Trace every transformation
Follow text through tokens, vectors, attention, logits, sampling, and the feedback loops that shape model behavior.
Large Language Models · from first principlesUnderstand the machine.
Understand the machine.
From first principles.
Start with the visible text-to-prediction loop, then add architecture, learning, serving, and safety. Formal mathematics stays available as optional depth after the mechanism is clear.
44 lessons · about 17 hours
THE LLM PIPELINE
“language”→tok_1842
next tokenmodelillustrative flow · not a measurement
Stop treating language models like magic.
Language models now shape how we write software, search knowledge, create media, and automate work. Understanding what happens beneath the chat box turns you from a spectator into someone who can build, evaluate, and challenge these systems with confidence.
Connect model to system
Join training, post-training, inference, retrieval, tools, evaluation, and serving into one coherent engineering picture.
Know what evidence can prove
Separate fluent demos from reliable capability, benchmark movement from real usefulness, and model behavior from system safety.
The finish lineLeave with a working mental model of the entire LLM stack—and the judgment to use it well.
From a next-token rule to a dependable LLM system.
Five cumulative phases add one layer of machinery at a time. The active diagram shows where each phase changes the system.
NEURAL FIELD GUIDE / COURSE ARC01 · 05
See how an LLM builds text
Follow text through pieces, representations, position, attention, layers, and next-token prediction before opening the formal mathematics.
You unlockExplain the complete text-to-prediction loop
Explain how the loop learns
Return to the visible mechanism and formalize representations, probability, feedback, parameter updates, and the tiny GPT build.
You unlockTrace one complete learning step
Create and shape the model
First create broad capability with data and compute; then shape the response policy with demonstrations, preferences, rewards, and safety data.
You unlockAudit open pre- and post-training recipes
Turn weights into a dependable system
Learn generation and memory before building retrieval, agents, evaluation, security, and operations around the model.
You unlockDesign and operate an evidence-grounded service
Choose an advanced branch
Distillation, LoRA, MoE, multimodality, and interpretability reuse the shared core but do not form one artificial dependency chain.
You unlockInvestigate the specialization relevant to your goal
Inside the course
One course. 7 connected frontiers.
Lessons 01–40 form the cumulative foundations-to-deployment spine. Lessons 41–44 are parallel advanced specializations: choose them for your goal rather than treating their order as a prerequisite chain.
01
Explain the learning loopFoundations
Build an intuitive account of representations, feedback, responsibility, and model updates before choosing the optional mathematics.
The payoffExplain one complete learning step and diagnose where it can fail.
02
Inside the machineArchitecture
Turn text into predictions, one transparent mechanism at a time.
The payoffTrace every tensor from text to next-token logits.
03
Build the base modelPre-Training
Learn how data, objectives, compute, and evaluation create capability.
The payoffDesign and audit a base-model training run.
04
Shape useful behaviorPost-Training
Transform a text predictor into a helpful, safer assistant.
The payoffChoose supervision and controls for an assistant behavior.
05
Run the modelInference & Serving
Control generation, memory, latency, throughput, precision, and reasoning budgets.
The payoffBuild an evidence-based serving and decoding policy.
06
Build dependable systemsApplications & Reliability
Design context, retrieval, agents, evaluation, security, and production operations.
The payoffShip a bounded, evaluated, observable LLM system.
07
Choose a specializationAdvanced
Compress, adapt, route, connect modalities, and investigate modern models.
Explore when relevantFollow the specialization that matches your model or research goal.
44connected lessons
44hands-on labs
40code notebooks
No black boxes. Build intuition, see the mechanism, then make the real engineering trade-offs.
Ready when you are
Leave with a working mental model of the entire LLM stack—and the judgment to use it well.
44 connected lessons, hands-on labs, and a complete end-to-end build.