AI Foundationspredict · compress · act

A FIELD GUIDE TO HOW MACHINES LEARN

Guess. Check.
Improve.

A machine makes a guess. It checks how wrong the guess was. Then it changes itself a little. Do that a few billion times and you get a model. This guide follows that one loop — from a single bit of information to an agent that acts.

Start the story →See the map

the whole field in one picture

One idea, seen six ways

Every topic in this guide is one of six jobs. They are not rivals. They are the same object looked at from six angles.

The same data, refracted into six jobs. A real topic can use several at once — that is why the field feels tangled until you see the prism.

the six jobs

What an AI is actually doing

RepresentWhat is the state?

Turn the world into numbers a machine can work with.

tokens, embeddings, world models

InferWhat follows?

Work out what is likely, given what it knows.

Bayes, causality, retrieval

LearnWhat should change?

Get a little less wrong by looking at data.

gradient descent, scaling laws

Search & planningWhat sequence?

Look through many options and keep a good one.

planning, MCTS, theorem proving

ActWhat to do now?

Choose an action and live with the result.

reinforcement learning, tools, agents

MeasureHow good, and how much?

Put a number on surprise, error, and cost.

entropy, cross-entropy, evaluations

the stack

Eight layers, from maths to institutions

A topic lives at one layer and is looked at through one or more jobs. The lower layers change slowly. The upper layers change monthly.

L0Foundations & theoryThe maths underneath everything.probability, optimisation, information

L1ModelThe shape of the thing that learns.transformer, CNN, diffusion, mixture of experts

L2DataWhat it reads.curation, synthetic data, licensing

L3ComputeWhat runs it.GPUs, memory bandwidth, power

L4TrainingHow it is taught.gradients, fine-tuning, RLHF

L5Serving & deploymentHow it answers once it is trained.decoding, batching, quantisation

L6Systems, agents & applicationsWhat it does for someone.tools, memory, sandboxes, coding agents

L7Ecosystem & institutionsWho builds it and who governs it.labs, open weights, policy

the map

Where everything sits

Six jobs across, eight layers down. Each box is filled from the graph — the terms are the concepts, and every one links to the chapter that explains it. Dashed boxes are open ground. Grey boxes do not apply at all — a GPU does not infer, and an institution does not learn.

Represent

Infer

Learn

Search & planning

Act

Measure

L0 Foundations & theory

Compression Prefix code Huffman coding Error-correcting code Vector Matrix Embedding space Frame problem

Probability Bayes' rule LikelihoodBayesian inferencePrediction Channel Prior Posterior

Loss function Generalization Overfitting No free lunch Bias–variance tradeoff Inductive bias Occam's razor Feedback

Turing machine Computability Halting problem Search Combinatorial explosion

Markov decision process

Bit Entropy KL divergence Divergence Channel capacity Source coding theorem Perplexity Intelligence Uncertainty Information Surprise Rate Noise Mutual information Dot product Turing test

L1 Model

Query, key, value Language model Reasoning model

Perceptron Backpropagation Weight Bias Multilayer perceptron Sequence model word2vec Value function Sparse autoencoder JEPA

open

Policy

Capacity Emergent ability

L2 Data

Token Byte-pair encoding Chunking Speech token

open

Contamination

L3 Compute

n/a

Memory bandwidth Interconnect Memory wall

L4 Training

Denoising

open

Exploration

Reward Model-based RL

Scaling law Compute-optimal

L5 Serving & deployment

KV cache

Classifier-free guidance Inference Batching Speculative decoding Chain of thought

open

Full-duplex

Test-time compute Inference scaling

L6 Systems, agents & applications

open

Retrieval-augmented generation Semantic search

open

Planning

Reinforcement learning Agent Agent harness Tool use Agent loop ReAct Code execution Workflow Durable execution Checkpointing Permission Least privilege Computer use Context management Loop policy

Benchmark Red teaming Human in the loop Tracing Drift Guardrail

L7 Ecosystem & institutions

n/a

Openness ladder Model family Fully open model Commoditisation Barbell

swipe to scroll →

Colour marks the job: cyan for information, violet for learning, magenta for decision. Dashed = open. Grey = not applicable.236 concepts, generated from one graph.

Open the full atlas →Every concept →Reading routes →

how we got here

Eighty years, in one line

Two winters, two thaws, and one architecture that ate the field. Twelve turns that explain the shape of the present — and why the money ran out twice before it worked.

