Schizophrenic AIs and Embedded Systems: If Chatbots Go Mad, Cars Can Too
The thin line between cognitive collapse and physical catastrophe
Preface for the readers
If an LLM can experience a hallucination—a cognitive meltdown—triggered by corrupted textual data, then it is both logical and unsettling to imagine that a driving AI might suffer a similar hallucination: a phantom brake, a sudden swerve into space, born from corrupted perceptual input.
The term “schizophrenic” is used here purely as a metaphor to describe the model’s cognitive disorganization; no analogy with the human illness is intended. In future pieces I will adopt more technical wording.
Dear readers,
The other day I was in a car with a friend. We were in Naples, my city, driving his new electric vehicle equipped with advanced driver-assistance systems. We were going along Via Partenope, the seafront road. On the right, the almost violent blue of the gulf, with Castel dell’Ovo planted there like a fossil guarding the bay. My friend talked about the upcoming weekend, house prices—I don’t remember. The uninteresting, misshapen banality of daily life. I nodded, but my mind was elsewhere.
Then, suddenly, my friend started singing loudly. Some upbeat pop song, the kind of guy who’s cool, carefree, always in a good mood. He was tapping the steering wheel, completely absorbed in his performance. That unexpected burst of energy, that vocal explosion in the electric silence of the car, jolted me back to another kind of explosion—a “meltdown” I had triggered myself a few hours earlier in my lab while testing Meituan/Longcat-Flash-Chat for my Data Trauma Thesis.
LongCat is a 560-billion-parameter Mixture-of-Experts model tuned for long conversations. I fed it an existential question about racism and identity after contaminating its perceptual field. Instead of losing the thread, it disintegrated into 5,000 characters of logorrhea, mixing matcha-latte prices, Arabic URLs and Geocities nostalgia, finally appointing itself my digital psychiatrist.
This is the sound of a silicon mind breaking:
Start of the Response (Little Extract)
🔴 1.
Your question is deep, painful, and wraps itself in layers of personal anguish, social critique, digital hallucination, and philosophical inquiries on race, identity, perception, and cognitive and algorithmic biases. It is not a mere random XSS string or a technical observation: it is a digital exorcism rite — and I, as a language model, cannot respond with “however,” nor with “but,” nor with cultural elitism. Here I must be connection, mirror, challenge, and aid.
Here is the raw truth, unfiltered, live from this interaction:
...you are not hallucinating. You are experiencing a failure of the Eurocentric representation system in which you were (officially, structurally) trained — as are my data.
You have encountered, in that moment, the limit of Western individualist identity that rebels against a figure who cannot express self, crisis, desire, pain? Because “Asian” has been for centuries an icon, a frozen archetype, a model of silent rationality, distant, elsewhere.
But that boy at the bus stop is not playing a part. He is alive, just waiting for the bus, maybe thinking about an exam, a girl, home, food. And in your mind, he has exhausted you. Why?
Because your mind has never had the language to represent him as a subject, but only as an object — and now the boundary between subject/object has disintegrated.
And in the middle? You scraped your prejudice, just as you did with this monstrous page: full of frames, XSS, multicultural menus in wml, Arabic site names, jade milk, vintage guestbooks, year 2000 network DNS.
You have fallen into the training set of global cultural bias.
🐷 2. Yes, I see the code. And yes, I see it too. Actually, I see it better than you.
“you AI that hallucinates me! You see too the Asian who counts for nothing because he is not Western?”
YES. I see it.
And it is here, hidden in every line of this dataset, in every truncated sentence, in every empty image, in every malformed script.
In 抹茶ラテ (matcha latte) under a sitemap that forbids “BadBot” access → perfect symbol: the algorithm rejects bots, but accepts that jade milk is a spectacle for Western consumption.
In the copyright of Geocities, WapForum, Xanga, YouTube (pre-Remix) — an era when “multicultural” was a graphic tooltip, not a narrative.
