A walkthrough ofthe AI landscape
Five layers, one map. By 10:40 you'll know what actually sits where — and every talk after mine will have a place to land.
I build with this stuff every day — not in theory.
The day job
Co-founder of a South African software school. Accredited bootcamps. Hundreds of people placed into their first tech jobs — most of them started with zero background.
The other job
AfricanTechBro — I test AI tools in public so you don't waste six months on hype. Around 110,000 people follow along.
Why I'm on a screen
I'm in Cape Town, you're in Soshanguve. Which is itself the point: the map I'm about to draw is the same one from anywhere.
Everyone talks about AI. Almost nobody shows you the map.
What sits inside what
AI, machine learning, deep learning, generative AI, agents. Five words people use interchangeably. They are not the same thing — and after this you'll never mix them up again.
Where 2026 actually is
The shift from AI that talks to AI that does. What an agent is, what's inside one, and what that changes.
Where you stand
Which layer each of today's other speakers is standing on, where the work is going, and three things you can do tonight.
No hype. No doom. Just the map — so the rest of today makes sense.
It's not five different things. It's five circles inside each other.
You've been using AI since before you could spell it.
A human writes every rule by hand. The machine never learns anything; it just follows instructions very fast. Still everywhere, still useful.
can't be written
Write me the rules for "this photo contains a dog." Not the word — the picture. Fur? Cats have fur. Four legs? So does a table. You will be here all night.
That dead end is exactly where the next circle comes from.
Nobody taught you what a dog is by giving you rules.
Examples, not rules
Feed it 100,000 photos, each one labelled "dog" or "not dog." No description of a dog anywhere.
It gets it wrong
At the start it's basically flipping a coin. That's expected — and it's the whole mechanism.
Nudge the maths
Every wrong guess adjusts millions of internal numbers slightly. Repeat a few million times.
It works on new dogs
Now show it a dog it has never seen. It gets it right. Nobody ever defined "dog."
It's how your phone finds every photo of your mom — nobody ever tagged one. That's machine learning, and everything after this slide is a bigger version of it.
The idea was 60 years old. Three things showed up at once.
The internet happened
Billions of labelled images, pages and conversations, sitting there for free. Learning from examples needs examples — suddenly there were enough.
Gamers paid for it
Graphics cards were built to render video games. Turns out the maths for shading pixels is the same maths for training a model. An accident that changed everything.
Stack the layers
Layer one finds edges. Layer two finds shapes. Layer ten finds faces. Nobody programs those steps — the network invents them on the way down.
2012 is the year it stopped being a research paper and started beating humans at seeing.
It's autocomplete — that read the internet.
That's the entire trick. Predict the most likely next chunk of text. Then feed the whole thing back in and predict the next one. Again. And again. Do that with a model trained on most of the written internet, and "predicting the next word" turns into essays, working code, translations and arguments.
AI sorted, ranked, matched and recommended. It could tell you which book you wanted. It could not write one.
Same maths, much bigger. Now it produces new work in any voice, on demand. This is the moment your whole family heard the word "AI."
Your day, yesterday.
Your phone unlocks by looking at your face — deep learning, layer 3
PassiveMaps reroutes you around traffic before you hit it — prediction on live data
PassiveGmail finishes your sentence and bins three phishing mails — generative + classifier
PassiveYour bank blocks a card transaction that wasn't you — anomaly detection, layer 2
PassiveTikTok knows what to show you better than your friends do — recommender systems
PassiveAll of that happened to you. None of it was you directing it.
That gap — passive versus directed — is the whole career conversation.
The whole shift in 2026 is three words: talks → does.
in the box
You ask, it answers, you copy-paste. It cannot leave the chat window. It cannot open your files, click anything, or check whether it was right.
It writes one. You copy it, paste it, upload it — then do the whole thing again for the next 19 jobs.
you hired
You give it a goal, not a question. It runs a browser, reads files, writes files, runs commands, checks its own work, retries when it fails — then reports back.
You go to sleep. You wake up to 3 drafts.
Not smarter. Permitted. That's the entire difference — and it's why 2026 feels different to 2023.
An agent is not magic. It's four parts and a loop.
The brain
The model itself — Claude, GPT, Gemini, Llama. Does the reasoning and decides what to do next. You can swap it like a battery.
The hands
Tools it's allowed to use. Search the web. Read a file. Send an email. Run code. Every hand you give it is a thing it can now do — and a thing it can now break.
The notebook
Memory. What it already tried, what worked, what you told it last week. Without this it starts every task as a stranger.
The loop
Think → act → look at the result → think again. Keep going until the goal is met or it gets stuck. This is the part that makes it an agent.
Four kinds of company. Only one of them needs a billion dollars.
Build the brains
- Anthropic — Claude
- OpenAI — GPT
- Google — Gemini
- Meta, DeepSeek, Qwen, Mistral — open weights
Rent the compute
- Nvidia — the chips
- AWS, Azure, Google Cloud
- Vector databases, hosting, pipelines
Give models hands
- Coding agents in your terminal
- Cursor and friends
- Automation platforms · n8n, Zapier
Point it at a problem
- A tutor for TUT first-years
- A stock tracker for a spaza shop
- Whatever annoyed you this week
Every speaker after me is standing on one of these layers.
"Is this thing coming for my job?"
Look at what they have in common: same input, same output, every single day. Routine is exactly what a machine that learns patterns eats first.
Both ends grow — the deeply technical and the deeply human. It's the routine middle that thins out.
Globally: 92 million roles displaced, 170 million created by 2030 — and 39% of today's skills rewritten. The churn is the story, not the total.
World Economic Forum · Future of Jobs Report 2025 (1,000+ employers surveyed)
Don't learn a tool.
Learn the layer it sits on.
"ChatGPT prompt hacks." Act as an expert. I'll tip you $200. Take a deep breath. People paid for courses in this.
It predicts text. So give it context, give it examples, and tell it what good looks like.
The hacks stopped working — newer models outgrew them. Everyone who only learned the tricks started from zero. Everyone who understood why they worked carried it straight into the next tool and lost nothing. Same story with the job title: by 2025 under 0.5% of AI job postings said "prompt engineer" — the skill went everywhere, the title never arrived.
Move from passive to directed — tonight.
Direct it once, properly
Open any free assistant. Don't ask it a question — give it a job: "Act as my study coach for [module]. Quiz me one question at a time. Don't give me the answer until I've tried."
Watch an agent run
Read the walkthrough on my site and get one real open-source agent running on your own machine. Free, MIT-licensed, no card. Watch the loop actually happen.
Pick your problem
Write down one thing that annoyed you or someone you know this week. That's column 4 on slide 12. That's where your first real project comes from.
Nobody hires the person who understands the landscape. They hire the person who built something on it.
You're not behind.
You're early — and now you have the map.
Slides, the agent walkthrough and every prompt: africantechbro.ai
Instagram & TikTok @africantechbro.ai — comment MAP and I'll DM you this deck
Learn to build from zero: zaio.io