i prepared this for an AI workshop at TUM Innovation Week 2026, long story short, it did not happen, they moved my session to a later time and i wasn’t gonna make it anyways. so here’s the two hours of me talking stuff, in words.
fair warning: there are some dad jokes in here. i refuse to remove them.
what is AI, really?
artificial intelligence is systems that can learn from data, recognize patterns, and make decisions — tasks that normally require human intelligence. three things: learn, reason, act.
- learn: from data and examples, not just rules
- reason: find patterns, make predictions
- act: generate text, images, code, decisions
AI is not magic. it’s very expensive, very complicated math running on very expensive computers. but fundamentally, it’s pattern matching at scale.
how we got here: the full timeline
AI is older than most university students. here’s the honest version of the history:
1956 — the term “artificial intelligence” gets coined at a summer workshop at Dartmouth College. a bunch of professors basically said “let’s make machines think” and then… didn’t. for about 40 years. that’s what we call the AI winters, when funding dried up because the hype didn’t match the results.
1997 — Deep Blue beats Garry Kasparov at chess. HUGE moment. Kasparov had said “a machine will never beat me” and then it did. in his defense, we’ve all said things we regret.
2012 — a neural network called AlexNet crushes an image recognition competition so badly that everyone suddenly remembered AI existed. this kicked off the deep learning revolution we’re still riding.
2017 — Google publishes the “Attention Is All You Need” paper. that’s the Transformer architecture, the T in GPT. without this paper, none of the tools you use today would exist.
2022 — ChatGPT drops and everyone becomes an AI expert overnight. “have you tried ChatGPT?” became the new “have you tried turning it off and on again?”
2023-24 — open source models catch up fast. Llama, Mistral, suddenly you could run AI on your laptop. also, AI agents started doing actual work: browsing the web, writing code, booking things.
2026 — where we are right now. agents are everywhere. companies are building AI employees. and you’re reading this, which means you’re ahead of a significant chunk of the population.
the big picture: 70 years of research, 5 years of breakthroughs, and AI is now a tool anyone can use. we went from “AI can play chess” to “AI can write your thesis” in one generation.
three types of AI (and which one we actually have)
people mix these up constantly, so let’s be clear:
narrow AI — what we have today. excellent at one task. ChatGPT is great at text, DALL-E is great at images, but neither can do what the other does and neither can go buy groceries.
general AI (AGI) — human-level reasoning across domains. still science fiction.
super AI — beyond human. very sci-fi. we’re nowhere near this.
everything you use today, every tool in this article, is narrow AI. specialists, not generalists.
AI is already in your life
you’ve been using AI every single day without realizing it:
| tool | what’s happening under the hood |
|---|---|
| Google Maps traffic | predicting congestion from millions of GPS signals in real-time |
| Netflix / YouTube recommendations | ML models learning your taste from watch history and similar users |
| Gmail spam filter | classifying emails using NLP, trained on billions of flagged messages |
| WhatsApp autocomplete | next-word prediction trained on language patterns |
| face unlock on your phone | a computer vision model mapping your face geometry in 3D |
| M-Pesa fraud alerts | anomaly detection flagging unusual transaction patterns instantly |
| Uber / Bolt surge pricing | demand forecasting models adjusting prices based on predicted ride requests |
none of these asked you to “use AI.” it just works in the background. that’s the point.
AI vs automation: the distinction that matters
this is the one that trips everyone up.
automation follows fixed rules. same input, same output, every time. your alarm clock is automation. Excel auto-calculating totals is automation. a traffic light on a timer is automation. no learning, no adapting, fast and reliable.
