Skip to content
AI LearningAtlas

Open learning system · Updated continuously

Understand the whole
LLM system.

Build a connected mental model across deep learning, transformers, training, GPU execution, distributed systems, serving, and evaluation.

Start the learning path Browse resources

English-primary guidance with Simplified Chinese summaries and client-side bilingual search.

SYSTEM ARCHITECTURE LLM system stack
  1. Serving Inference · scheduling · SLOs
  2. Model & training Architecture · data · optimization
  3. Runtime Kernels · compilers · distributed
  4. Hardware Compute · memory · network
Higher layers depend on the layers below; evaluation continuously tests assumptions across the stack.
8connected modules
A–Csource authority ratings
EN + 中文bilingual discovery
$0required to begin

LEARNING PATH

Eight modules. One learning route.

Move from first principles to production decisions. Every module ends with checkpoints, self-tests, and a concrete artifact.

01 Fx Foundations

Reason with tensors, gradients, probability, and measurable experiments.

Build the base → 02 NN Deep Learning

Train and diagnose neural networks with deliberate optimization choices.

Study the loop → 03 Attn Transformers

Connect tokenization, attention, architecture, and representation.

Trace attention → 04 LM LLM Training

Design data, pretraining, post-training, and alignment gates as one governed pipeline.

Plan a training run → 05 SLO Inference & Serving

Balance KV cache, quantization, batching, throughput, latency, and reliability.

Budget a service → 06 GPU Kernels & Compilers

Connect hardware execution to kernel correctness and performance evidence.

Optimize a kernel → 07 Dist Distributed Systems

Reason about collectives, topology, parallelism, and failure boundaries.

Design the topology → 08 Eval Evaluation

Build reproducible comparisons that expose quality, safety, and limitations.

Prove what works →

LATEST HOTSPOTS

Fresh signals, without the hype.

New work from official creator and project feeds, sorted newest-first. Popularity is not authority; use each item as a lead, then inspect the evidence.

Loading the latest official feeds…

CURATED, NOT COLLECTED

Find the right material faster.

Browse by format, then return to the learning path for sequence and context. Canonical metadata records authority, access, and rights.

01CoursesStructured lectures, labs, and assignments. 02DocumentationPrimary references, guides, and tutorials. 03VideoTechnical demonstrations and lecture channels. 04PodcastsPractitioner context and current discussion. +Paid recommendationsClearly separated, disclosed, and paired with free alternatives.

BUILT FOR TRUST

Sources you can inspect.
Claims you can challenge.

The Atlas publishes original synthesis, keeps canonical provenance in structured data, and defaults unknown reuse rights to link-only. Corrections and limitations are part of the learning system.

Editorial policy → Source & rights policy →
01

Source-groundedOfficial and high-authority material is prioritized and rated.

02

Rights-awareFree access is never treated as permission to redistribute.

03

Evidence-ledCheckpoints turn reading into explanations, estimates, and experiments.

START WITH THE SYSTEM

Build knowledge that connects.

Take the diagnostic, choose one checkpoint, and create the first artifact in your full-stack LLM learning path.

Open Start Here
Explore connectionsGraph and backlinks

Backlinks

Graph View