The brain
Model
The AI that does the reasoning. Pick one; swap it later.
Oslo Vibe Coding · a hands-on class
Vibe coding,
from scratch.
One hour. You will see how this works, where AI is heading, and how to build with your own tools. No experience needed.
What you'll leave with
1The philosophy.What vibe coding is, and the one skill that matters.
2The map.Where AI is, distilled from 13 talks by the people building it.
3Your own setup.Your AI coding tools running, with one thing built.
01
The philosophy
What we do when we vibe code, and why it works now.
The idea
Building by describing.
You say what you want in plain words. An AI writes the code and runs it. You look at the result and steer. You build first, and pick things up as you go.
Andrej Karpathy named this in 2025.
The loop
Describe. Build.
Run. Steer.
You go around this loop until the result matches what you meant. You direct; the model types. Most of your work is the steering.
The one skill that matters
Taste, not syntax.
You do not memorize a language. You build judgment: you decide what good looks like, then check that you got it. You can start tonight.
“Shipping code is going to zero. What is not going to zero is the taste to build something good.”
Diana Hu · Y Combinator
Be honest about this
Where it breaks.
The model writes code that looks right and can be wrong. It does not feel the mistake. Work in small steps, run it every time, read the parts that touch data or money, and check before you ship.
This is what separates a demo from a thing that works.
02
The map
Where AI is now, distilled from 13 talks by the people building it.
A map for later
The frontier, distilled.
Stanford's CS 153 put the people building modern AI in one room: Nadella, Altman, Jensen Huang, Ben Horowitz. There are short notes on the site distilled from all 13 talks. Skim the map now; watch the talks on your own time.
oslovibecoding.tech/frontierThe whole stack
Modern AI is layers.
Trust & safetyWhether AI keeps permission to operate. (Sullivan, Nadella, Midha)
Capital & marketsMoney and relationships organize the work; the context you own is the moat. (Horowitz)
ProductsTiny teams turn models into apps people use. You live here. (Singhal, Tan & Hu, Luma, ElevenLabs)
ModelsCompute and data become AI that reads, sees, hears, and acts. (Altman, Nadella, Blattmann)
ComputeChips and datacenters that turn power into thinking. (Huang, Vahdat)
EnergyPower plants and grids. The deepest constraint. (Nolan)
From the talks · 1 of 2
The bottleneck is not code.
Writing software is the cheap part now. Energy, compute, and knowing what to build are the scarce parts. The frontier is a race for power and chips.
One gigawatt of AI costs about $40B and takes 2 to 3 years. Energy is the one thing money cannot rush.
Amin Vahdat · Google
so whatIf code is cheap, your ideas and judgment are the valuable part. That is you.
From the talks · 2 of 2
Small teams, big output.
One person with AI does what used to need a department. The unit of building is shrinking to a few people, or one.
“You can do the work of about 500 to 1,000 people.”
Garry Tan · Y Combinator
so whatYou do not need a company or permission. One person and a laptop is a real team.
03
Build your setup
The hands-on part. Laptops out. We set this up together.
Use this model
Your setup has four parts.
The hands
Tools
It reads files, writes files, and runs commands.
The memory
Context
What it knows: your folder, your goal, your rules.
The engine
The loop
Describe, act, verify, steer. You drive it.
Claude Code, Cursor, and similar tools share these four parts.
Step 1 · pick a tool
Install one tool.
Pick one now. We help you set it up in the room.
Claude CodeCursor
# Claude Code (a terminal, needs Node 18+)
$ npm install -g @anthropic-ai/claude-code
$ claude # then sign in
$ npm install -g @anthropic-ai/claude-code
$ claude # then sign in
# Cursor (an app)
download from cursor.com
open it, then open a folder
download from cursor.com
open it, then open a folder
Step 2 · give it context
Open a folder.
State your goal.
Make an empty folder and open it in your tool. Say your goal in one sentence. For a bigger project, add a short rules file with the facts the tool should use.
Good context turns a lucky guess into a useful answer.
Step 3 · run the loop
Ask, check, steer.
Ask for the smallest useful version. Check what it made. Ask for the next small change. Watch the diff and reject what you did not want.
› Make a one-page site with my name,
a short bio, and a links section.
# after you look:
› Bigger heading. Dark background.
Add my GitHub link.
a short bio, and a links section.
# after you look:
› Bigger heading. Dark background.
Add my GitHub link.
Good first habits
↺One change at a time.Big asks hide mistakes.
▷Run it every time.See it work, do not assume.
✓Save when it works.A point you can go back to. Ask us how.
?Read the risky parts.Anything that touches data, money, or accounts.
One we built · try it
Live Oslo departures.
One HTML file asks the free Entur open data API for the next departures and shows them. No backend. No key. You could build this tonight.
oslovibecoding.tech/oslo-departures.html# the whole recipe
1. pick an open API (Entur)
2. ask it one question
3. draw the answer on a page
POST api.entur.io/.../graphql
{ stopPlace { estimatedCalls } }
1. pick an open API (Entur)
2. ask it one question
3. draw the answer on a page
POST api.entur.io/.../graphql
{ stopPlace { estimatedCalls } }
Now you · we build together
You have the idea.
Start building.
Take our departures idea or bring your own. Pick a small version. Get it working. We are in the room the whole time.
Oslo departures · Entur
Weather · Yr
City bikes · Bysykkel
Your own idea
Stuck for a few minutes? Wave someone over.
Tonight's recap
1You direct; the model types.Describe, build, run, steer. Taste over syntax.
2Ideas and judgment are the scarce part.So a beginner with taste can build useful things.
3Your setup is model + tools + context + loop.And it runs on your laptop now.
Now go build.
No one codes alone. oslovibecoding.tech