AI learns patterns from data. it handles new situations, gets better over time, and makes predictions. same ingredients twice and you might get a slightly different dish, because it learned from the last time.
the best real-world systems combine both. automation handles the obvious stuff, AI handles the tricky situations.
the M-Pesa case study
people assume M-Pesa’s fraud detection is all AI. it’s actually both working together, and that’s the whole point.
layer 1 is automation: hard-coded rules. try to send Ksh 500k to a new number at 3am? that gets flagged by a simple IF statement. no machine learning needed.
layer 2 is AI: it builds a profile of YOUR normal behavior. you usually send money to family on Fridays? great. you suddenly send 5 transactions to 5 new numbers in 10 minutes on a Tuesday? that’s unusual for YOU, even if no single rule is broken. that’s the AI layer.
the takeaway: most real-world systems are a sandwich. automation on the outside (fast, cheap), AI on the inside (smart, adaptive). knowing which layer to use is the skill.
the AI tools landscape
for everyone (no coding required)
| category | tools |
|---|---|
| chat & writing | ChatGPT, Claude, Gemini, Perplexity |
| images & design | Midjourney, DALL-E, Canva AI, Adobe Firefly |
| video & audio | Runway, CapCut AI, ElevenLabs, Descript |
| productivity | Notion AI, Microsoft Copilot, Google Gemini |
| learning | Khan Academy AI (Khanmigo), Duolingo Max |
Claude and ChatGPT are free to use. Canva AI is free tier. you don’t need a budget to start.
for developers
| category | tools |
|---|---|
| coding assistants | GitHub Copilot, Cursor, Claude Code, Windsurf |
| APIs & platforms | OpenAI API, Anthropic API, Hugging Face, Replicate |
| frameworks | LangChain, LlamaIndex, CrewAI, AutoGen |
| local / open source | Ollama, LM Studio, Llama 3, Mistral, Qwen |
| MLOps & data | Weights & Biases, MLflow, DVC, Label Studio |
Ollama lets you run Llama 3 on your laptop. no API key, no internet connection, no sending your code to a server. privacy-first AI, local.
prompting: the CRAFT framework
the quality of your prompt directly determines the quality of the output. vague prompt gets vague results. “write me something about business” will get you generic garbage.
here’s the framework that fixes that:
C — context: give background info. who you are, what this is for.
R — role: tell the AI what role to play. “act as a senior software engineer reviewing code” changes how it responds entirely.
A — action: be specific about what you want. write, analyze, compare, summarize.
F — format: specify output format. bullet points, table, 200 words, markdown, JSON.
T — tone: set the vibe. professional, casual, academic, fun.
the difference in practice:
bad prompt:
write me something about business
good prompt:
you are a business analyst. write a 200-word LinkedIn post about why
Mombasa-based startups should adopt AI for inventory management.
tone: professional but approachable. include one real example.
night and day difference. every element you add narrows the output toward what you actually want.
a common mistake to avoid: too vague (“help me with my homework”), too long (a wall of text), or no format specified (and you get an essay when you wanted bullet points).
AI agents: beyond chatbots
most chatbots just answer questions. you ask, they respond. end of story.
AI agents are different. an agent can perceive its environment, make decisions, and take actions to achieve goals, autonomously. the key difference is autonomy and tool use. agents can:
- browse the internet
- read and write files
- send emails
- execute code
- chain multiple steps together
a chatbot is a waiter. takes your order, brings your food, that’s it. an AI agent is a personal chef. checks what’s in the fridge, plans a meal, goes shopping if needed, cooks it, plates it, and sends you the grocery receipt.
real agents in the wild: Claude Code (writes, tests, and deploys code from your terminal), Devin (autonomous software engineer that takes a ticket and ships a PR), AutoGPT / CrewAI (multi-agent teams collaborating on complex tasks).
key takeaways
AI is a tool, not magic. understand what it can and can’t do, and you’ll use it 10x better than everyone treating it like a black box.
better prompts = dramatically better results. the CRAFT framework gives you the structure. context, role, action, format, tone.
AI agents are the next frontier. they act, they plan, they use tools. chatbots answer questions; agents get work done.
you can build AI-powered tools today. open APIs, open source frameworks, free tiers — the barrier is lower than ever.
start using AI now. the gap between people who use AI effectively and people who don’t is growing. people who learn to use it well will have a significant advantage in the next few years.
resources to keep going
- chatgpt.com — start chatting with AI
- claude.ai — Anthropic’s AI assistant
- learnprompting.org — free prompting course
- huggingface.co — open source AI models and datasets
Written and Authored by Chris