LatentLens · Mastery
LLM Mastery
From one token to a production AI system. Every topic answers the same ten questions and follows the same A to M format, using the real Qwen3-8B-Base architecture wherever a fact can be verified.
New to all of this? Start with the Storybook (level 0) for intuition. Progress is stored only in this browser.
The path
- 10/14 mastered · not startedTransformer foundationsTrace one token through Qwen3-8B · after level 0
- 1.1 Autoregressive language modelingNot started
- 1.2 TokenizationNot started
- 1.3 EmbeddingsNot started
- 1.4 The residual streamNot started
- 1.5 RMSNorm (and Q/K norm)Not started
- 1.6 Linear projections (xW)Not started
- 1.7 Queries, keys and valuesNot started
- 1.8 MHA, MQA and GQANot started
- 1.9 RoPE: rotary position embeddingNot started
- 1.10 Causal attention, end to endNot started
- 1.11 The MLP (SwiGLU)Not started
- 1.12 Residual connections and depthNot started
- 1.13 Final norm and LM headNot started
- 1.14 Softmax and decodingNot started
- 20/9 mastered · not startedTraining and post-trainingFrom corpus to gradient update to assistant · after level 1
- 30/3 mastered · not startedInference and servingFrom request to prefill, KV cache and decode · after level 1
- 40/2 mastered · not startedRAG and retrievalGround answers in retrieved text · after level 1 and 3
- 50/3 mastered · not startedAgents and toolsLoops, tools, memory and MCP · after level 4
- 60/3 mastered · not startedInterpretability and latent statesRead and test what is inside · after level 1
Levels 7 to 10 continue in the Security Lab; level 11 is the research workspace.
Practice
- 10-minute daily reviewTen questions chosen from what you missed or have not reviewed lately.
- I keep confusing these31 side-by-side pairs, from token vs embedding to MCP client vs server.
- Follow one request end to endOne email-assistant request through the UI, the model and every trust boundary.
- Interview questionsSix levels: explain, trace, compare, debug, threat-model, design.
- Final mastery testTen free-response parts, from drawing the model to Professor Regan's research question.
How it all connects
Five lanes of one system. Pick any step to see where it sits and what it touches in the other lanes.
How an LLM computes
How it learns
How it runs
How it is used
How it fails and is secured
Residual stream
Where am I? How an LLM computes, step 5 of 10.
What comes before? Embedding.
What comes after? 36 blocks: norm, attention, MLP.
Connections to other lanes (highlighted):
Open the lesson →- RAG: Retrieved text becomes tokens whose states mix into the residual stream through attention.