workshop//01
01 / 18
A beginner's guide to AI and vibe coding

The World
of AI

What it is. Who makes it. How to use it.

scroll to explore
The Foundation

It all starts with Machine Learning.

Classical software is explicit: a programmer encodes rules. ML inverts this. You define an objective, feed in data, and an optimization algorithm — gradient descent — adjusts millions of internal parameters (weights) until the model minimizes its error.

No rules are written. Behavior emerges from optimization. The same core method — define a loss function, backpropagate gradients, update weights — underlies every major AI system in use today.

How Training Works
01 — Data
Feed the model massive amounts of labeled or unlabeled examples — text, images, code, sensor data, game states.
02 — Objective
Define what "good" looks like via a loss function — predict the next token, reconstruct a clean image from noise, maximize a game score.
03 — Optimize
Gradient descent iteratively updates billions of weights to minimize error. Run this long enough at sufficient scale and complex capabilities emerge.
The Two Breakthroughs

ML existed for decades.
These two things changed everything.

Modern AI didn't come from one invention. It came from two distinct advances — one in architecture, one in alignment — that together turned raw prediction into something that feels like intelligence.

2017 — Google

Transformer
Architecture

Before the Transformer, AI read text sequentially — word by word, left to right. The breakthrough was attention: letting the model look at all words simultaneously and learn which ones matter most to each other.

In the sentence "The animal didn't cross the street because it was too tired" — attention is what tells the model that "it" refers to "animal," not "street." That ability to track relationships across context is what makes LLMs feel like they understand.

Vaswani et al., "Attention Is All You Need," 2017
2022 — OpenAI

RLHF

A trained model is technically just a very good text predictor. It doesn't yet know how to be helpful or safe. Reinforcement Learning from Human Feedback is the fine-tuning step that shapes it into an assistant.

Human raters compare pairs of model responses and pick the better one. Those preferences become a reward signal, and the model learns to produce outputs humans rate highly. This is what turns a raw predictor into something that follows instructions, avoids harm, and feels like it has a personality.

Why Claude feels different from GPT-4 — same architecture, different RLHF.
04
One Method, Many Architectures

ML is the umbrella.
The models beneath it vary wildly.

The same optimization principle underlies completely different model families — each with a different architecture, training signal, and class of problems it solves well.

Most AI tools you use today combine more than one. ChatGPT and Claude are LLMs fine-tuned with RL. DALL-E is a diffusion model guided by a text encoder. Sora combines both.

Model Families
Large Language Models (LLMs)
Transformer-based. Trained to predict the next token across text, code, and data. Claude, GPT-4, Gemini.
Diffusion Models
Trained to reverse a noise process. Start with static, iteratively denoise into an image or video. Stable Diffusion, Midjourney, Sora.
Reinforcement Learning (RL)
Trained via reward signals from an environment. No labeled data — the model discovers strategies by trial and error. AlphaGo, robotics, RLHF fine-tuning.
Vision Models (CNNs / ViTs)
Trained on images for classification, detection, segmentation. CNNs dominated for a decade; Vision Transformers now lead on benchmarks.
Multimodal Models
Combine architectures to handle text, images, audio, and video together. GPT-4o, Gemini 1.5, Claude 3 — the current frontier.
The Landscape

AI isn't a chatbot —
it's everything

The tools you actively choose to use are just the tip of the iceberg.

💬

Text & Chat

Conversation, writing, analysis, research, advice.

ClaudeChatGPTGemini
🎨

Image Generation

Type a description, get a picture. Any style.

MidjourneyDALL·EFirefly
🎬

Video Generation

Text or images turned into moving clips.

SoraRunwayPika
🎵

Audio & Music

Full songs, voice cloning, voiceovers.

SunoElevenLabsUdio
💻

Coding

Write, fix, and build entire apps.

Claude CodeCursorCopilot

Productivity

Meetings, emails, files, tasks — automated.

Notion AICopilotOtter.ai
🎮

Gaming

Smarter opponents, generated worlds, lifelike NPCs.

Inworld AINvidia ACE
🏥

Healthcare & Science

Reads scans, folds proteins, discovers drugs.

AlphaFoldTempus
🚗

Autonomous Systems

Self-driving cars, drones, and physical robots.

WaymoTeslaFigure
🔍

Search & Discovery

AI-powered search with real-time cited answers.

PerplexityGoogle AI
🛡

Security & Fraud

Detects threats, flags fraud, protects systems.

DarktraceCrowdStrike
🌍

Translation

Real-time translation, subtitles, multilingual content.

DeepLGoogle Translate
🏦

Finance

Trading, forecasting, risk analysis, and financial research.

