A Walkthrough of the AI Landscape

TL;DR

Twenty-five minutes, one job: hand the room a map.

Most people use AI, machine learning, deep learning, generative AI and agents as if they're five competing things. They're five circles inside each other. Once you can see the nesting, AI news stops being noise and the whole rest of the day's programme has somewhere to land.

This talk opens the event, so it's built as the spine for every talk that follows — slide 13 puts each of the day's other speakers onto the map by name.

The map (the one slide to photograph)

Circle	Since	What it actually means
Artificial Intelligence	1950s	The whole field. Any machine doing something that looks like thinking. Mostly hand-written rules.
Machine Learning	1990s	Stop writing rules. Show it examples, let it find the pattern itself.
Deep Learning	2012	Stacked layers of maths on huge data. The point where it started actually working.
Generative AI	2022	It stopped only sorting things and started making them — text, images, code.
Agents	2024	It stopped only answering and started doing. The live edge, right now.

Ask anyone who says "we use AI" which circle. Nine times out of ten they don't know.

The three ideas the talk actually turns on

1 · Nobody wrote the rulebook for "dog." Try to write out the rules for recognising a dog in a photo. Fur? Cats have fur. Four legs? So does a table. Nobody has ever finished that rulebook — so we stopped trying, and showed machines 100,000 labelled examples instead. Every wrong guess nudges millions of internal numbers slightly. A few million rounds later it recognises a dog it has never seen. That surrender is where machine learning comes from, and everything above it is a bigger version of the same trick.

2 · The gap is passive vs directed. Face ID at 06:40, the Maps reroute at 07:15, the fraud block at 13:00, the For You page at 20:00. All AI. All of it happened to you. None of it was you directing it. The entire career conversation lives in that gap — nothing else.

3 · Agents aren't smarter. They're permitted. A chatbot is the intern in the chat box: you ask, it answers, it can't leave the window. An agent gets a goal instead of a question — it browses, reads and writes files, runs commands, checks its own work and retries. Same model underneath. What changed is that we gave it hands.

An agent is only as safe as the hands you hand it. Permissions are the design decision, not an afterthought.

Anatomy of an agent — four parts and a loop

Part	What it is
🧠 Brain	The model — Claude, GPT, Gemini, Llama. Swappable like a battery.
🔧 Hands	Tools it may use: search, files, email, code. Every hand is a thing it can do and a thing it can break.
📓 Notebook	Memory. Without it, it meets you as a stranger every morning.
🔁 Loop	Think → act → check the result → think again. This is the part that makes it an agent.

Who builds what — so you can read AI news without drowning

Column	Who	Cost of entry
1 · Model labs	Anthropic, OpenAI, Google, and the open-weight crowd (Meta, DeepSeek, Qwen, Mistral)	Billions. About ten organisations on earth.
2 · Infrastructure	Nvidia, AWS, Azure, Google Cloud	Picks and shovels — sells to everyone else.
3 · Harnesses	Coding agents, Cursor, n8n, Zapier	Where most of 2026's real engineering jobs sit.
4 · Applications	A tutor for first-years. Stock tracking for a spaza shop. Whatever annoyed you this week.	Free tools, a laptop, and a problem you understand better than someone in San Francisco does.

Columns 1 and 2 need a billion dollars. Column 4 is wide open.

The rest of the day, on the map (slide 13)

This talk runs first, so its real job is to make the next four make sense:

Time	Speaker · talk	Which layer
10:15	Akhil — A Walkthrough of the AI Landscape	The whole map
10:40	Ayanda More — Breaking In: your first tech role	Columns 3 & 4 — the people who give models hands and point them at problems
11:05	Boikokobetso (Mr FingerZ) — Cybersecurity & career paths	The permissions problem, weaponised. Attackers get agents too.
12:20	Nonhlanhla Magagula — Responsible AI for real problems	The judgement layer — not what it can output, but whether it should ship
12:45	Simanga Mchunu — Machine learning & competitions	Straight back to circle 2 — the maths under all of it

Not five separate talks. One walk across one map.

