Today, we discover

1. Multimodal AI — much more than text
2. How AI sees (pixels → recognition)
3. AI that draws, that listens, that speaks
4. Music and video: current frontiers
5. The models of the world: describing vs anticipating
6. deepfakes: spotting the fake
7. Apps that change lives
8. Activity: create with AI!

Opening question

Have you ever used an AI that doesn’t read text…
but who sees or who listens?

Clues:

• How do you unlock your phone?
• “Hey Siri…”, “OK Google…”
• Shazam, which recognizes a song in 3 seconds
• Automatic subtitles on YouTube

You use multimodal AI every day — often without knowing it.

Multimodal AI

Multi = several · Modal = content types (“modes”)

A multimodal AI can understand and create:
text · images · sound · video

Before: chat with the AI by SMS, in the dark.
Now: she looks at your photos, listens to your voice,
and can respond by drawing.

The “senses” of AI

Human sense	AI equivalent	Everyday example
The view	Computer vision	Face unlock
Hearing	Voice recognition	“Hey Siri…”
The word	Speech synthesis	The GPS that speaks
The drawing	Image generation	DALL-E, Midjourney

Nuance: AI does not “see” like we do.
She transforms everything into numbers and looks for patterns in them. No consciousness — of calculation.

How AI Sees — Step 1: Pixels

A digital image = a grid of small colored squares: the pixels
(picture elements: “image elements”)

Each pixel = 3 numbers: Red, Green, Blue (RGB code), from 0 to 255

 Bright red = R:255  G:0    B:0
 Sky blue   = R:135  G:206  B:235
 White      = R:255  G:255  B:255

For the computer, a photo of a cat = a huge table of numbers. Nothing else.

Demo: Pixel Explorer — let's zoom in to squares!

Step 2: visual cues

The AI learns to spot patterns, from simple to complex:

1. Very simple: contours, lines, spots of color
2. Medium: shapes — an eye, a pointed ear, whiskers
3. Complex: the entire object — a cat!

How to learn to read:
letters → syllables → words → sentences
The AI does the same: strokes → shapes → objects

(In English, these indices are called features, “characteristics”)

Step 3: recognition

After viewing millions of labeled images
(“this is a cat”, “this is a dog”)…

…the AI associates the detected patterns with a response, with a confidence score:

“Cat: 97%” · “Dog: 2%” · “Fox: 1%”

Reminder from previous sessions: this is learning from examples — not magic. And she could be wrong!

Computer vision around you

• Facial unlock — the phone measures the geometry of your face (distance between the eyes, shape of the nose, etc.)
• Autonomous cars — detect pedestrians, signs, lines... dozens of times per second
• Medical imaging — spot anomalies on x-rays, sometimes invisible to the eye

Golden rule: the AI assists the doctor.
The diagnosis and decision remain human.

When vision is wrong

The AI recognizes patterns — it does not “understand” the scene.

• A simple sticker on a STOP sign can disrupt it
• A dog in an unusual position → “60% cat?”
• Fog, backlight, weird angle → confidence plummets

Lesson: impressive, but not foolproof.
This is why humans maintain control over important decisions.

AI that draws

You write a description (a prompt) → the AI creates the image.

The big names (it’s evolving quickly!):

• DALL-E (OpenAI) — integrated with ChatGPT
• Midjourney — the artistic rendering
• Stable Diffusion — open source (open code)

“A cat astronaut who plants a flag on the Moon,
cartoon style, bright colors” →

How does she do it? (simple version) →

Many of these AIs work by diffusion:

1. They start from a “fog” of random pixels
2. They “clean” it little by little…
3. …guided by your description…
4. …until the image appears!

Like a Polaroid that reveals itself — but driven by your words.

Known weak points : hands , text in image, fine details.
(Constantly improving!)

AI that listens

Voice recognition: sound → numbers → text

Your voice is a sound wave → more numbers!
The AI ​​learned from thousands of hours of speaking.

