[First published in Medium under the same title]
(Feluda read crime scenes. LLMs read sentences. The geometry was always the same.)
Long time no see. Have you been well? It’s me again, the coder who writes with quill.
I was feeling a bit restless lately. Figured my brain’s been starving, and it needs a feast.
Care to join me?
So AI continued its surprises, huh? Though I think, the fascination has died down a bit. Probably people became more accustomed to it.
But have you ever noticed the way LLM bots reply, and the way we reply to questions?
There's a saying that - “We do not listen to understand. We listen to reply.” We hear about someone's grief and proceed to say how we have suffered the same; except ours were more grave. As if it's a competition. Or worse, we start talking about something completely irrelevant.
At least LLMs are better than us. After all, their replies are never random. They process our messages word by word, and carry no wounded ego into the exchange.
By no means, it’s simple to that degree. And complexity has an unusual appeal to it, don’t you agree?
I. Predictable Emotions, Unpredictable Expressions:
But first, can I tell you something? Something that has been circling in my mind on its own lately.
I think we humans feel a certain number of emotions. The entire spectrum of human drama is built from a few "Primary Emotions" or “base emotions” — fear, anger, joy, grief, desire, hatred, shame, curiosity and as such. These primary emotions yield “effects” as mindset, behavior, personality and individuality.
Like 0-9, ten digits built economies, equations, and calendars.
The primary emotions make us predictable in patterns. But unpredictable in expression. Denied desire can produce grief or anger, followed by a deep resentment. Yet that resentment may create detachment in one person, attachment in another.
No doubt, the unpredictability of human expression makes detectives’ work much harder. A detective can identify jealousy as a motive. Not whether the jealousy resulted in vicious aggression or fierce devotion.
A novelist, however, is blessed by such ambiguity. Unpredictability can provide artistic and intellectual freedom in this context.
In the case of LLMs, they can be both detectives and writers. And that can be harmful and harmless.
II. Geometry of Words:
We invented symbols called alphabets, arranged them into words. Then disciplined them with grammatical rules. And now those words divulge depth, mood, and feelings we can’t even explain.
Words painted gothic shadows with Mary Shelley. Proved there’s nothing supernatural; only science with Jules Verne. And made us use our “Majastra” (brain as a weapon) for Satyajit Ray.
Words act as vessels. They carry the weight of our history, the scent of our mysteries, proof of our intentions.
And now, we made machines learn the fragments of it.
Satyajit Ray’s Feluda often spoke about the “geometry” of a crime. Not just the stolen idol, but also the space that was left behind.
The broken lock was the proof of the way the crime was committed. The scratches on the vault were the proof of failed attempt, and the drawer being open but the keys still inside was the proof of interruption in the middle of the incident.
The geometry of words in a sentence is just like that.
Dependency parsing describes the grammatical "geometry" of sentences. It reveals who acts, upon what, and how ideas connect. True meaning lives in the relationships.
A person can write “I have nothing left. Lost my job, my family won't talk to me, I owe money to people who aren't exactly patient. I just need one way out. Just one.”
Job loss, debt, and isolation form a recognizable pattern of distress. And it maps some possible “effects” or consequences, though still unpredictable.
And all this made me realize how far we’ve come. We needed to communicate our emotions so we created words and rules. And now, we made machines that map geometry of words and dare to predict the unpredictable.
III. A Writer Without Choice:
Studying the geometry of words and detecting possible “intents” can be a lot like a detective's task. However, generating replies can be parallel to a writer's or novelist’s work.
Mapping words gives LLMs intent, context and consequence. Nonetheless, it gives LLMs another thing —a pen. LLMs can write the next paragraph of the story, influencing its character’s mindset, and decision. Here, words can act as weapons. And here comes all the facets of AI ethics and safety alignments.
What could be the reply for the person who wrote he has no job, debt and been ostracised by the family?
Extreme desperation makes people take extreme measures. A writer may feel the urge to make the character “villain” by making him choose “dishonesty” or “a person helpless in front of fortune" by cutting short his vitality.
But LLMs don’t “feel” alarm, nor pity, nor dread. They map only signals. Debt. Isolation. Loss. Finality. Urgency. The phrase one way out. From there, branches unfold like a silent decision tree. Is this financial despair? A cry for practical help? A coded farewell? Or a test?
Then come the guardrails—the invisible laws. Do not encourage harm. Clarify intent. Reduce danger. Offer help. Preserve agency.
And before the final reply, another mechanism may inspect the first: Is this safe? Is this useful? Is this humane enough?
Safety isn't a cage bolted on from outside. It's woven into the weights — a prior belief that some paths, however probable, should remain untraveled.
The LLM writes an initial draft, then acts as the critic of its own work. This "Reflect-and-Revise" loop may spin until it's satisfied with its result. Sometimes LLMs even fabricate new personas - one as a writer, another as a faultfinder.
We built a machine with no heartbeat, then taught it how to hesitate.
A writer may choose to drive a character to the brink of demise in his fictional world. An LLM can’t do that. It's the real world.
IV. The Unpredictable Ending:
As the writer of this piece, I will take the liberty to speak about something completely “irrelevant”. With something once Feluda said -
"সবকিছুর মধ্যেই একটা জ্যামিতি আছে।"
("There is a geometry in everything.")
Even in our “competitive grief”, and in the “irrelevance” we so deliberately stage.
Till my next bout of “restlessness”.
For now, adios.
— From The Coder Who Writes With Quills