aitia (αἰτία) is Greek for cause — the answer to "why". A knowledge graph in which the edges carry the reasons: most maps of a field give you the nodes, but what is almost never written down is why one idea leads to the next. Press / to search.
Mathematical ground that more than one field stands on. A concept belongs here once a second domain needs it, not before.
Where to put the samples, and how the error is distributed when you cannot afford enough of them. — 9 nodes
What the machine is actually good at, and which of its limits a technique is fighting. — 7 nodes
What finite precision does to an otherwise correct algorithm. — 6 nodes
Rendering: how light is transported, what surfaces do to it, and how any of that is made to fit in a frame budget.
The quantities everything else is defined in terms of. — 11 nodes
Algorithms that solve the rendering equation, and the estimators they are built from. — 20 nodes
Models for what a surface does to light that hits it. — 30 nodes
Transport through a volume rather than across a surface. — 9 nodes
Getting a usable approximation of the above inside a frame budget. — 18 nodes
Visibility from the light, and the artifacts of every way of caching it. — 8 nodes
Reconstruction, in space and over time. — 12 nodes
Trading variance for bias so that low sample counts are shippable. — 5 nodes
Representing surfaces and finding where a ray meets them. — 25 nodes
Turning radiance into pixels a display can show. — 8 nodes
Learned scene representations, and their reconstruction lineage. — 7 nodes
Deciding what can be seen before paying to shade it. — 6 nodes
The aperture and the shutter, and the effects that come from each being finite. — 4 nodes
Training and architecture: the lineage of optimisers, normalisation, sequence models and generative methods.
How the weights actually get set, and what each step of that lineage fixed. — 23 nodes
Carrying information across a sequence, and the successive removal of bottlenecks. — 6 nodes
Learning to sample from a distribution rather than to label one. — 5 nodes
Fitting training and inference into the memory and bandwidth actually available. — 8 nodes
Pretraining, scaling and alignment of models over text. — 11 nodes
Learning from a reward signal rather than from labelled answers. — 11 nodes
Splitting a model and its optimizer state across devices once neither fits on one. — 6 nodes
A closed vocabulary. Every edge but part-of must say
why — an edge without a reason is rejected by CI.