CITYPOPY · DISTRICT 04

Welcome to The LLM Creation Factory

Discover how AI brains are trained. Explore giant neural reactors, data rivers, token storms and prediction simulators powering AI City.

01

Collecting Internet Knowledge

Drones lift books, websites, code and conversations into the factory's intake tower.

Real-Life Scenario

A customer support model learns from past tickets, help-center docs, and chat logs before it can answer users accurately.

Tap a source. Drones lift it into the factory.

Total tokens collected: 0 M
Knowledge Intake Tower

LLMs learn from enormous collections of text — books, websites, code, and conversations.

02

Tokenization Machine

The Token Slicer 9000 chops your text into tiny pieces the model can count.

Real-Life Scenario

When you type 'Can I get a refund?', the model does not read it as one chunk; it processes token pieces to understand intent.

Before AI learns language, text is sliced into smaller pieces called tokens.

Token Slicer 9000
ThefutureofAIisbright
8 tokens · token IDs would map each to a number the model understands.
03

Pattern Learning Reactor

Repetition strengthens neural pathways. The reactor glows brighter as patterns repeat.

Real-Life Scenario

In e-commerce support, the model sees thousands of pairs like 'Where is my order?' and replies that include tracking updates. After enough examples, it learns this relationship and answers shipment questions more reliably.

Feed repetitive examples. Watch neural pathways strengthen.

Pattern strength: 20%
Neural Reactor Core

AI learns statistical relationships between tokens — not human understanding.

04

Prediction Training Simulator

Guess the next token. Billions of times. That's basically how training works.

Real-Life Scenario

Autocomplete in email works this way: it predicts likely next words from patterns in language, then ranks the best option.

AI predicts the next token:
The sky is ___
blue78%
green14%
pizza8%

Next-token prediction

During training, the model is shown billions of half-sentences and asked to guess what comes next. Confident, correct guesses earn reward; wrong guesses are corrected.

Repeat this billions of times and the model becomes scary good at the next-word game — which, it turns out, is most of what language is.

05

Loss & Correction Center

Wrong answer? Correction beam fires. The model gently adjusts itself.

Real-Life Scenario

If a coding assistant suggests a broken snippet, feedback loops nudge the model so future suggestions are more syntactically correct.

Correction Beam Console
“The sky is …” → guess: pizza
target: blue · loss: 1.00

Mistake → adjust → improve

Each wrong prediction sends a correction signal back through the network, gently nudging its connections so the same mistake is less likely next time.

Imagine a million tiny dials, each turned a hair’s width per example. After billions of examples, those dials encode language itself.

06

Neural Weight Factory

Weights are the model's long-term memory of every pattern it ever saw.

Real-Life Scenario

Recommendation systems store learned preferences in weights, helping decide what products or videos users are most likely to click next.

Slide to strengthen learned connections.

Avg weight strength: 50%

Weights are the millions/billions of numbers a model learns. They are the memory of every pattern it has ever seen.

Weight Storage Grid
07

Fine-Tuning Lab

Take a base brain and specialise it: travel guide, coder, storyteller, doctor.

Real-Life Scenario

A hospital fine-tunes a base model on clinical writing style so summaries use medically accurate terms and safer phrasing.

Pick a specialty pack:

Fine-tuning takes a base model and teaches it the style and knowledge of a narrower domain.

Specialist AI Output
Pack light, book early, and chase sunsets.
08

Inference Engine

Trained model meets real prompt — output streams token by token.

Real-Life Scenario

When a user asks a chatbot for a trip plan, inference is the live moment where the model turns prompt + learned patterns into a response.

A trained LLM generates output one token at a time, each chosen using the patterns it learned during training.

Inference Stream
// awaiting prompt…
09

Hallucination Zone

When patterns are weak, the model can confidently invent things. Stay curious!

Real-Life Scenario

A legal assistant may cite a non-existent case when uncertain, which is why high-stakes workflows must verify every factual claim.

Ask the AI:

What is the capital of France?
Confidence Cloud
awaiting prediction…
10

AI Scaling Chamber

More data, more compute, bigger model — the brain grows and grows.

Real-Life Scenario

As traffic grows from 1,000 to 1,000,000 users, teams scale GPUs, optimize latency, and tune costs to keep responses fast and affordable.

Training Data30%
Compute Power30%
Model Size30%
Estimated capability: 30%
11

Final Training Mission

Make the right calls. Light up AI City.

Real-Life Scenario

Shipping a real AI product means balancing accuracy, safety, cost, and reliability before launch day, not after incidents happen.

Train AI City’s New Mega Brain

🧠

Mission in progress…

Complete every step to power up AI City’s new mega brain.

You now understand how LLMs are trained.

Data → tokens → predictions → corrections → weights → fine-tuning → inference.

Unlock AI City Pro Membership

Unlock all 13 interactive districts with a yearly subscription.

Razorpay · UPI, Card, Netbanking