Most writing about AI either flatters the reader or drowns them. This course does neither. It teaches the actual mechanism: the formulas, the worked arithmetic, and the measured before-and-after numbers from the papers that moved the field. Every claim traces to a primary source, and every number is the one the source reported.
When a lesson names a formula, it walks through what the formula computes, term by term, then works a numeric example with real figures.
Parameter counts, error rates and benchmark scores come from the paper that reported them, named inline and listed in full at the end of each lesson.
Model names, prices and leaderboard positions age fast. Where a lesson depends on one, it is flagged as a dated snapshot rather than presented as permanent fact.
The order is load-bearing. Each tier uses vocabulary the tier before it established, so the agent and multi-agent material at the end can be written at full depth rather than stopping to re-explain embeddings. Fifteen lessons are published, through the first three tiers. The rest are drafted and are being prepared for the web.
Five nested terms with the exact boundary between each, and the three shifts that built the field: symbolic AI to statistics, statistics to AlexNet, AlexNet to the transformer.
02Training as a deterministic search. Loss functions, the gradient descent update rule, and a two-parameter model learning a line across three hand-calculated iterations.
03Neuron to dense-layer matrix math, why nonlinearity is required, and a full backpropagation worked example.
04Why recurrent networks hit a wall, the full query-key-value derivation, multi-head attention and positional encoding.
05What a token is, why some words trip up models, and why pricing is quoted per token.
06Pretraining, the scaling laws, and a first look at how a raw predictor becomes an assistant.
07The word2vec lineage, cosine similarity worked through, and what an embedding dimension actually buys.
Zero, one and few-shot prompting formally defined, plus an audit of which popular prompting tricks have primary-source backing.
10Chain-of-thought mechanics and the measured benchmark jump it produced.
11Why attention cost grows with the square of the input, and why a context window is not memory.
12Vision transformers, contrastive image-text alignment, speech recognition and the diffusion process.
13Why next-token prediction does not optimize for truth, and which mitigations have measured effect.
14Chunking trade-offs, vector index mechanics, and reranking.
Retrieval failure modes, grounding accuracy, and the vector-store risk category.
Request and response shapes, sampling parameters, streaming, rate limits and retries.
The observe, reason, act loop, tool definitions, and least-privilege rules stated early rather than deferred.
A documented bias case study with full error-rate tables, and near-term harms separated from speculative risk.
Benchmark scoring mechanics, judge-model bias modes, contamination, and the benchmark illusion.
The memory arithmetic of full fine-tuning, low-rank adaptation worked through, and a decision matrix.
The full alignment pipeline with a worked reward-model trace, and the newer closed-form alternatives.
Sub-word algorithms compared, glitch tokens, and the cross-lingual cost penalty.
Soft-target transfer, a full benchmark table for distilled variants, and the deployment economics.
The memory-bandwidth bottleneck and measured quality cost at each precision level.
Cache arithmetic, three worked hardware sizing scenarios, and a full cost-latency example.
Attention variants and mixture-of-experts routing, with specific models treated as replaceable case studies.
Six named reasoning architectures, each with a worked example or a measured result.
Tool-calling contracts compared, the Model Context Protocol, and tracing taxonomy.
The harness and scaffold stack, and the failure modes that appear only in long sessions.
Named team architectures with a full ablation table, and the token-multiplier arithmetic.
A graph-theory refresher, the static versus dynamic topology trade-off, and a full failure taxonomy.
The confused-deputy mechanics of indirect prompt injection, sandboxing, and policy models.
Three evaluation axes applied across current tools. Written as a dated snapshot and a living document.
