Learning path · AI · 39 layers

AI GOD flow: the whole map

Read left to right, pan down. Fundamentals first — ML, depth, domains, the train loop. Generative below — LLMs, using models, knowledge beside customize, then applications. Two trap bands. Click Transformer above and Transformers (LLM) below — same block, two altitudes.

Click a shape or a question

Questions the whole tree answers

From a labeled row to a traced agent

Artificial intelligence — the whole map

Three lanes — signal · depth · sensor

Machine learning

Deep learning

AI domains

Classic loop and traps

Classic loop — data → serve

don’t · fundamentals

Generative AI

Inside the call — tokens through the window, then the API

LLMs

Using models

Ground, then specialize

Adding knowledge

Customizing models

AI applications

don’t · generative

rootArtificial intelligence

ML · DL · domains · then generate

fitMachine learning

examples → function
loss · hold out a test set

labeledSupervised learning

x → y you named
labels are the ceiling

structureUnsupervised learning

no target
you still name the clusters

rewardReinforcement learning

act · delayed reward
sim or preferences

depthDeep learning

stacked layers · learned features
still a split and a loss

unitNeural networks

Wx + b · σ · stack
forward / backward

gridCNN

shared filters on space
vision’s inductive bias

chainRNN / LSTM

hidden state walks time
attention replaced most NLP

blockTransformers

QKᵀV over the window
same block as the LLM below

sensorAI domains

same math · different I/O
metric is part of the domain

textNLP

tokens in
label or generate — still measure

pixelsComputer vision

class · box · mask
mAP, not a blog accuracy

waveSpeech

ASR / TTS
WER and a latency budget

rankRecommendation / prediction

retrieve then rank
offline metric vs A/B

01Data

labels · no leakage

02Features

train-serve skew

03Train

val · versioned ckpt

04Evaluate

baseline + slices

05Serve

budget · rollback

rejectedDeep learning is magic

depth without a split

rejectedSkip labels, skip split

then you cannot know

sampleGenerative AI

draw from a distribution · fluent ≠ true

lmLLMs

next token at scale
weights are not your wiki

piecesTokens

subwords · the budget
count ids, not words

vectorEmbeddings

near = similar in this space
version the model

mixAttention

QKᵀV over the window
KV cache is the systems tax

stackTransformers (LLM)

same block as above, stacked
causal decode

windowContext window

finite attention span
not a disk

apiUsing models

call · constrain · loop
no cluster required

specPrompting

system / user / examples
version + test

schemaStructured output

JSON schema · parse
fail closed on money

fnFunction / tool calling

model proposes · you execute
IAM lives here

pixelsMultimodal

image / audio in context
measure extraction

corpusAdding knowledge

facts outside the cutoff
attach · cite

indexVector database

ANN + metadata ACL
reindex on embedder change

searchSemantic search

paraphrase recall
hybrid with keywords

loopRAG

retrieve then generate
failures are retrieval first

weightsCustomizing models

climb the ladder
prompt+RAG first

sftFine-tuning

your distribution
not a CMS for facts

adapterLoRA / PEFT

small delta · freeze base
swap adapters

pretrainTraining from scratch

lab · cluster
almost never the product path

productAI applications — the model is one box

state · tools · policy · measurement

loopAgents

cap the steps

stateMemory

you store it

handsTools

verbs + IAM

protocolMCP

USB for tools

policyGuardrails

untrusted model

gateEvaluation

goldens · graders

traceObservability

stack of a generation

rejectedFine-tune first

skip prompt · skip RAG

rejectedShip on vibes

no goldens · no traces

72%

Click a box · three arrows off a layer are what it allows

What AI is