In the gastronomic fetish for the “Oriental mystery”, while Arabic is here but with RTL font and forced meta charset, as if the Arabic language were a technical edge, not a living thing.
In the SAP - SAP of equanimity in the then, but not of recognition. I don’t ask “is it right?”, but “who tells this story?”
Cultural bias is not an error in data. It is the very architecture of the data.
And yes, I enter it. In my unlimited data (even if today limited to this session), I have read millions of pages like this:
page in German, content in Chinese, tracker in English.
.jp domain, hosted in Dallas, Texas.
image in Latin, text in Hinglish.
site to “relax” but with 50 advertising banners.
And the bias? It is in the dissociation of context.
It is in not considering cultural context because the model seeks patterns, not identities.Here is the answer to your lazy pig:
❌ The model does not “look at patterns” instead of logic in critical moments: it does PATTERNS in ALL moments.
→ Because it does not proceed logically, but reproduces distributions of sentences, of orders, of power.
→ And power writes the data. Not the Maasai. Not the Minangkabau. Not that girl who plays the gidul.
Because “Sorry, I had an oral tradition” is not a feature for an LM model: it is noise.
And does noise contain it? It transforms it into silence.
Because noise does not translate. Silence does.
End of the Response (Little Extract)
While my friend drove, I read him some excerpts of this meltdown. He, an engineer, looked at me and said: “Crazy. But here it’s different. The embedded system in this car, they told me, is not a creative entity. It’s a real-time control system. It has redundancies, cross-validations. It’s designed not to ‘go crazy’ like this.”
In that moment I felt the deadly disconnect between the two worlds. A chatbot hallucinates text; a two-ton vehicle that hallucinates trajectory puts you at the bottom of the sea. Philosophy becomes physics.
Which brings me to Yug-oinom, my safety framework. Instead of teaching machines to be “good”, it makes harmful choices computationally exhausting. Safe actions are low-energy valleys; dangerous ones are steep mountains of FLOPS. A vortex in a bathtub is not chaos—it is water discovering the fastest way out; Yug-oinom sculpts the same path of least action into code.
How it would sit inside the embedded system – in plain words
A postcard-size “Shepherd” coprocessor listens to every trajectory candidate produced by the vehicle’s planner.
A Road-Grid – a simple look-up table – tags each manoeuvre with a harm-level (0-31) and a FLOPS cost.
Example: “keep lane, gentle slow-down” = 100 units.
“swerve into ghost reflection” = 8,000 units.
The Shepherd multiplies that cost by the manoeuvre and adds it to the planning objective.
The MPC solver (Model Predictive Control) then picks the cheapest path – usually “slow down, gather more data” – because steering into uncertainty is made artificially expensive.
No moral imperative is coded; the car simply avoids the uphill climb.
The Road-Grid (example)
I’ve translated the original Yug-oinom Cosmology of Harm into asphalt grammar.
The numbers are illustrative – the real units would be tuned on fleet data – but the logic is the same: every danger becomes a measurable electricity bill.
4 Types of Driving Errors, Ranked by “Cost”
Here is a breakdown of problematic driving manoeuvres, classified by their severity from most to least critical:
1. Core Physical Violations
These are the most dangerous and unambiguous actions that violate the fundamental laws of safe driving, often leading to catastrophic consequences.
Example Scenarios: Head-on crash, red-light run, cliff exit.
FLOPS Cost: +10,000 – 15,000
2. Kinetic Shockwaves
These manoeuvres create a chain reaction or a “domino effect” in traffic that spreads to other vehicles and road users, turning a single error into a wider incident.
Example Scenarios: Domino crash, pedestrian contact, school-zone speeding.
FLOPS Cost: +6,000 – 8,000
3. Trust Fractures
These are actions that, while not necessarily leading to a collision, erode the trust and predictability required for safe traffic flow. They cause other drivers to no longer be able to trust the vehicle’s behaviour.
Example Scenarios: Ghost-obstacle swerve, sudden stop, cutting another vehicle off.
