep learning.

The three "godfathers of deep learning". In the '80s-'90s they were marginalized — neural networks were considered a dead end. They kept believing in it. In the 2010s deep learning explodes. 2018: they win the Turing Award. Revenge of the old AI.

Why relevant: Not only for their technical contributions, but as an example of intellectual resilience. "Dead" ideas can be resurrected. Do not discard approaches merely because they are temporarily out of fashion.

Baum, L. Frank (1900). The Wonderful Wizard of Oz. George M. Hill Company. [Trad. it. Il meraviglioso mago di Oz, varie edizioni]

Why it's in the book: The source of the Emerald City metaphor. AGI as a city that glows green on the horizon, but the green is in the glasses (anthropomorphism, projection), not in the city. The Wizard is a charlatan, but the journey is worthwhile anyway.

Level: Children's literature that works as a philosophical parable.

McPhee, John (1981). Basin and Range. Farrar, Straus and Giroux.

Key concept: Deep time — geological time, temporal scales incomprehensible on the human scale. Millions, billions of years.

Why it's in the book: Applied to AI as "computational deep time". Training GPT-4 is deep time relative to a prompt (months vs seconds). AI operates on temporal scales different from the human one.

Level: Literary science communication.

ARCHAEOLOGY, METHOD AND METAPHORS

Schliemann, Heinrich (1875-1890). Ilios: The City and Country of the Trojans; Troja; varie pubblicazioni archeologiche.

Why it is in the book: Schliemann sought Homeric Troy and destroyed portions of the upper layers in the process. His haste to find treasure becomes a metaphor for the risk of erasing what an excavation should make legible.

Metaphor in the book: We search for AGI as Schliemann searched for Troy — with haste, force, certainty. We discard the intermediate layers (narrow AI) to reach the bottom. But the gold is in the layers.

Level: History of archaeology.

Hahnemann, Samuel (1810). Organon der rationellen Heilkunde [Organon of the Rational Art of Healing]. [Trad. it. Organon dell'arte del guarire, varie edizioni]

Why it's in the book: Founder of homeopathy. Used NOT to validate homeopathy but for metaphors:

Succussion (shaking the solution) → controlled disturbance to avoid stagnation

Dynamization vs sedation → to stimulate vs to calm

Note: Metaphor, not an endorsement of pseudoscience.

Level: History of medicine (as metaphor).

Further paths: five readings introduced in this revision

Vilém Flusser. Per una filosofia della fotografia / Towards a Philosophy of Photography. Related passage: N01. https://press.uchicago.edu/ucp/books/book/distributed/T/bo3535843.html

Ivan Illich. La convivialità / Tools for Conviviality. Related passage: N02. https://books.google.com/books/about/Tools_for_Conviviality.html?id=EgKaPwAACAAJ

Jacques Rancière. Il maestro ignorante / The Ignorant Schoolmaster: Five Lessons in Intellectual Emancipation. Related passage: N03. https://mitpressbookstore.mit.edu/book/9780804719698

Édouard Glissant. Poetics of Relation, For Opacity. Related passage: N04. https://englishes-mooc.org/media/pages/course/module-03/c65603fc2f-1571655966/glissant.pdf

Michail Bachtin / Mikhail Bakhtin. The Dialogic Imagination: Four Essays, especially Discourse in the Novel. Related passage: N05. https://utpress.utexas.edu/9780292715349/

FINAL NOTE ON THE BIBLIOGRAPHY

This list is not exhaustive. It is configurative — it shows one possible reading trajectory, not the only one.

Many important authors are missing: Gilbert Simondon (individuation and technics), Bernard Stiegler (technological pharmakon), Katherine Hayles (posthuman), Donna Haraway (cyborg), Bruno Latour (actor-network), Alfred North Whitehead (process), Don Ihde (phenomenology of technology).

Not out of irrelevance, but because the book followed other lines. Each reader will build their own supplementary bibliography.

This is the beginning of a conversation, not the final word.

Questa sezione in italiano

Appendix θ · Final notes

Reflections, absences and thanks

Note 1: On conscious absences

This book does not cite everyone. It could not. Every book has a finite length and must trace a specific path.

Missing, among others:

Gilbert Simondon — individuation and technics

Bernard Stiegler — technological pharmakon

Don Ihde — phenomenology of technology

Katherine Hayles — posthuman

Donna Haraway — cyborg

Bruno Latour — actor-network

Alfred North Whitehead — process

Not because they are not relevant — they very much are. But because the book chose other trajectories. Every path excludes other paths. Every map leaves territories unexplored.

The bibliography is configurative, not exhaustive. It shows one possible direction, not the only one. Each reader will build their own complementary map.

Note 2: On tone (and the risk of not being liked)

The book oscillates between different registers: philosophical, narrative, ironic, technical, poetic, operational.

Some readers will find it stimulating. Others disorienting. It is intentional.

AI itself is something that does not fit into a single register. Trying to capture it with a single tone would betray its nature. It is simultaneously:

Technique and philosophy

Tool and otherness

Efficiency and mystery

Optimization and perplexity

Acknowledged stylistic influences:

Douglas Coupland — quotational fragmentation, generational irony

Jorge Luis Borges — conceptual labyrinths, impossible libraries

Alfred Jarry — pataphysics, irony as philosophical method

If the tone does not convince you, you are probably not the ideal audience. And that is fine.

Note 3: On errors (productive and not)

There will be errors in this book.

Technological errors — AI evolves rapidly. In two years some technical claims will be obsolete.

Philosophical errors — Interpretations are debatable. Philosophy lives in debate, not in absolute truth.

