Skip to main content

The atlas on one map

Everything in the atlas on one mind map: the layers and their nodes, and under each node the papers, people, labs and milestones it connects to. Every box opens its section.

HOW IT WORKS · L0–L3WORKING WITH IT · L4–L6Knowledge atlasL0Math and CSfoundations−Math: learn it when you need itMathLinear algebraLinear algebraProbability and StatisticsProbability and StatisticsInformation TheoryInformation TheoryOptimizationOptimizationL1ML and deep learningbasics−Classical machine learning: learning from examplesClassical machine learningLoss and backpropagationLoss and backpropagationPyTorch: train a network yourselfPyTorchGeneralizationGeneralizationClassic architectures: CNN, RNN / LSTM, residualsClassic architecturesL2LLM core−LLM: how the model worksLLMTokenizerTokenizerTransformerTransformerPretraining and scalingPretraining and scalingFine-tuning: teach it your taskFine-tuningPost-training: shaping with rewardsPost-trainingDecoding and inferenceDecoding and inferenceLong context and MoELong context and MoEMultimodalMultimodalL3Systems andinfrastructure−GPU basics: why it's used, and where the bottleneck isGPU basicsServing: running models yourselfServingParallel training: why one card is not enoughParallel trainingThe economics of compute: where training money goes, how inference is pricedThe economics of computeKubernetes: running it at the customer's siteKubernetesL4Applicationengineering−Calling a model API: token, streaming, tools, costCalling a model APIPrompting: instructions you can measurePromptingStructured output: making the model's answers consumable by programsStructured outputEmbedding: turning meaning into vectorsEmbeddingRAG: Look Up First, Then AnswerRAGAgent: a model with tools in a loopAgentMCP: One Standard Socket for Every SystemMCPEvals: knowing whether it got betterEvalsObservability: know what is happening in production, and feed it back into evalsObservabilityCost: every call has a price tagCostSecurity: the model acts on text it should not trustSecurityTypeScript: the language of front ends and SDKsTypeScriptL5The field, part I:history−AI History: How to Read the Three WavesAI HistoryThe first wave: symbols and expert systemsThe first waveThe Second Wave: Statistical Learning and Deep LearningThe Second WaveThe third wave: scaling and LLMThe third waveL5The field, part II: thefrontier−Weekly reads: the frontier feed (≤ 5)Weekly readsLook up when needed: frontier reference list (≤ 15)Look up when neededOpinions and predictionsL6Careers and themarket−Role families: how AI jobs are dividedRole familiesJD requirements: how to read the skill wordsJD requirementsPay: record sources, not numbersPayProfiles−AI Product Manager (sketch)AI Product Manager(sketch)FDE / AI EngineerFDE / AI EngineerInference and training infrastructure (sketch)Inference and traininginfrastructure (sketch)ML Engineer (Sketch)ML Engineer (Sketch)More in the atlas−Amazonlab or company · 1994AmazonMetalab or company · 2004MetaCursorlab or company · 2022Cursor

Drag to move. Pinch, or hold Ctrl or ⌘ and scroll, to zoom. A circle with a number opens its branch; Shift-click opens everything under it. Levels open the whole map to that depth (or press its number).

core: other nodes build on it, and the profiles' paths run through itHover or tab to a node to trace it:what it needswhat it opens uplinked

Translated from the author's Chinese notes by a model; the Chinese page is the original.