Bloomberg AIKenshoVertus
⚖️

Legal

Contract review, case research, and legal workflow automation.

HarveyCasetext
👗

Marketing & Fashion

Ad copy, trend forecasting, and AI-powered styling.

JasperAdCreative.ai
Full Interactive Platform Reference
The bigger picture

AI is already embedded in everything you use.

The tools on the previous slide are the ones you consciously choose to use. But AI has been quietly running in the background of your daily life for years — you just didn't call it that.

What's changed isn't that AI arrived. It's that it became visible, conversational, and creative — and now anyone can use it, not just engineers.

📱
Your phone
Face ID, autocorrect, Siri, photo sorting — all AI
🎵
Spotify & Netflix
Every recommendation is an AI decision
🛒
Amazon & Instagram
What you see first? AI chose that
💳
Your bank
Fraud alerts, credit scores, loan approvals
📧
Your inbox
Spam filters, priority sorting, smart replies
🗺
Google Maps
Traffic prediction, ETA, rerouting — all AI

"AI won't replace you. But the version of you that knows how to use AI will be dramatically more capable than the version that doesn't."

05
Why Now

AI went from
science fiction to
your phone in
about two years.

Three Things Converged

The breakthrough wasn't one invention. It was three.

Compute
GPU clusters at unprecedented scale made training trillion-parameter models economically feasible for the first time.
Data
Decades of digitized human knowledge — books, code, scientific literature, the web — became training data at internet scale.
Architecture
The Transformer (Vaswani et al., 2017) unlocked parallel training and long-range attention — making scale actually work.
100M
ChatGPT users in 2 months
$200B+
invested in AI in 2024
47+
major platforms today
Three Types of AI

Not all AI thinks the same way

Understanding these three types helps you know what any AI tool can — and can't — do for you.

01 — Predictive

Predicts
what comes next

Trained on patterns to forecast an outcome. It doesn't understand — it calculates probabilities.

Everyday examples
📧 Gmail's spam filter
🎬 Netflix recommendations
🏦 Credit card fraud detection
⌨️ Autocomplete on your phone

"Based on what's happened before, here's what's likely to happen next."

02 — Generative

Creates brand
new content

Given a prompt, it generates something that didn't exist before — text, images, music, video, code.

Everyday examples
💬 Claude, ChatGPT, Gemini
🎨 Midjourney, DALL·E
🎵 Suno, ElevenLabs
🎬 Sora, Runway

"Tell me what you want, and I'll create something new for you."

03 — Agentic

Takes action
on your behalf

Goes beyond answering — it actually does things. Opens apps, browses the web, writes files, sends emails.

Everyday examples
💻 Claude Code — builds apps
📧 AI that reads & replies to email
🗓 Books your calendar for you
🛒 Orders groceries automatically

"Don't just tell me — let me go handle it for you."

💡

Most tools today combine all three. Claude is generative when you chat with it, and agentic when you use Claude Code. The lines are blurring fast.

A New Way to Build

Vibe
Coding

Coined by Andrej Karpathy — one of the founders of OpenAI — in early 2025 to describe a new way of building software that anyone can do.

Instead of writing code yourself, you simply describe what you want in plain English — and the AI builds it. You don't need to know syntax, frameworks, or computer science. You just need to know what you want.

"You describe it. Claude builds it. You take the credit."

The old way

Learn Python. Study HTML. Understand databases. Take courses for months. Write hundreds of lines of code. Debug errors. Start over.

🤓
Vibe coding

Open Claude Code. Type "Build me a web app that tracks my family's weekly schedule and sends reminders." Watch it appear.

💡

The skill isn't coding anymore — it's knowing what to ask for and recognizing when it's right. That's a skill you already have.

Claude Code

You don't need a terminal. You don't need VS Code.

Claude Code used to be a tool only developers could use — living in the command line, requiring installs and technical setup. That's changed.

Anthropic now offers a Desktop App that gives you the full power of Claude Code through a regular window — just download and open it like any other app on your computer.

The terminal and the desktop app are just two different doors into the same room. Same AI, same power — the desktop app is just the friendlier entrance.

Getting Started

Four steps. No tech background needed.

1

Go to claude.ai

Create a free account with your email address.

2

Upgrade to Claude Pro

$20/month — this unlocks Claude Code. The free plan doesn't include it.

3

Download the Desktop App

Available for Mac and Windows. Install it like any normal app.

4

Open it and start talking

Describe what you want to build in plain English. That's it.

No terminal. No VS Code. No coding experience required. Just a conversation.

Under the Hood

The three Claude models

Whether you use Claude in the browser, the desktop app, or the terminal, you're talking to one of three underlying AI models. Think of them as different settings on the same engine.

