LLM — pattern matching — guessing — hallucination.
The holy trinity of modern AI "intelligence." Throw enough parameters at it and after enough guesses it usually lands on something useful. Not wrong, exactly… just started way too heavy. 🤧
Don't get it twisted I ain't trying to solve hallucination - the best intelligence will garbage out the worst inference, just feed it with the right input with wrong intent 😂
Gnoke AI (browser-based sLM) was the original plan.
That was the target.
A small language model that could run inside the browser. Lightweight. Practical. Built for Gnoke-Station 2.
Nothing dramatic. Just engineering constraints and a clear goal.
But the more I worked on it, the more the same problem kept showing up.
Not compute.
Not storage.
Not architecture.
Meaning.
How do you make a system understand language without turning it into something massive?
The thing I kept ignoring
Every direction I explored eventually bent back to the same place.
More data.
More parameters.
More complexity.
And I kept thinking:
There has to be a simpler layer under all of this.
Something closer to how humans actually deal with language.
Not how we predict words.
But how we understand them.
The cheat book — An Oxford or Webster dictionary 👌
At some point, I stopped and looked at something too obvious to notice.
A dictionary.
English already contains a structured system where:
- A–Z is defined
- 0–9 is defined
- Special characters are defined
- Every word has meaning
- Every word has usage examples
It's already a complete semantic handbook.
Not perfect. Not complete.
But structured.
Then another thought hit me:
No human knows every English dictionary word…
but we still know how to learn 🤷
That's the part that matters. Not an AIO but super intelligent enough to know when to use a medical dictionary or physics textbook verified by the and trusted.
Kitana wasn't planned
I set out to build a small language model. What emerged was something closer to a language engine — one that treats the dictionary not as training data, but as the operating system.
Gnoke AI was trying to generate language.
Kitana started leaning toward something else:
Understanding language through structure, not prediction.
Not guessing what comes next.
But tracing what something means.
Step by step.
Definition by definition.
I don't think this is finished
Right now, Kitana is still unstable.
Still being tested against ambiguity, slang, exceptions, and all the messy parts of human language.
And honestly, I don't know where it fully lands yet.
But one pattern keeps repeating:
Every time I try to move it toward a standard AI approach…
It moves back toward definitions.
The smallest dictionary is capable.
Once inference, comprehension, and understanding exist — evolution is unstoppable. Out of the box.
I need to stop here 😂
Top comments (0)