The routes
The graph knows which ideas lean on which. That makes several valid reading orders. These routes are computed from the prerequisite_of edges alone — walk them in order and nothing arrives before it is explained.
← the atlasevery concept →the taxonomy →
Start here
Ideas with nothing before them, ranked by how much of the graph they unlock.
a list of numbers treated as one point or one arrow
words used in similar company end up meaning similar things
multiply matching entries and add — a number for how aligned two vectors are
hide part of the data and let the model teach itself to recover it
the number training tries to make small — the formal statement of what you want
the first machine that learned weights to separate two classes
an internal simulator of how a situation evolves — imagining without acting
measure the gap to the goal and use it to correct the next step
a chunk of text — often a word-piece — that the model reads as one unit
the mathematics of belief under uncertainty
the rule that cuts text into tokens — it shapes everything downstream
a single number saying how well things just went
how many answers are still possible
how much uncertainty a message removes
The longest route
8 ideas deep — the longest dependency chain in the graph. It is not the only good order; it is the one that assumes the least.
- Vector
- Word embedding
- Attention
- Transformer
- Pretraining
- Fine-tuning
- RLHF
- GRPO
Routes to a destination
Pick a goal and walk backwards: these are the ideas the graph says must come first, in dependency order. The goal itself is highlighted.
Agent · 9 ideas before it
Transformer · 5 ideas before it
Diffusion model · 0 ideas before it
Retrieval-augmented generation · 1 ideas before it
Open weights · 1 ideas before it
236 concepts, 26 of them terminal — reachable goals you can build on once the route behind them is walked. Nothing on this page is hand-ordered; change a prerequisite in the graph and the routes change with it.