<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <title>LLM Field Notes learning map</title>
  <id>https://humblemat810.github.io/llm-wiki/feed.xml</id>
  <link href="https://humblemat810.github.io/llm-wiki/feed.xml" rel="self" />
  <updated>2026-07-12T00:00:00.000Z</updated>
  <subtitle>Runnable notes for turning documents into useful, inspectable systems.</subtitle>
  <entry>
    <id>https://humblemat810.github.io/llm-wiki/notes/agents.html</id>
    <title>Tools make a system</title>
    <link href="https://humblemat810.github.io/llm-wiki/notes/agents.html" />
    <updated>2026-07-12T00:00:00.000Z</updated>
    <summary>How does a model take action? An agent is a control loop around a model: decide, call a tool, observe, and decide again. Reliability comes from constrained actions, state, checks, and a clear stop condition.</summary>
  </entry>
  <entry>
    <id>https://humblemat810.github.io/llm-wiki/notes/attention.html</id>
    <title>Attention is a lookup</title>
    <link href="https://humblemat810.github.io/llm-wiki/notes/attention.html" />
    <updated>2026-07-12T00:00:00.000Z</updated>
    <summary>How does context enter the computation? Queries ask what a position needs, keys advertise what is available, and values carry information back. Scaled dot products turn those matches into weights over the context.</summary>
  </entry>
  <entry>
    <id>https://humblemat810.github.io/llm-wiki/notes/embeddings.html</id>
    <title>Meaning in a vector</title>
    <link href="https://humblemat810.github.io/llm-wiki/notes/embeddings.html" />
    <updated>2026-07-12T00:00:00.000Z</updated>
    <summary>How does a model store meaning? An embedding is a learned coordinate system. Similar uses often land near one another, but dimensions are not a hand-labeled dictionary. The geometry is shaped by training data and the objective that created it.</summary>
  </entry>
  <entry>
    <id>https://humblemat810.github.io/llm-wiki/notes/evaluation.html</id>
    <title>Measure the thing</title>
    <link href="https://humblemat810.github.io/llm-wiki/notes/evaluation.html" />
    <updated>2026-07-12T00:00:00.000Z</updated>
    <summary>How do I know if it works? Evaluation is a decision instrument. Define the user task, collect representative examples, record acceptable failures, and measure changes with a repeatable harness.</summary>
  </entry>
  <entry>
    <id>https://humblemat810.github.io/llm-wiki/notes/finetuning.html</id>
    <title>Teach, don&apos;t just prompt</title>
    <link href="https://humblemat810.github.io/llm-wiki/notes/finetuning.html" />
    <updated>2026-07-12T00:00:00.000Z</updated>
    <summary>When should I change the weights? Fine-tuning changes behavior through examples. It is useful for repeatable format, style, or task behavior; it is not a magic database update.</summary>
  </entry>
  <entry>
    <id>https://humblemat810.github.io/llm-wiki/notes/inference.html</id>
    <title>One token at a time</title>
    <link href="https://humblemat810.github.io/llm-wiki/notes/inference.html" />
    <updated>2026-07-12T00:00:00.000Z</updated>
    <summary>What happens at inference? Generation is a loop: run the model, turn logits into a distribution, choose a token, append it, and repeat. A key/value cache avoids recomputing old context.</summary>
  </entry>
  <entry>
    <id>https://humblemat810.github.io/llm-wiki/notes/knowledge-graphs.html</id>
    <title>A graph is a hypothesis</title>
    <link href="https://humblemat810.github.io/llm-wiki/notes/knowledge-graphs.html" />
    <updated>2026-07-12T00:00:00.000Z</updated>
    <summary>How does a document become a representation you can improve? A knowledge graph is not a magical summary of a document. It is a compact, inspectable hypothesis about which concepts matter, how they relate, and what evidence supports each claim. That makes it useful: a person can c</summary>
  </entry>
  <entry>
    <id>https://humblemat810.github.io/llm-wiki/notes/production.html</id>
    <title>Ship the boring parts</title>
    <link href="https://humblemat810.github.io/llm-wiki/notes/production.html" />
    <updated>2026-07-12T00:00:00.000Z</updated>
    <summary>What survives contact with users? A production system is a set of promises: response time, cost, availability, data handling, and behavior when the model is wrong or unavailable. The fallbacks and observability are part of the product.</summary>
  </entry>
  <entry>
    <id>https://humblemat810.github.io/llm-wiki/notes/rag.html</id>
    <title>Give it a library</title>
    <link href="https://humblemat810.github.io/llm-wiki/notes/rag.html" />
    <updated>2026-07-12T00:00:00.000Z</updated>
    <summary>How can a model use my data? Retrieval gives a model relevant context at request time instead of asking its weights to contain every fact. Search, chunking, citation, and freshness often matter more than a clever final prompt.</summary>
  </entry>
  <entry>
    <id>https://humblemat810.github.io/llm-wiki/notes/scaling.html</id>
    <title>Scale changes the game</title>
    <link href="https://humblemat810.github.io/llm-wiki/notes/scaling.html" />
    <updated>2026-07-12T00:00:00.000Z</updated>
    <summary>Why does scale work? More parameters alone are not a strategy. Useful scaling balances model capacity, data, and compute so the model keeps finding structure instead of memorizing noise.</summary>
  </entry>
  <entry>
    <id>https://humblemat810.github.io/llm-wiki/notes/tokens.html</id>
    <title>Tokens are the interface</title>
    <link href="https://humblemat810.github.io/llm-wiki/notes/tokens.html" />
    <updated>2026-07-12T00:00:00.000Z</updated>
    <summary>Why can&apos;t the model see words? A language model receives a sequence of token IDs, not words. A tokenizer turns bytes or characters into a vocabulary of reusable pieces. That boundary controls context length, cost, spelling behavior, and how much information can fit into a request</summary>
  </entry>
  <entry>
    <id>https://humblemat810.github.io/llm-wiki/notes/training.html</id>
    <title>Loss is a compass</title>
    <link href="https://humblemat810.github.io/llm-wiki/notes/training.html" />
    <updated>2026-07-12T00:00:00.000Z</updated>
    <summary>What does &apos;learning&apos; mean here? Loss compresses prediction error into a signal that can move parameters. It is useful because it is differentiable, not because it perfectly captures what a person values.</summary>
  </entry>
  <entry>
    <id>https://humblemat810.github.io/llm-wiki/notes/transformers.html</id>
    <title>The transformer stack</title>
    <link href="https://humblemat810.github.io/llm-wiki/notes/transformers.html" />
    <updated>2026-07-12T00:00:00.000Z</updated>
    <summary>What is the machine made of? A transformer repeats a compact block: mix information across positions with attention, transform each position with an MLP, and preserve a clean path with residual connections and normalization.</summary>
  </entry>
</feed>
