Four lenses to tell every "AI" apart — so headlines stop sounding like magic.
AI for DummiesModule 1
A spam filter, ChatGPT, a chess engine, Midjourney, and Tesla Autopilot
are all called "AI."
They work completely differently. This lesson gives you four questions to tell them apart.
Lesson arc: Capability → Output → Modality → Where it runs → Recap
Lesson 02 · Slide 2
Four lenses — one for each question
1
How capable?
Narrow → General → Super Where we are today? Narrow only.
2
What does it produce?
Predictive (classifies & forecasts) Generative (creates new content)
3
What kind of content?
Text · Image · Audio · Video Multimodal (several at once)
4
Where does it run?
Cloud (company servers) Local (your own machine)
Any AI tool you encounter — including ones invented next year — can be described with these four questions.
Lesson 02 · Slide 3
Lens 1: How capable? — Three tiers
Narrow AI (ANI)EXISTS TODAY — every tool you can name
Good at one kind of task. "Narrow" does not mean weak — it means specialised. A chess engine is superhuman at chess. Ask it to write a poem: blank.
General AI (AGI)HYPOTHETICAL — does not exist yet
As flexible as a human across any task. Seriously debated: some researchers say 5 years away, others say decades, others say the concept is ill-defined.
Superintelligence (ASI)SCIENCE FICTION — makes great movies
Smarter than all humans at everything simultaneously. Nothing to download, nothing to interact with. Exists only in speculation and movie scripts.
Lesson 02 · Slide 4
Narrow AI is everywhere right now
Every one of these is Narrow AI — each brilliant at its specific job, clueless outside it.
♔ Stockfish / AlphaZero
Chess & Go. Superhuman at the board. Cannot book you a flight.
🔑 Gmail Spam Filter
Classifies billions of emails / day. Terrible at writing a reply.
💬 Claude / ChatGPT
"Broad" language models. Still narrow: they can't drive your car.
🚗 Tesla Autopilot
Reads cameras, radar, maps. Cannot compose a text message.
🎧 Suno / Udio
Composes music from a text prompt. Cannot play chess.
📸 Midjourney
Paints images from words. Cannot answer a customer email.
The rule: whenever a headline says "AI can now…" — fill in "narrow AI got better at one specific thing."
Lesson 02 · Slide 5
The capability timeline — where we actually are
Every AI product you can subscribe to or download today sits in the green zone. The orange and red zones are speculation.
Lesson 02 · Slide 6
Lens 2: What does it produce?
📊 Predictive AI
Answers "which bucket?" or "what number?"
Input → classification or forecast.
This is the old, quiet, everywhere kind. Running silently in every major digital service for 20+ years.
spam/not-spamfraud detectionweather forecastchurn predictionmedical diagnosis AI
✨ Generative AI
Creates new content that didn't exist before.
Prompt → text / image / audio / video / code.
This is the new, loud kind. Exploded publicly in 2022 with DALL-E and ChatGPT.
ChatGPTClaudeMidjourneySunoSoraCopilot
Important: Both types are still Narrow AI. "Generative" describes the output format, not the capability level. A music-generating model cannot also write code.
Lesson 02 · Slide 7
Predictive AI — it's been around for decades
You encounter predictive AI dozens of times a day without noticing it.
🔑 Gmail & Outlook spam filter
Classifies each incoming email: spam or not-spam. Processes billions of messages per day. Model answer is a single bit.
💳 Stripe & PayPal fraud detection
Scores each transaction as fraudulent or legitimate in <100ms. Saves billions of dollars annually.
Generators: Sora, Runway, Pika Analysis: Gemini 1.5 Pro
⚡ Multimodal
Handles several types at once. Photo + text → answer about the image. Leading models in 2025–26.
A text-only LLM cannot "see" a photo, no matter how good it is at writing. Match the tool's modality to your task.
Lesson 02 · Slide 10
Multimodal in action — photo + question → answer
This is how Claude, GPT-4o, and Gemini handle mixed inputs. The model processes text and image together.
Most leading models in 2025–26 are multimodal by default. A text-only model (like an older Llama variant) simply cannot accept an image input — it's a hard limit, not a setting.
Lesson 02 · Slide 11
Lens 4: Where does it run?
Same idea, two very different places — with real consequences for privacy and cost.
☁️
Cloud (Closed)
Model runs on the company's servers. You connect over the internet.
ClaudeChatGPTGemini
🏠
Local (Open)
Model downloaded to your computer. Nothing leaves your machine.
LlamaMistralGemmaQwen
The distinction matters for privacy. Cloud: your prompts travel to a company server. Local: zero bytes leave your machine — conversations are completely private.
ollama run llama3.2 Downloads and runs in one command. Fast, scriptable.
LM Studio (GUI)
Graphical app, ChatGPT-style interface. No terminal needed. Mac/Win/Linux.
✓ Pros
100% private — nothing leaves your machine
Free per query after download
Works fully offline
You control the model version
💻 Hardware reality check
8 GB RAM — small models, slowly 16 GB RAM — comfortable for 7B models 32 GB+ — large models at good speed GPU / Apple M-chip — makes it much faster Modest laptop? Stay on cloud chat.
Lesson 02 · Slide 14
When closed vs open actually matters
☁ Choose Cloud when…
You want the most capable model available
You're starting out — zero friction
You need multimodal (text + image) easily
Your hardware is modest
You're not working with sensitive data
🏠 Choose Local when…
You're pasting confidential contracts or medical notes
You need it to work offline (travel, lab, no internet)
You run it thousands of times — per-query cost adds up
You want to modify or fine-tune the model itself
You're in a regulated environment (GDPR, HIPAA)
Example: A lawyer pasting a client contract should use local AI. A student asking for essay feedback is fine on cloud.
Lesson 02 · Slide 15
Quick check — classify this tool
You open Spotify. It shows you a "Recommended for you" playlist of songs it thinks you'll enjoy. What kind of AI is behind that feature?
Lesson 02 · Slide 16
All four lenses — combined mental map
Apply all four to any AI story. Here are three examples:
Tool / story
Capability
Output
Modality
Where
Gmail spam filter
Narrow
Predictive
Text
Cloud
Claude (claude.ai)
Narrow
Generative
Multimodal
Cloud
Midjourney
Narrow
Generative
Image
Cloud
Whisper (local)
Narrow
Predictive
Audio
Either
Tesla Autopilot
Narrow
Predictive
Video
Local
Llama (Ollama)
Narrow
Generative
Text
Local
Notice: every row has "Narrow" in the capability column. Because all real AI tools today are narrow AI. The theory page has an interactive grid to filter all 12 examples.
Lesson 02 · Slide 17
Recap — what you can do now
The four questions
Capability: All real AI is Narrow. AGI and ASI are hypothetical.