Skip to content

Articles

Talks and essays, written out in full.

Our talks, expanded into proper essays you can actually read. Honest, in-depth, and free, like everything else here.

A 5-part primer

How do LLMs work? Start the series →

/llms
Harsh Trivedi6 min readfrom a 15-page talk

Spec-driven development: write the plan before the code

At our 5 August session, we explored a simple shift for building with AI: decide what the software must do, write those decisions down, and only then ask the agent to build it.

Read the article
Harsh Trivedi7 min readfrom a 21-page talk

Loop engineering: from prompts to /goal

At our July 9 drop-in, we looked at the work hiding behind a good AI answer: the loop that tries, checks, adjusts, and knows when to stop. Then we put it to work on Oslo's live city-bike data.

Read the article
Akshat Jain6 min readfrom a 22-page talk

Our first vibe coding class at Spaces

Our vibe coding drop-in moved to Spaces in central Oslo this week. Here is the whole session for anyone who missed it: what vibe coding is, where AI is heading, how to set up your own tools, and two small apps we built from open Oslo data.

Read the article
Akshat Jain8 min readfrom a 32-page talk

What cheap AI does to the developer's job

Two old ideas, one from AI research and one from economics, explain why writing code got cheap and what that means for someone learning to build today.

Read the article
Akshat Jain8 min readfrom a 14-page talk

I spent $25,507 on AI coding agents in twelve weeks

A twelve-week receipt from real agentic engineering, and what the numbers actually mean for your own spend.

Read the article
Oslo Vibe Coding8 min readKarpathy’s deep dive · part 1 of 5

How an LLM is trained: the internet in a box

Pretraining is the first stage of building something like ChatGPT. By the end of this piece you will understand how raw internet text becomes a model that can continue any document, and why that model is not yet an assistant.

Read the article
Oslo Vibe Coding8 min readKarpathy’s deep dive · part 2 of 5

Inside the model: predicting the next token

What actually sits behind the chat box is one giant fixed math function. This piece explains what that function is, how it generates text one token at a time, and why that shape sets a hard limit on what it can do in a single step.

Read the article
Oslo Vibe Coding8 min readKarpathy’s deep dive · part 3 of 5

From base model to assistant

A base model is an internet-document simulator. Post-training turns it into ChatGPT by teaching it to imitate human labelers who write ideal answers. Once you see the model as a simulation of those labelers, its behavior stops feeling magical.

Read the article
Oslo Vibe Coding8 min readKarpathy’s deep dive · part 4 of 5

Why LLMs make things up, and what fixes it

A language model always emits a plausible next token, even when it has no idea. Once you see why, you can tell when to trust it, when to hand it a tool, and when to check its work.

Read the article
Oslo Vibe Coding9 min readKarpathy’s deep dive · part 5 of 5

Teaching models to think: RL and reasoning

After a model has read the internet and copied expert answers, one stage is left. Reinforcement learning lets the model practice, discover its own way to solve problems, and start to think before it answers.

Read the article