Prerequisites & knowledge map
Everything below is a concept Wiring Language Models leans on that ISN'T taught inside the course itself — not the whole of another course, just the specific skills.
Deep Learning / Neural Networks
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
Matrix multiplication
A layer's forward pass is a matrix-vector product plus a bias.
required -
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
College Algebra
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
Expected value
Entropy is the expected value of surprise.
helpful
Calculus I — Differential & Integral Calculus
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
-
required
English Grammar and Sentence Construction
-
required
-
required
-
required
-
Nouns
A tokenizer's "word-level" option is exactly the human notion of a word this skill already names -- knowing what a word IS (and that punctuation and word boundaries aren't the same thing) makes the character/word/subword trade-off concrete rather than abstract.
helpful -
Verbs
Subword tokenization frequently splits a verb into a stem plus an inflectional ending (e.g. "walk" + "##ing") -- recognizing that split as a genuine word-formation pattern, not an arbitrary cut, is easier having already named verbs and their forms.
helpful -
Where a sentence begins and ends
A tokenizer's sequence-length trade-off is stated per sentence or document; knowing where a human reader would say a sentence starts and ends gives the character/word/subword comparison a real unit to measure against.
helpful
Computer Vision
-
Bounding boxes, IoU, and non-max suppression
Text detection is a detection-shaped problem -- it reuses ch.13's own bounding-box/confidence machinery to locate text regions before recognition ever runs.
recommended -
Contrastive image-text embedding (the CLIP-style mechanism)
computer-vision teaches how contrastive image-text embeddings are actually trained; this course only needs to use the resulting shared vector space for retrieval.
recommended -
The OCR pipeline -- text detection, recognition, and document layout
computer-vision owns the OCR pipeline (text detection then recognition) in full; this course only needs the working-depth "OCR-plus-LLM" pattern, not the pipeline's own internals.
recommended -
Zero-shot classification and retrieval via a vision-language model
computer-vision's own VLM skill goes deeper on zero-shot classification/retrieval via a vision-language model than this course's own working-depth treatment needs to.
recommended