Machine Learning: The Complete Architecture
DeepDeepest No. 003. Static text for crawlers and readers without JavaScript. The interactive page loads above when the bundle runs.
What this piece establishes
Machine learning is a pipeline from data and features through model architecture, training, evaluation, serving, and agent tooling. The taxonomy runs from classical models through CNNs, transformers, mixture of experts, RAG, graph neural networks, and edge inference.
Training pipeline
Data collection and curation, feature engineering, model selection, training infrastructure, loss and optimization, evaluation metrics, then fine-tuning and alignment (including RLHF). Each stage has different failure modes. Garbage in the first stage cannot be recovered in the last.
Transformers
Self-attention lets every token condition on every other token in the window. That is the mechanism behind modern language models, and it is also why context windows and serving cost dominate production economics.
Graph Neural Networks
GNNs pass messages over edges instead of over a sequence. They are the right primitive when the data is a graph: molecules, knowledge graphs, or a JavaScript syntax tree. A later catalog piece (pipeline No. 036) applies this to storefront scripts.
RAG and serving
Retrieval-augmented generation grounds generation in documents you control. Edge inference moves a smaller model next to the request. Both are architecture choices about where state and compute live, not product features.
Read next
No. 008 AI Agent Runtime and No. 018 Anthropic Claude Architecture pick up serving, tools, and loops.
