INTERACTIVE COURSE FAMILY
Learn mathematics the hands-on way — 0 → Hero
Five free, dependency-free, browser-native courses: every atomic concept ordered from zero to mastery, with live demos you operate (RREF steppers, eigen hunters, quantizers you tune live, refusal-direction surgery on a real micro model) — plus field context, motivation, formalisms, ELI5 rescues, exercises and quizzes per concept. No accounts. Progress lives in your browser.
01 · CORE MATH
Linear Algebra
213 atomic concepts across 18 stages — from fields and Gaussian elimination through eigen-decomposition, SVD, Jordan forms, to special matrices & the full identity catalog. Interactive RREF stepper, eigen-direction hunting, SVD sphere→ellipsoid theater in 2D and 3D.
0⊃fields → vectors → systems → matrix algebra → inverses → determinants → spaces → bases → maps → eigen → inner products → factorizations → spectral → duality/tensors → structure → special & identities → mastery
02 · NEURAL NETWORKS
Neural Networks
193 concepts across 16 stages — from single neurons through backprop, optimization (Adam/GRPO's siblings), CNNs, sequence models, transformers, embeddings, and generative modeling. Live training loops in your browser: watch networks actually learn.
0⊃math interface → neuron → MLPs → losses/optim → backprop → regularization → loss families → CNNs → sequences → transformers → embeddings → generative → scale train → interpret & safety → advanced → mastery
03 · DEPLOYMENT GRADE
Quantization
172 concepts across 12 stages — the complete craft of making neural nets small enough to matter: bit formats, uniform PTQ (GPTQ), rotations (QuaRot), codebook VQ, QAT and extreme 1–2 bit methods (BitNet lineage), with measured W4/W3/W2 numbers from the literature.
0⊃bits & walls → affine map → granularity → PTQ objectives → rotations/incoherence → codebooks → QAT → extreme low-bit → KV+attention → formats & deployment → failure science → mastery
04 · INTERPRETABILITY
Abliteration
120 concepts across 9 stages — residual-stream surgery on the refusal direction: geometry of the stream, the 2024 single-direction finding's discovery & lineage, runtime projection, permanent weight edits (incl. norm-preserving biprojection), variants, and honest accountability for what ablation breaks.
0⊃geometry → refusal phenomenon → direction finding → runtime ablation → permanent surgery → variants & soundness → damage science → ethics & safety → mastery
05 · ALIGNMENT ERA
Post-Training
156 concepts across 11 stages — the modern behavior-installation stack: SFT, Bradley-Terry reward models, canonical PPO-RLHF, the DPO family, RLVR/GRPO reasoning-era, stability engineering, LoRA anatomy, data engineering, and measurement of the behavioral delta.
0⊃frame → SFT → reward models → PPO-RLHF → DPO family → reasoning RL (GRPO) → stability → PEFT → data engineering → measurement → mastery
06 · TRAINING PHYSICS
Optimization
140 concepts across 10 stages — training as a dynamical system: gradient mechanics, heavy-ball physics, stochastic gradient physiology, Adam mechanics, schedules & Edge-of-Stability, Grokking, muP, Muon and the numerical anatomy of training runs.
0⊃frame → GD mechanics → SGD physiology → adaptive methods → schedules & stability → geometric anatomy (muP/Grokking) → numerics of backprop → new-era optimizers (Muon/Schedule-Free) → measurement science → mastery