Understand
7 min Context as working memory
Everything a model knows about your situation, it knows because it is in the window right now. Managing that window is a design job.
WorkBuild
Open landmark →EXPLORE THE ATLAS
Places to build
judgment.
Switch your lens to change examples and recommendations. The underlying concepts stay shared and trustworthy.
WHAT THIS ISA landmark is one idea you can learn in about five minutes, ending in a decision you make. Rings group them by how far along you are, not by topic. See the legend
22 landmarks visible
Understand
7 min · EUData provenance
Where the training data came from, whether anyone agreed to it, and whether you can tell what a piece of content is.
WorkBuildLead
Open landmark →Understand
8 min · EUDrawing the system
The model is one box among many. Drawing the whole path — and where it can fail — is what turns understanding into judgment.
WorkBuildLead
Open landmark →Understand
7 min Embeddings
Text turned into coordinates, so that “nearby” means “similar in meaning” — the basis of search that doesn’t need matching words.
WorkBuild
Open landmark →Understand
8 min Fine-tuning and alignment
A raw pretrained model doesn’t answer questions — it continues text. Everything that makes it an assistant happens afterwards.
WorkBuildLead
Open landmark →Understand
7 min How a model uses tools
A model can’t run code or send email. It can emit a structured request that something else executes — and that gap is where safety lives.
WorkBuildLead
Open landmark →Understand
8 min · EUHow evaluation works
Without a fixed set of cases with known answers, “it seems better” is the only quality signal you have — and it isn’t one.
WorkBuildLead
Open landmark →Understand
8 min How models are trained
Pretraining is a long, expensive pass over a fixed corpus. What was in that corpus — and what wasn’t — shapes everything after.
WorkBuildLead
Open landmark →Understand
7 min How multimodal models work
Images, audio, and video become tokens too. That’s why one model can handle them — and why it fails at them in specific ways.
LifeWorkBuild
Open landmark →Understand
8 min Inside a transformer
Attention lets every token be interpreted in the light of the others. That mechanism is most of what makes modern models work.
WorkBuild
Open landmark →Understand
7 min Memory and state
Models are stateless. Every “it remembers me” feature is a store outside the model, re-inserted into the window each time.
LifeWorkBuildLead
Open landmark →Understand
6 min · EUModel cards and documentation
The document that says what a model was built for, tested on, and known to be bad at — and how to read one that isn’t saying much.
WorkBuildLead
Open landmark →Understand
7 min · EUOpen and closed models
“Open” usually means downloadable weights, not open source. The difference decides what you control and what you inherit.
WorkBuildLead
Open landmark →Understand
7 min Predicting the next token
A language model produces one token at a time from a probability distribution. Everything else it appears to do sits on top of that.
LifeWorkBuild
Open landmark →Understand
7 min Reasoning and extended thinking
Some models will “think” before answering — producing intermediate steps, or spending extra computation, to do better on hard problems. The steps help, but they are generated text, not a proof.
LifeWorkBuild
Open landmark →Understand
8 min Retrieval-augmented generation
Fetch the relevant passages first, then answer from them. It’s the main way to make a model current, specific, and citable.
WorkBuildLead
Open landmark →Understand
6 min Seeing inside a model
A model is billions of numbers with no built-in explanation of itself. Interpretability is the research effort to read those internals — and it's why “ask it why” is not the same as knowing why.
LifeWorkLead
Open landmark →Understand
7 min Size, distillation, and quantisation
The same capability can be delivered at very different sizes and costs. Knowing how explains why a model runs on a laptop at all.
WorkBuild
Open landmark →Understand
6 min Tokens
Models don’t read letters or words. They read tokens — and that choice explains a surprising number of their failures.
LifeWorkBuild
Open landmark →Understand
7 min Uncertainty and confidence
A model’s tone carries no information about whether it is right. Getting a usable uncertainty signal takes deliberate work.
LifeWorkBuildLead
Open landmark →Understand
7 min What a benchmark measures
A leaderboard score is a real measurement of a specific thing — almost never the thing you actually need.
WorkBuildLead
Open landmark →Understand
7 min What happens at inference
Running a model is a sequence of two very different phases. Knowing which one you’re paying for explains most of what you feel as speed and cost.
WorkBuild
Open landmark →