Machine learning
The root page of this wiki: fitting a model to data so it performs a task no one wrote rules for. The spoke treats machine learning as an engineering discipline rather than a body of theory — what it costs to train, what it takes to adapt, and what breaks between a demo and a deployment.
The three layers this corpus covers
A model's life runs through stages, and the founding sources land on three of them:
| Layer | What it asks | Founding source |
|---|---|---|
| Algorithm | How does a model fit data at all? | mlalgorithms → classical-ml-algorithms |
| Adaptation | How is a pretrained model turned to a new task? | finetuning-explained-kdnuggets, unsloth-amd-support → fine-tuning |
| System | What gets built around the model so it earns its keep? | ml-system-design-case-studies, ai-projects-fail-infrastructure-people → ml-system-design, demo-to-production-gap |
The gap in the middle is the founding one: nothing here yet covers training a model from data end to end — data collection, labelling, pretraining, evaluation. The corpus jumps from the fitting procedure to adapting somebody else's finished model. See synthesis.
Closed since, in both halves. Data collection and labelling arrived 2026-08-08 and after (training-data-quality); the same page now also holds the opposite strategy — generating supervision from simulated worlds instead of annotating observed ones (code-as-world-report) — which is where a task nobody can annotate gets its labels.
Underneath the three layers
information-theory-inference-learning-algorithms added a fourth thing the corpus lacked — not a stage in a model's life but the theory beneath all of them. Fitting a model is inference over its parameters (bayesian-inference), the loss you minimize is a likelihood, model complexity is priced by the evidence rather than by a penalty you choose, and the same message-passing machinery decodes an error-correcting code (information-theory). It is the spoke's only source that explains why any of this works, and its only T1. It is also twenty-three years old and says nothing about what training costs.
Where the rest of it lives
The hub splits a model's life across spokes, and this one is deliberately train-time and
design-time. Serving mechanics (sampling, KV cache, batching) are ../llm-inference-wiki; the
provider and pricing market is ../llm-providers-wiki; agent frameworks are
../agentic-tooling-wiki; running the production platform is ../platform-ops-wiki. Full
boundaries in this spoke's CLAUDE.md.
Related
classical-ml-algorithms · fine-tuning · ml-system-design · demo-to-production-gap · bayesian-inference · information-theory · monte-carlo-methods · training-data-quality · executable-world-representation · synthesis