Most Powerful

Opus 4.6

The deepest thinker. Best for complex reasoning, nuanced writing, and hard problems where quality matters more than speed.

Complex analysisResearchLong documents
Best Balance

Sonnet 4.6

The everyday workhorse. Fast, smart, and capable — handles the vast majority of tasks beautifully. What you'll use most.

CodingWritingDaily tasks
Fastest

Haiku 4.5

Lightning quick for simple tasks. Great when you need a fast answer, a short summary, or a quick rewrite.

Quick answersSummariesSpeed

These same three models power every way you access Claude — the website, the mobile app, the desktop app, and the terminal. The interface changes; the intelligence underneath doesn't.

Opus 4.6, released February 5th 2026, is state-of-the-art on several evaluations including agentic coding, multi-discipline reasoning, knowledge work, and agentic search.

The Skill That Replaced Prompt Engineering

Just talk to it
like a person.

For a brief moment, "prompt engineering" was a career. People sold courses. Companies posted job listings. There were frameworks, formulas, tricks.

Then the models got smarter — and all of that became unnecessary.

Then

Memorize syntax. Use the right incantations. Format your prompt correctly or it won't work.

Now

Be clear. Give context. Say what you actually want. Push back when it's wrong. That's it.

The best interface for AI is the one
you've been using your whole life —
plain English.

Web Apps

Build in a browser.
No setup required.

Web apps run in a browser — on any phone, laptop, or tablet. Claude Code's Desktop App can build these with zero technical setup on your end.

You can even make them feel like a real app on your phone — add to your home screen, works offline — these are called Progressive Web Apps (PWAs). No App Store, no Xcode, no $99/year developer fee.

No Xcode
No terminal or VS Code
Works on any device
Free to publish
Can add to home screen as a PWA
Native iOS Apps

The App Store path is a whole different adventure.

Native iOS apps — the kind that live in the App Store — require a separate set of tools that even Claude Code can't skip.

🖥 Xcode — Mac only
Apple's development tool. Required to build and compile iOS apps. Does not run on Windows.
🍎 You must own a Mac
No way around this. iOS development is locked to Apple hardware.
💳 Apple Developer Account — $99/year
Required to publish anything to the App Store.
⌨️ Terminal time — unavoidable
Even with Claude Code helping, you'll end up in the command line at some point.

A Note on the Design Process

Good things get built
in loops, not lines.

✏️
Design
Describe what you want in plain English
Prototype
AI builds a working version in seconds
👀
Test
Use it. Break it. Notice what's missing
🔁
Iterate
Tell AI what to fix. Repeat until it's right

The calendar app you just saw wasn't built in one prompt — it was shaped over many. Each round of feedback made it sharper. That back-and-forth is the skill.

Workshop

Let's build
something real.

In the next 30 minutes, we're going to build a real, working app — together, using Claude Code. No coding experience needed. You just need to know how to talk.

30 min
Build time
💬
0 lines
Code you'll write
1 real app
You'll walk away with
The App

The Brearley
Smart Calendar

The school calendar PDF has everything — but it's a wall of text. This app turns it into a filterable, color-coded calendar so you can see exactly what matters to your daughter's grade.

No more scrolling through dates that don't apply. Filter by grade, filter by type, and see only what's relevant — at a glance.

Source: Brearley School Calendar PDF ↗

What it does

One calendar.
Filtered your way.

🎓
Filter by Grade
See only events relevant to your child's class
🔴
Color-coded by type
Red = days off · Blue = events · Green = deadlines
📅
Month view or list view
Toggle between a calendar grid and a clean list
📲
Works on your phone
Add to home screen as a PWA — no App Store needed
🔔
Coming soon: reminders
Get notified the day before anything important
Workshop Steps

What you'll say to
Claude Code

Step 1 — 3 min
Feed it the PDF
"Here is the Brearley school calendar PDF. Parse all the events, dates, grades, and types into structured data."
Step 2 — 5 min
Build the calendar
"Create a web app that displays these events as a calendar. Clean design, easy to read on a phone."
Step 3 — 8 min
Add filters
"Add a filter by grade so I only see events for my daughter's class. Add a second filter by event type — days off, school events, deadlines."
Step 4 — 5 min
Color code it
"Color code the events: red for days off, blue for school events, green for deadlines. Make it look polished enough to share with other Brearley parents."
💡

This is a real app you'll actually use. When the audience sees something built from their own school's data, it stops feeling like a demo and starts feeling like a superpower.

You already have
everything
it takes.
Now you have the tools.

Curiosity, taste, and knowing what you need — that's the real skill. AI just gives you a way to act on it, faster and bigger than ever before.

ReginaFloresMir.ai