Where the work is going

Per the WEF Future of Jobs Report 2025 (1,000+ employers surveyed):

• Shrinking fastest by 2030: data entry clerks, bank tellers, cashiers, postal clerks, admin assistants — every one a bundle of routine tasks. A machine that learns patterns eats routine first.
• Growing fastest: AI & ML specialists, big data, security specialists, software developers, fintech engineers — and nurses, teachers, skilled trades.
• The pattern: both ends grow — the deeply technical and the deeply human. The routine middle thins out.
• Globally: 92 million roles displaced, 170 million created by 2030; 39% of today's skill sets rewritten.

The one line to keep

Don't learn a tool. Learn the layer it sits on.

In 2023 the internet decided "prompt engineer" was the job of the future. By 2025, under 0.5% of AI job postings carried that title — the skill spread into every role, the job title never really arrived. Bet on a tool and you get stranded. Understand the layer and you just move up it.

Tonight — three things, 30 minutes

1. Direct it once, properly (5 min) — open any free assistant and give it a job, not a question:

Act as my study coach for [MODULE].

Quiz me on [TOPIC], one question at a time.

Don't give me the answer until I've tried. When I'm wrong, explain
why in simple words — then test me again with a similar question
until I get it right.

Keep score. At the end, tell me my weakest area and what to revise next.

2. Watch an agent run (20 min) — follow Run Your First AI Agent and get one real open-source agent running on your own machine. Free, MIT-licensed, no card. Watch the loop happen with your own eyes.

3. Pick your problem (5 min) — write down one thing that annoyed you, or someone you know, this week. That's column 4. That's your first project.

The only rule: nobody hires the person who understands the landscape — that took 24 minutes. They hire the person who built something on it.

Run of show — 17 slides · 25 minutes

Time	Slides	Block
0:00–2:45	1–3	Open · who's talking · the promise
2:45–5:15	4	The map — five nested circles (the slide they photograph)
5:15–9:45	5–7	Layer 1 rules → the dog problem · Layer 2 learning from examples · Layer 3 why 2012
9:45–13:15	8–9	Layer 4 generative — autocomplete that read the internet · your day yesterday (passive vs directed)
13:15–17:00	10–11	Layer 5 agents — talks → does · four parts and a loop
17:00–20:30	12–13	Who builds what · the rest of today, mapped
20:30–23:00	14–15	Where the work goes (WEF) · don't learn a tool, learn the layer
23:00–25:00	16–17	Three things tonight · close + back for the 13:10 panel

Full speaker notes with timings are baked into the deck — press S.

Delivered remotely — the notes that matter

The talk is given live from Cape Town to a room in Soshanguve, which changes a few things and they're built into the deck:

• No audience-response moments that need hearing — the room's audio won't carry. Ask for hands, not answers.
• No live demo. Nothing in the talk depends on a connection holding: the deck is a single self-contained HTML file with no external fonts, images, or scripts.
• Confirm sight and sound before slide 1 — "can everyone at the back read this?" — and wait for the organiser's thumbs up.
• Speak ~15% slower than in-person. Remote audio eats consonants.
• Hand back cleanly at 10:40. Ayanda is straight after; Q&A is a panel at 13:10, not part of this slot.

If time slips

• Never cut slide 4 (the map) or slide 10 (talks → does). Everything else is negotiable.
• At 5 minutes remaining with slides left: compress 7 (why 2012) and 12 (who builds what) to one line each.
• Slide 13 (the day mapped) is worth more to the organisers than any other slide — protect it.
• Slide 14 (WEF numbers) can be reduced to "routine shrinks, both ends grow" if you're truly out of time.

Sources

• World Economic Forum, Future of Jobs Report 2025 — 170M roles created, 92M displaced by 2030; 39% of skill sets transformed. Digest
• International Labour Organization (2025) — roughly one in four workers is in an occupation with some GenAI exposure; transformation far more likely than replacement. Publication
• "Prompt engineer" reality check — arXiv 2506.00058 (2025): under 0.5% of AI job postings carry the title; the skill diffused into every role.
• 2012 as the deep-learning inflection: the ImageNet/AlexNet result — the point deep learning moved from research paper to state of the art in computer vision.

Find Akhil

Comment MAP on the latest @africantechbro.ai post and this deck plus the agent walkthrough come back as one DM.

• Instagram / TikTok: @africantechbro.ai
• Guides and talks: africantechbro.ai
• Learn to build from zero: zaio.io