What Is AI? — Interactive Guided Visual Flow

CLICK ANY BOX BELOW TO EXPLORE Matryoshka Model

1. Computer Science Field

2. Artificial Intelligence (AI) Selected Layer

3. Machine Learning (ML) Method

4. Deep Learning (DL) Technique

5. LLMs & Generative AI

Click inside any color box to see what it means Nested Breakdown

CONCEPT BREAKDOWN Layer 2 of 5

Artificial Intelligence (AI)

Goal: Systems performing human judgment tasks

What it actually is

A field of computer science building systems that perform tasks requiring judgment, perception, or reasoning.

Crucial Difference

Traditional software executes rules hand-written by a programmer. AI learns its own internal rules (parameters) from examples.

Step 1: The Hierarchy

STEP 2 OF 5

The AI Data Pipeline Flow

Follow raw data from initial examples to final output.

Node 1 / 6

1. Data

Examples

2. Vectors

Embeddings

3. Training

Parameters

4. Model

Frozen Base

5. Inference

Run Output

6. App & Feedback

Prediction

PIPELINE STAGE 1

Data & Training Examples

Dataset = Raw Training Examples

Data quality and representative coverage bound model capabilities. Raw text, pixels, or audio examples are assembled into datasets.

STEP 3 OF 5

Training vs. Inference Divergence

Parameters Change vs. Parameters Locked

Training Phase

OFFLINE / EXPENSIVE

Weights Updating

Gradient Optimization Loss: 0.812

Heavy Compute: Runs over weeks on thousands of expensive GPUs.
Changes Model: Internal parameter weights continuously adjust to fit patterns.

One-time version creation phase

Inference Phase

ONLINE / CHEAP

Weights Frozen

Forward Pass Execution Latency: 14ms

User Input Frozen Model Output

Fast & Scalable: Repeated calls for millions of users running on static weights.
No Permanent Memory: Chat memory passes conversation history into the context window, not parameter updates.

Repeated runtime deployment phase

NEXT-TOKEN PROBABILITY ENGINE Step-by-step assembly

Input Context Context:

"Artificial intelligence learns statistical patterns from data"

Calculated Vocab Distribution:

FACTUALITY INSIGHT

Why Generated Content Can Be Wrong

Generative AI does not search a fact database. It computes probability distributions over token sequences to find plausible continuations.

Plausibility ≠ Truth

Fluent, confident, and correct are three separate properties. A model can produce a sequence that is completely fluent yet factually false (hallucination).

STEP 5 OF 5

When to Use AI (Architectural Matrix)

Matching capabilities to risk & cost

1. Hand-Coded Rules

Logic is stable, fully known, and must be 100% deterministic & auditable.

e.g., Tax formulas, even/odd math

2. Exact Database

Exact facts must be retrieved without risk of plausible hallucination.

e.g., Bank balances, user accounts

3. Pure AI System

Patterns are too complex for human rules; error rate is cheap or tolerable.

e.g., Spam filters, language translation

4. Hybrid Loop (Best)

AI drafts, ranks, or flags $\rightarrow$ Human expert reviews and takes accountability.

e.g., Legal drafting, medical triage

Congratulations! You've Completed the 5-Step AI Tour

You now possess a precise, hype-free working mental model of Artificial Intelligence.

WHAT IS AI? GUIDED VISUAL TOUR

START HERE: What are you looking at?