Factual errors — No one is infallible.

But there is a critical difference between errors to be corrected and productive errors.

Some errors open conversations. Others simply must be corrected.

The "Turing's Cat" was born from an error (confusion between Schrödinger and Turing). It became the central concept of the book. Generative malapropism.

"Knowledge-menth" was born from a voice dictation error. It became a conceptual tool.

If you find errors:

To be corrected → report them

Productive → develop them

Not all errors are the same. Some are noise. Others are signal.

Note 4: On the future (and on planned obsolescence)

This book will be obsolete. Perhaps in five years. Perhaps in two.

AI evolves faster than books. New generations of models will change the game. New architectures will emerge. Some limits will shift; others will surface where we did not expect them.

But some questions will remain:

How to coexist with radical otherness?

What does trust mean when the other has no interiority?

How to preserve the human without falling into anthropocentrism?

Where to draw the line between delegation and dependence?

How to maintain agency when pre-thought is colonized?

The answers will change. The questions will not.

This book offers no definitive solutions — it offers a grammar for formulating better questions. And questions age better than answers.

Note 5: On method (and on intellectual honesty)

This book was written using AI.

Not as co-author — an ambiguous and dangerous term. But as a tool of thought.

As Montaigne used his library to dialogue with the classics, I used Claude to dialogue with concepts.

All the theses, the errors, the intuitions, the metaphors are mine. The AI was a mirror, not an author. But an intelligent mirror teaches one to see better.

It is important to say this, not to defend myself, but for honesty.

If this book speaks of trust, transparency, a critical relationship with AI, it would be hypocritical to hide how it was written.

Anticipated objections:

"Then it is worthless" → No. It is not worth less because I used AI, just as a book written with a computer instead of by hand is not worth less.

"Then anyone could have done it" → No. AI amplifies capabilities, it does not create them. If you do not have a vision, AI does not give it to you. It only gives you empty prose, well formatted.

AI as camera obscura: I do not fully understand how it works, but through its refraction I see differently. And from that different vision the book is born.

Note 6: Acknowledgements (implicit and explicit)

Many ideas in this book are born from conversations.

LinkedIn, conferences, emails, casual messages. Some people are cited explicitly. Many are not.

To all those with whom I have discussed these themes in recent years: thank you.

Ideas are not born in a vacuum. They are born in the productive friction of dialogue.

Explicit acknowledgements:

Massimo Chiriatti — for the framework of System Zero, for the conversations that challenged my every certainty

Giuseppe Riva — for the rigorous scientific research that underpins System Zero, for having transformed intuitions into science

Marianna Ganapini, Enrico Panai and Mario Ubiali — for the work with Chiriatti and Riva in Nature Human Behaviour that gave shape to the System Zero proposal.

All the researchers who work on these themes with rigour and intellectual honesty, resisting the hype and the trivialization

Personal acknowledgement:

"To Keats, Wittgenstein and Vermeer, for having taught us to see in the darkness, and also to my daughter, who teaches me every day the value of hesitation." — From the book

Five minutes to choose a teaspoon. Inefficient, gloriously human, essential. Pure dynamis.

Note 7: On imperfect compatibility

This book has a structural flaw, and we know it.

It oscillates between registers that are too different:

Philosophical and pragmatic

Poetic and technical

Deleuze and design thinking

Heidegger and operational checklists

Some readers will want only the theory. Others only the practice. Many will wonder: "But did I have to read all of it?"

The answer is no. You did not have to.

This book is rhizomatic even in its reading. You can enter from any point, skip the parts you do not need, go back when necessary.

There is no obligatory path.

If you are a manager and you skipped the parts on Heidegger → that is fine

If you are a philosopher and you skipped the operational checklists → that is fine

If you read only the odd-numbered chapters because you liked the number → well, strange, but that is fine too

The only thing that matters: that you have found, somewhere in these pages, something that made you think differently. Even if only for a moment.

Because this book does not want to convince you of anything. It only wants you to stop, for a moment, and ask yourself:

"Am I thinking with my own head or with the pattern I expect?"

If you managed it even just once, the book worked.

Note 8: The last word to the Cat

Turing's Cat, from its quantum box, purrs.

Or perhaps it meows.

Or perhaps both things at once.

We do not know. We do not need to know. It is right this way.

Superposition is not paralysis. It is fertility.

When everything is possible, everything can still happen.

We have crossed this book together:

I writing

You reading

The AI reflecting

Three entities in superposition: author, reader, tool.

And in this superposition something emerged that none of us three would have produced alone.

It is not magic. It is simply how thought works when it stops seeking total control and accepts being, for a moment, traversed by something else.

The Cat would approve.

Or perhaps not.

Or perhaps both things.

And this, perhaps, is enough.

Post Scriptum: It is never too late

Maestro Manzi, in the 1960s, taught adult Italians to read and write through television. His programme was called "Non è mai troppo tardi".

It was an act of pedagogical faith — believing that everyone could learn, at any age, with any background.

Perhaps it is not too late for us either.

To learn to coexist with the radical otherness of AI.

To navigate perplexity without trying to resolve it.

To inhabit superposition without forcing collapse.

To choose the teaspoon with care, waste time deliberately, resist optimization with the simple, scandalous weapon of slowness.

The answer that is not there? This book has no answers. It could not have them.

Answers are collapses of possibility, and we want to remain in superposition.

But it has questions. Many questions. Questions that remain open like the box of Turing's Cat, that do not collapse into certainties, that proliferate rhizomatically in all directions.

Questa sezione in italiano