/ NOTEBOOK 01 · FOUNDATIONS

Let’s open the AI toolbox.

Let’s look inside, piece by piece. No need to memorize names: just understand which question each technology answers.

/ HOW IT LEARNS

Three ways to learn from data.

Imagine we want to detect a fault. We can show labeled examples, let the system find clusters, or let it try actions and receive a signal. From there come these three paths.

Supervised

Learns from examples with a known answer: fraud/not fraud, price, category, or demand.

Classification · regression

Unsupervised

Searches for structure without prior labels: groups, similarities, topics, or outlier behaviors.

Clustering · reduction

Per reinforcement

Learns through actions, rewards, and consequences within an environment.

Control · planning

/ VISUAL ARCHIVE · OPEN SCIENCE

Intelligence begins by observing the world.

Real scales, signals, and patterns before turning them into data.

Electron micrograph of a geological sample with rounded structures — Matter under the microscopeNASA Images

Two satellite views of Hurricane Juliette and the height of its clouds — Atmosphere turned into dataNASA Images

Transmission spectrum chart of exoplanet HAT-P-11b — Data, models, and spectraNASA Images

Bright magnetic loops over the Sun’s surface — Solar magnetic fieldsNASA Images

/ WHAT IT CAN DO

Now let’s look at what it can do.

The learning method explains the training. To choose a real‑world tool, it’s best to ask for the capability we need.

PERCEPTION

Turn signals into information

Vision, audio, sensors, and documents.

PREDICTION

Estimate likely results

Demand, risk, maintenance, and anomalies.

GENERATION

Create conditioned content

Text, code, image, voice, and video.

DECISION

Compare options under rules

Optimization, planning, and recommendation.

ACTION

Use tools in an environment

Agents, robots, and automations.

/ GENERATIVE AI

We’ve reached the tools that create.

A foundation model learns general relationships from lots of data. Then we can steer it with an instruction, examples, or our own documents. That’s where acronyms that sound hard appear, but they describe fairly concrete pieces.

LLM

Large-scale language models. They work with tokens and excel at text, code, and structured reasoning.

VLM

Vision-language models. They relate images to language to describe, compare, or answer about a scene.

Outreach

Generative family that learns to reconstruct signals from noise; widely used in image, audio, and video.

Embeddings

Numeric representations that place similar content close together and enable semantic search.

/ AGENTIC SYSTEMS

What changes when AI can act?

Until now the tool interpreted or generated. An agent adds a goal and access to tools: it can look up a datum, prepare a draft, or execute a step. Its autonomy isn’t magic; it depends on the permissions we grant.

01ObjectiveWhat outcome it seeks

02ContextWhat it knows and remembers

03ToolsWhat it can view or change

04PolicyWhat limits it must respect

05VerificationHow it proves the result

Agent governance answers who is responsible, what data and permissions the system uses, when it must stop, and how it proves the result.

/ SELECTION CRITERIA

We close the notebook with a decision.

We already know the pieces. To choose, we go back to the starting point: what we want to improve, what information we have, and what could go wrong.

01
Define the result
Time saved, error reduced, revenue or quality.
02
Govern the data
Owners, availability, quality, sensitivity, lineage, and usage rights.
03
Limit the risk
Impact of an incorrect response or an unauthorized action.
04
Compare alternatives
Rules, classic software, small models, or generative systems.
05
Test, measure, and assign owners
Evaluation, oversight, and stop criteria before scaling.

NEXT STOPCurrent models and computer vision