Map the Complete LLM System and the Next Subjects

Separate base architecture, pretraining, post-training, inference, retrieval, tools, evaluation, policy, and infrastructure, then route each layer to its curriculum owner.

An LLM product is more than a decoder checkpoint. A useful system map prevents a retrieval result, safety policy, serving optimization, or post-training method from being mistaken for a base-architecture property.

Separate the Layers

LayerExample questionLater owner
base architectureHow do records mix across positions?this subject / frontier architectures
pretraining data and objectiveWhich tokens and targets shaped the base model?language-model training and data
post-trainingHow are instructions or preferences learned?post-training and alignment
inference policyHow are logits converted into outputs?this subject / AI Engineering
retrieval and toolsWhat external evidence or actions are supplied?retrieval and agent systems
evaluationWhich behavior was measured and under what conditions?evaluation and safety
product policyWhich inputs or outputs are allowed?product safety and governance
infrastructureHow is the model placed, batched, and served?AI Engineering

One product change may cross several rows. Report each changed layer.

Reconstruct the Subject Path

Language Modeling defined sequences, conditional probabilities, tokenization, objectives, and evaluation. This subject built attention, Transformer blocks, decoder-only language models, training, generation, caching, and interpretation.

The next subjects can now add depth without hiding their assumptions:

  • AI Engineering: kernels, accelerators, precision, quantization, distributed training, serving, and operational measurement;
  • Post-training and Alignment: instruction tuning, preference methods, reward modeling, and behavior shaping;
  • Retrieval and Agents: external memory, tools, planning loops, permissions, and failure recovery;
  • Multimodal Models: modality encoders, fusion, objectives, datasets, and evaluation;
  • Frontier Architectures: state-space and other alternative sequence mixers;
  • Evaluation, Interpretability, and Safety: stronger behavioral and causal evidence, risks, and governance.

Q1. Route a system question

A team keeps the checkpoint fixed but adds a vector database and cites sources in generated answers. Which layer changed, and which later subject owns the depth?

Choose the layer and later subject

Select one choice, then check.

Hint
The system fetches information before or during generation.
Solution
The retrieval-and-tools layer changed. Retrieval and agent systems own the design, measurement, permissions, and failure modes in depth.
Not attempted
Review

Not marked done.

Finish with a Claim Audit

For any modern LLM claim, state the changed layer, invariant components, analytical consequence, empirical measurement, comparison baseline, source boundary, and remaining unknowns. That habit is more durable than memorizing a list of current model names.

Pause and reflect

In your own words, note what you understood, what remains unclear, or what you want to revisit. The note stays with this lesson.

0 of 1 exercises marked done

Review

Not marked done.

LLM PrimerMap the Complete LLM System and the Next Subjectshttps://llmprimer.com/transformers/variants-scaling-and-boundaries/map-the-complete-llm-system-and-the-next-subjects© 2026 LLM Primer