1936The Turing machineAlan Turing asks what a machine can compute at all — and finds limits.1943The artificial neuronMcCulloch and Pitts model a brain cell as simple arithmetic.1948Information theoryShannon measures information in bits. This guide starts here.1956Dartmouththe term "artificial intelligence" is coined 1969Minsky & Papertpublish "Perceptrons"; perceptron funding stalls 1973Lighthill reportUK AI funding collapses 1986BackpropagationRumelhart, Hinton and Williams show networks can learn their own hidden features.1997LSTMA gated loop remembers over long stretches, easing the vanishing gradient.2012AlexNetthe thaw, and the boom that never stopped 2017The transformer"Attention Is All You Need" replaces recurrence with attention.2022ChatGPTPreference tuning turns a text model into a product. The public arrives.2025DeepSeek-R1Open reasoning weights land, with a cost shock and a new rivalry in the open.

swipe to scroll →

See all 54 moments →

open ground

What we still don't know

These are not footnotes. They are the live edge of the field.

Will more compute keep working, or hit a wall?
Why do some skills appear suddenly as models get bigger?
Can we see what a model is really doing inside?
How do we know an answer is true, not just fluent?
What happens when the training data runs out?
Can a model keep learning without forgetting?
How do we make AI want what we actually meant?
Who is responsible when an agent acts?

the story

54 chapters, in order

If the map is the territory, this is the walk across it. Each chapter is short, plain, and carries a diagram. Prefer the arc by year? See the timeline →

act I · The Core Question

01What Is Intelligence, Really?intelligence · agent

02Intelligence Is Compressioncompression · prediction

act II · The Substrate

03The Bit: A Yes or a Nobit · uncertainty

04Entropy: The Price of Surpriseentropy · surprise

05Codes: Saying More with Lesssource coding · prefix code

06Channels: Pushing Through Noisechannel · noise

07Distance Between BeliefsKL divergence · mutual information

act III · The Instruments

08Belief, Updatedprobability · Bayes' rule

09Everything Is a Vectorvector · dot product

10Rolling Downhillloss function · gradient descent

11Why Memorizing Failsoverfitting · underfitting

12No Free Lunchinductive bias · no free lunch

act IV · The Machines

13Can Machines Think?Turing machine · computability

14The First Neuronneuron · perceptron

15Why Silicon Got Good at Thisparallelism · GPU

act V · The First Attempts

16Rules All the Way Downsymbolic AI · knowledge representation

17The First WinterAI winter · combinatorial explosion

18Feedback: The Other Treefeedback · control theory

act VI · The Connectionist Turn

19Learning from Blamehidden layer · multilayer perceptron

20Seeing with Windowsconvolution · convolutional neural network

21Memory in a Looprecurrent neural network · LSTM

22Words as Coordinatesword embedding · word2vec

23Learning from Rewardreinforcement learning · policy

act VII · The Attention Revolution

24Attention Is All You Needattention · transformer

25Breaking Language into Piecestoken · tokenizer

26The Big Readpretraining · self-supervised learning

27More Is Differentscaling law · compute-optimal

28Teaching Tastefine-tuning · instruction tuning

29Noise into Imagesdiffusion model · denoising

act VIII · The Alignment Layer

30Opening the Black Boxmechanistic interpretability · feature

31Getting the Goal Rightalignment · specification

32How Do We Even Know?benchmark · evaluation

act IX · The Agent Infrastructure

33From Answer to Actionagent loop · ReAct

34Hands and Function Callsfunction calling · tool schema

35Memory Outside the Weightsretrieval-augmented generation · vector database

36A Port for ToolsModel Context Protocol · MCP server

37Many Hands, One Joborchestration · workflow

38How Tokens Get Servedinference · KV cache

39A Safe Place to Actsandbox · permission

40Watching the Looptracing · observability

act X · The Frontier

41Thinking Before Answeringchain-of-thought · reasoning model

42Models of the Worldworld model · model-based RL

43The Open QuestionAGI · superintelligence

act XI · The Open-Weight World

44What 'Open' Meansopen weights · open-source AI

45DeepSeekmixture of experts · multi-head latent attention

46The Chinese Labsmodel family · Qwen

47The Other Open LineagesLlama · Mistral

48The Compute Squeezeexport controls · Huawei Ascend

49Licences and Rulesopen-weight licence · acceptable use policy

50Running One YourselfGGUF · LoRA

51The Open Frontiercommoditisation · open-weight risk

act XII · Beyond Text

52The Harnessagent harness · harness engineering

53Image Modelslatent diffusion · diffusion transformer

54Omni-Modelsomni-model · any-to-any