Home · Browse

Everything on the site, in one place

All 40 sections grouped by theme, the timeline they sit on, and the order they depend on each other in. ⌘K if you already know what you want.

All 40 sections Roadmaps Labs Timeline Learning order Scope words Beyond this site

The learning graph

All 40 sections

Grouped by theme. The numbers are a reading order, not a ranking.

Foundations

Math & Stats Foundations

Linear algebra, probability, information theory.

Full

Classical Supervised ML

Regression, trees, boosting, SVMs — still wins on tabular.

Full

Unsupervised & Self-Supervised

Clustering, VAEs, contrastive learning. No labels.

Full

Neural Nets & Vision

Neural Network Fundamentals

Backprop, activations, normalization, residuals.

Full

CNNs & Vision Foundations

Convolutions, ResNet, detection, segmentation.

Full

Modern Vision Foundation Models

CLIP, SAM, ViT-based VLMs.

Full

Sequence, Attention & LLMs

Sequence Modeling Pre-Transformer

RNNs, LSTMs, seq2seq — the bottleneck attention removed.

Full

Attention & Transformers

Query, key, value. The mechanism behind almost everything.

Full

Bag-of-words to BERT to instruction tuning.

Full

LLM Architecture & Training

Decoder-only stacks, scaling laws, RLHF and DPO.

Full

Generative & Multimodal

RAG, Agents & Reasoning

Why LLMs hallucinate, why retrieval helps, how agents plan.

Full

Generative Models

GANs, VAEs, diffusion.

Full

ASR, TTS, speaker ID, real-time voice.

Full

Fusing vision, audio and text into one system.

Full

Decision & Retrieval Systems

Reinforcement Learning

MDPs, Q-learning, PPO — and what RLHF borrows.

Full

Recommenders, Ranking & Search

Retrieval, ranking, rerank — and the feedback loop that poisons it.

Full

Time Series & Forecasting

Why you cannot shuffle. Stationarity and honest backtests.

Full

Message passing as the one primitive.

Full

Embodied & Frontier

Scientific & Structured AI

Equivariance, protein structure, PINNs.

Full

3D, Spatial AI & Autonomous Driving

SLAM, BEV, sensor fusion, closed-loop eval.

Full

Robotics & Embodied AI

VLA models, sim-to-real, diffusion policies.

Full

Learned simulators and compounding error.

Full

Systems, Safety & Interview

Efficient AI & Systems

FlashAttention, quantization, parallelism.

Full

ML Engineering & MLOps

Feature stores, drift, canary deploys.

Full

Evaluation, Reliability & Safety

Calibration, hallucination, red-teaming, LLM-as-judge.

Full

2026 Frontier Map

Reasoning models, agents, world models.

Full

Interview Mastery

Question types, drills, how to structure an answer.

Full

GPU Architecture, CUDA & Distributed Training

SIMT, warps, memory hierarchy, ring all-reduce.

Full

Hands-on

Fine-tuning LLMs in Practice

When not to. LoRA vs QLoRA by VRAM.

Full

Running Models Locally

Compute the VRAM yourself, then pick a format.

Full

Preference Optimization: RLHF, DPO & GRPO

Reward models, the DPO derivation, GRPO.

Full

Inference & Serving

LLM Inference & Serving

Prefill vs decode, TTFT, continuous batching.

Full

LLM Inference at Scale

Disaggregation, KV quantization, multi-LoRA, FP4.

Full

Real-Time Voice AI

The turn-latency budget, endpointing, barge-in.

Full

Diffusion & Video Inference

Cutting the sampling loop, caching, video cost.

Full

The Labs about the labs →

Forward, backward, then gradients checked numerically.

Build a BPE vocabulary merge by merge.

Seven attention mechanisms on one worked example.

Temperature, top-k and top-p on the same row.

Quantization Lab

INT8 error cell by cell, and what one outlier costs.

Model Atlas

Model Atlas: Language Models

Frontier, open-weight and on-device LLMs compared.

Full

Model Atlas: Speech & Audio

STT, TTS, separation, diarization — with the metrics explained.

Full

Model Atlas: Vision & Generative Media

VLMs, detection, OCR, image and video generation.

Full

Grand Challenges

The Millennium Prize Problems

Seven problems, $1M each, six still open.

Full

AI in Industry: What It Solves

AlphaFold's Nobel — and the claims that did not survive.

Full

00 · The timeline

Seven eras, 1950s → 2026

Scroll sideways.

1950s – 1970s

Symbolic roots

Perceptron
Symbolic AI & expert systems
Bayesian foundations
Classical control theory

1980s – 1990s

The statistical turn

Backpropagation popularized
Decision trees
SVMs
HMMs & graphical models

2000s

Kernels & ensembles

Kernel methods
Boosting & random forests
CRFs
Matrix factorization

2012 – 2017

Deep learning breaks out

AlexNet
Word2Vec / GloVe
Seq2seq + attention
GANs, ResNet

2017 – 2020

The Transformer era begins

Transformer, BERT, GPT
ViT
Contrastive learning
Diffusion beginnings

2020 – 2023

Scale and alignment

Scaling laws
Instruction tuning, RLHF
Multimodal foundation models
Diffusion explosion

2023 – 2026

Reasoning & agents

Reasoning models, tool use
Long context, efficient attention
Multimodal-native models
Video/world models, VLA robotics

Suggested path

Core learning order

If you are starting cold, this is the dependency order.

Math→ Classical ML→ Unsupervised→ NN Fundamentals→ CNN/Vision→ RNN/Seq2seq→ Attention→ Transformers→ NLP/LLMs→ Generative→ Multimodal→ Speech→ RL→ RAG/Agents→ Rec/TS/Graph→ 3D/Spatial→ Robotics→ World Models→ Efficient AI→ MLOps/Eval/Safety→ 2026 Frontier

Vocabulary check

AI vs ML vs deep learning vs foundation models vs agents

Nested, not synonymous.

Scope, nested from broadest to narrowest

AI — any system that performs tasks we'd call intelligent, including hand-coded expert systems, no learning required. ML — a subset that learns its behavior from data. Deep learning — a subset of ML using multi-layer neural networks specifically. Foundation models — a subset of deep learning: large models pretrained on broad data, adaptable to many downstream tasks. Agents aren't a subset at all — they're a system built around a foundation model, wiring it to tools, memory and a planning loop.

Beyond this site

Watch, practice, discuss

Creators worth following.

Creators worth following

Priyam Mazumdar — Exploratory Data Adventures

PyTorch and deep learning built up from the ground up, aimed squarely at demystifying the field.

Transformers, LLaMA, and modern architectures implemented from scratch, line by line.

Reads the actual papers, start to finish, and unpacks the parts that don't hold up.

Plus 3Blue1Brown, StatQuest, and Andrej Karpathy — linked directly from the "Go deeper" section of the pages they're most relevant to.

LeetCode, but for ML — implement linear regression, attention, backprop, and dozens more from scratch, implemented from scratch.

Start solving →

My Notes — Browse

Free notes

Highlights on this page