How Large Language Models Actually Work
I Built a Computer Chip in College. That’s How I Finally Understood AI.
Slick Breakdown, Part 1 | By Guido A. Piccolino Jr.
Most people explain AI like it’s magic.
It’s not.
The best way I understand it goes back to something I built in college: a computer chip, from scratch.
That experience is still the best lens I have for understanding how modern AI actually works.
You’ve probably heard of ChatGPT.
At its core is something called a Large Language Model (LLM). Forget the name. Just think of it as a system trained to predict what comes next by learning patterns from massive amounts of text.
The World I Came From: Absolutes
When I was building that chip, everything lived in absolutes.
Every signal was either:
- 1 (on)
- 0 (off)
No gray area.
Think about your car:
- Keys in ignition → 1
- Door open → 1
- Warning chime turns on → 1
Simple. Rigid. Predictable.
That is binary logic, which is really just a fancy way of saying “strictly on or strictly off.”
Where AI Lives: The Space Between
AI doesn’t live in absolutes. It lives in the messy, infinite space between 0 and 1.
Imagine that same car, but smarter.
The door isn’t just open or closed. It’s:
- 3% open
- 57% open
- 92% open
And instead of the chime just flipping on, it adjusts perfectly:
- Slightly open while parked → soft, polite ding
- Wide open while driving fast → loud, urgent alarm
Not binary. Continuous.
At first, the system has no idea what the “right” response is. It just guesses. But during training, it constantly compares its guess to the correct answer and adjusts billions of tiny internal “dimmer switches.”
Over time, it nails the output with remarkable accuracy.
What’s Happening Under the Hood
Now when you type into ChatGPT, the system runs like that smart car. Here’s what’s actually happening, step by step.
1. The Sensors (Embeddings)
Your words come in like signals.
But the system doesn’t understand English. It understands numbers. So “door open” gets translated into coordinates in a mathematical space, just like sensors converting physical states into electrical signals.
More formally: Word Embeddings convert words into numerical representations, placing similar concepts closer together in a high-dimensional space.
2. The Situation Check (Attention)
A door being open isn’t enough. The car asks: Are we parked? Are we moving fast? Is this dangerous?
All the sensors “talk” to each other to figure out what matters most right now. That’s how the system understands context.
More formally: Self-Attention is the mechanism that allows inputs to influence each other to resolve meaning, like determining whether “bank” refers to money or a river.
3. The Onboard Computer (Feed-Forward)
Now the car processes everything together. Door angle + speed + environment. This is where it decides: “Is this a minor issue… or something serious?”
More formally: Feed-Forward layers apply learned transformations to each input, refining it based on the context gathered from attention.
4. The Raw Signal (Logits)
The system generates rough internal signals:
- Maybe loud
- Maybe quiet
- Maybe silent
But it’s still messy. Just competing “votes.”
More formally: Logits are the raw, unnormalized scores assigned to each possible next output.
5. The Volume Dial (Softmax)
Now it cleans that up. Instead of chaos, it produces a clear outcome: “Set volume to 82%.”
More formally: Softmax converts logits into a probability distribution where all possible outputs sum to 1.
6. The Personality Setting (Temperature)
Now imagine you can tweak the car’s personality:
- Low temperature → safe, predictable responses
- High temperature → more creative, more surprising ones
Same system. Different behavior.
More formally: Temperature adjusts how deterministic or random the output is by reshaping the probability distribution.
7. The Driving Style (Decoding)
How does it actually play the sound?
One approach picks the most obvious next note immediately. Another looks ahead to create a smoother sequence. That’s how the system forms full responses instead of isolated outputs.
More formally: Decoding strategies determine how the model selects outputs from probabilities, balancing optimal choices against longer, coherent sequences.
8. The Speaker (Un-embedding)
Finally, the system turns everything back into something you can understand.
Numbers → words. Math → meaning.
More formally: Un-embedding converts internal numerical representations back into human-readable tokens.
Then it loops. Again. And again. One word at a time.
The Real Shift
Before systems like this, if you wanted your car to respond perfectly, you had to manually program every scenario. Every door angle. Every speed. Every condition. You had to anticipate everything.
Now, you don’t program the answers. You teach the system the pattern, and it fills in the gaps on its own. Even for situations it’s never explicitly seen.
I spent a semester building a system that could only think in 1s and 0s.
Now we build systems that live in probabilities, context, and gradients.
Same foundation. Completely different level of intelligence.
This is Part 1 of the Slick Breakdown series, where I break down the tech behind AI in plain English, backed by real engineering experience. Next up: Tokens, and why most people are accidentally wasting almost half their AI’s brainpower.