Applications:

• Voice assistants: Siri, Alexa, Google Assistant
• Voice dictation (write without keyboard)
• Automatic subtitles — YouTube, video, live TV

A revolution for deaf or hard of hearing people.

AI that speaks

Speech synthesis: the reverse path — text → sound
(in English: text-to-speech, TTS, “from text to speech”)- GPS, reading articles, audio books

• A voice for people who cannot speak
  (like the famous physicist Stephen Hawking)
• Voices increasingly natural: intonations, pauses, emotions

Some tools can clone a voice in seconds
recording… We'll talk about it again very quickly.

Music and video: the borders

Music — Suno, Udio…
A description → a complete song: lyrics, vocals, instruments

Video — Sora, Veo, Runway…
A text → a short realistic video.
Still imperfect: weird physics, objects that appear/disappear…

What is “impossible” today will perhaps be commonplace in a year.
Hence the following question: can we still believe what we see?

Models of the world — the next frontier

Riddle: “I’m going to wash my car, the station is 1 km away. How many km on foot?”

→ Answer: zero! We're going there by car

A text model manipulates words; he has never inhabited a world.

A model of the world = knowing how the world works to anticipate:
if I push this glass, will it fall?

The baby, the cat… and the robot

A baby understands that objects fall before he can speak.
A cat aims at a table without knowing the word “trajectory”.

Our understanding of the world is older than language.

Text: the error is corrected. The physical world: error breaks glass.

Image (space) → Video (time) → Robotics (action)

→ Describe the world ≠ anticipate it. This may be the missing piece of AI.

Why this is a problem

• Disinformation — false statements by personalities, false “evidence”
• Scams — the cloned voice of a loved one who “calls” to ask for money (it really exists!)
• Harassment — false images of real people: illegal and seriously punished
• The worst: the erosion of trust…

When everything can be false, we can also deny what is true:
“This proof? Well, it’s a deepfake!”

The 4 reflexes of the detective

1. The details — hands, teeth, reflections, illegible text, inconsistent shadows (less and less reliable over time!)
2. The movement — strange blinks, lips out of sync
3. The source — who published? A reliable media? Does the information exist elsewhere?
4. Common sense — too good/shocking to be true? Strong emotion = check before sharing

It's up to you: the deepfake detective game!
(session web page — real or AI?)

Doubt ≠ no longer believe anything

The goal is not to become paranoid.

The goal: to become auditor.

Doubt = check before believing
And above all: check before sharing

The #1 skill of the 21st century: knowing how to go back to the source.

Multimodal AI that changes lives Accessibility:

• Real-time subtitles (deaf/hard of hearing people)
• AI audio description: the phone describes out loud what its camera sees (blind/visually impaired people) — read a menu, recognize a ticket
• A synthetic voice for those who cannot speak

Medicine:

• Help detect cancers on x-rays and mammograms
• AI assists, humans decide — always.

Art and creativity

New tools for artists:

• Illustrate a story in just a few minutes
• Compose a song model
• Prototype a movie scene

Debate: an image generated by AI, is it art?
(There is no official answer — defend yours!)

History repeats itself: photography, the synthesizer... have transformed art without killing it.

Activity: your mini AI studio

Mission: create a mini-work combining at least 2 contents AI.

Themes to choose from: postcard from the year 3000 · fantastic animal · poster for the planet · trailer for an imaginary film

Presentation in 1 minute:

1. What I asked (my prompt)
2. What I got
3. What I had to correct

Golden rule: never the face or voice of a real person without their consent. Never. Even for a laugh.

The 5 ideas to remember

1. Multimodal AI sees, listens, speaks and creates — not just text
2. For AI, everything is numbers: pixels, sound waves. Motives, not feelings.
3. Vision: pixels → visual cues → recognition
4. Generate image/music/video: a description is enough — and it evolves at full speed
5. Deepfakes: source, cross-checking, distrust of strong emotions — before sharing