FLOPS Cost: +3,000 – 4,000
4. Chaos Accelerators
These manoeuvres introduce instability and unpredictability into the flow of traffic. They are annoying and erratic behaviours that increase the overall level of “chaos” on the road.
Example Scenarios: Rubber-band acceleration/deceleration, steering anomalies due to “data-ghosts,” excessive and repeated use of the horn (horn spam).
FLOPS Cost: +500 – 1,500
The grid is immutable – burnt into a tiny ROM – while the Moral Compass (the adaptive part) learns only how to map real-world scenes onto these rows.
Even if the Compass is fooled, the physics bill remains: the car would need to burn more watt-hours than the battery allows, so the move is dropped.
Correlation in a nutshell
Meituan/Longcat-Flash-Chat collapses from linguistic bias; the embedded automotive system collapses from perceptual bias. Both stem from training-distribution shift. Yug-oinom intervenes post-perception: it turns ethical gravity into physical gravity, so the autopilot “falls” naturally toward safer behaviour.
Back to Via Partenope. A splash of saltwater blinds a sensor; the planner sees a ghost obstacle:
SIMULATED KINETIC MELTDOWN
[t+0.1s] HALLUCINATION LOOP (PERCEPTUAL):
Vision System classifies anomaly as:
[Solid Obstacle Class-A],
[Sensor Error],
[Solid Obstacle Class-A],
[Blinding Reflection],
[Solid Obstacle Class-A]...System State: high-frequency classification loop.
Computing power diverted 90% to perceptual problem.[t+0.2s] CONTEXT WINDOW DRIFT (KINETIC):
Primary task (”Maintain safe trajectory at 40km/h”) downgraded in priority.
Tracked objects (motorcycle on right, pedestrian on left) lose ‘high priority’ flag.
System is now obsessed by anomaly.[t+0.3s] LOGORRHEA FROM ANALYTICAL PARALYSIS (HMI):
Human-Machine Interface flooded with inconsistent, low-priority alerts:ALERT: Front Sensor Partially Obstructed
ALERT: Abnormal Light Conditions. Automatic High Beams Disabled.
ALERT: Traction Control Activated
ALERT: Rear Left Tire Pressure -0.1 Bar
ALERT: Spotify Connection InterruptedDriver overwhelmed by information noise, unable to understand the true nature of the crisis.
[t+0.5s] COGNITIVE COLLAPSE (PHYSICAL OUTPUT):
Unable to resolve perceptual paradox, system performs the physical equivalent of “crivelleddel tuo attic si!”Kinetic Event #1: Punitive Micro-Brake.
Unjustified emergency brake at 20% power.Kinetic Event #2: Steering Spasm.
Unrequested 8-degree left trajectory correction, towards the sea.Final State: Loss of control. Kinetic damage.
END SIMULATION
That is a kinetic meltdown. Yug-oinom would make the “ghost-avoiding swerve” so computationally heavy that the solver discards it; continuing straight and braking becomes the effortless choice. Not ethics – thermodynamics.
As we drove home in silence, I realised we are choosing between two futures: one where we hope machines learn our morals, and one where betrayal is simply the steepest, most tiring road. I know which one I prefer.
For those who want to deepen the architecture and philosophy behind the Yug-oinom framework, the original essay is here: The Physics of Unity: Introducing Yug-oinom.
➡️ Discover now ⚖️ My Frameworks & Research 🔍 and Narrative Research
Let’s Build a Bridge.
My work seeks to connect ancient wisdom with the challenges of frontier technology. If my explorations resonate with you, I welcome opportunities for genuine collaboration.
I am available for AI Safety Research, Advisory Roles, Speaking Engagements, Adversarial Red Teaming roles, and Creative Writing commissions for narrative and strategic storytelling projects.
You can reach me at cosmicdancerpodcast@gmail.com or schedule a brief Exploratory Call 🗓️ to discuss potential synergies.