labs

> the live galaxy

Inside theIntelligence galaxy.

10 constellations, 2,143 charted planets, 290 on the required route, one continuous path: from the linear algebra and calculus every model assumes, through optimization and the deep-learning stack, to the transformer and the language models and agents it runs. This is the route to modern AI, built from the floor up.

Each constellation below is a region of this galaxy. Every planet is mapped and prerequisite-gated. Lessons are being authored along the required route and open as they pass review. Underneath, an engineering track turns the theory into code you actually ship.

> 10 constellations · 2,143 planets · 290 required · 2,196 lessons

The map,end to end.

Each card is a constellation: a self-contained body of theory and practice with a fixed lesson count, sequenced so nothing arrives before its prerequisites. Tap any card to expand its systems and lessons. Start at 01 and walk the cards, or enter the map and let the diagnostic place you.

> real-world stack

Tools you’llactually use.

By the time you finish, you’ll have shipped real code in the tools that production teams actually use. Not abstract pseudocode: the concrete stack.

  • NumPyLinear algebra. Tensors. The bedrock.
  • PyTorchYour main framework. From scratch to production.
  • JupyterThe notebook you’ll actually live in.
  • Hugging FacePretrained models, datasets, hub.
  • JAXFunctional autodiff. Research-grade speed.
  • CUDAGPU intuition. Why kernels matter.
  • W&BTrack experiments. Stop losing runs.
  • FastAPIServe your model. Deploy or perish.
  • DockerReproduce environments. Ship anywhere.
  • GitVersion control. Branch. Diff. Revert.

> engineering track · 9 projects

Theory is a map.Code is the territory.

Nine projects, built from scratch in a notebook a real engineer would write. Each one binds to specific systems: you derive it, then you ship it.

  1. 01foundationnext

    Linear Regressor in NumPy

    Build a linear regressor end-to-end: fit by normal equations, then by gradient descent. Plot residuals, check assumptions, beat a baseline.

    [ stack · NumPy ]
    • NumPy implementation, no scikit-learn
    • Train/val/test split
    • Residual analysis script
    • Comparison vs. sklearn ground truth
    theory 01 · 02project 01
  2. 02foundationnext

    Logistic & Softmax Classifier from Scratch

    Binary then multiclass classification by hand: derive the gradient, train with mini-batches, plot decision boundaries.

    [ stack · NumPy ]
    • Binary logistic on 2D toy data
    • Softmax on Fashion-MNIST
    • Numerical gradient check
    • Confusion matrix + calibration plot
    theory 01 · 02 · 03project 02
  3. 03intermediatenext

    Optimizer Lab

    Implement SGD, Momentum, RMSProp and Adam from scratch. Race them on convex and non-convex toy problems and a small MLP.

    [ stack · NumPy ]
    • Each optimizer as a 30-line class
    • Convergence curves on Rosenbrock & quadratic bowl
    • Train an MLP four ways
    • Ablation report
    theory 03project 03
  4. 04intermediatequeued

    Neural Network in 100 Lines

    A feedforward network with explicit backprop, no autograd. Train it on MNIST and explain every line.

    [ stack · NumPy ]
    • Single-file MLP in < 120 lines
    • Backprop derivation matching the code
    • 95%+ MNIST test accuracy
    • Initialization + activation ablation
    theory 01 · 03 · 06project 04
  5. 05intermediatequeued

    Tiny Autograd

    Build a minimal reverse-mode autodiff engine: Value class, computational graph, .backward(). Then re-use it to train the same network as project 4.

    [ stack · pure Python ]
    • Value class with +, *, **, exp, log, ReLU, tanh
    • Reverse-mode .backward()
    • Same MLP rewritten on top of it
    • Side-by-side gradient match with PyTorch
    theory 01 · 06project 05
  6. 06intermediatequeued

    Convolutional Network from Scratch

    Code Conv2D, pooling, batch norm and a residual block by hand. Train a small ResNet on CIFAR-10.

    [ stack · PyTorch (no nn.Conv2d) ]
    • Conv2D forward + backward without im2col
    • BatchNorm with train/eval modes
    • Residual block module
    • 60%+ CIFAR-10 with no pretrained weights
    theory 06 · 07project 06
  7. 07capstonequeued

    Mini-Transformer

    Implement scaled dot-product attention, multi-head attention, a transformer block. Train a tiny language model on Tiny Shakespeare.

    [ stack · PyTorch (low-level) ]
    • Single-file decoder-only transformer
    • KV-cache for inference
    • Greedy + temperature sampling
    • Loss curve + generated samples
    theory 06 · 07 · 08project 07
  8. 08capstonequeued

    Build a Small GPT

    Scale up the mini-transformer. Train a 10M-parameter GPT on a curated corpus and serve it locally.

    [ stack · PyTorch ]
    • BPE tokenizer from scratch
    • ~10M-param GPT trained on a custom corpus
    • Inference server (FastAPI)
    • Eval on a small held-out suite
    theory 01 · 06 · 07 · 08project 08
  9. 09capstonequeued

    Diffusion from Scratch

    Implement DDPM training and sampling. Train on a small image set and watch noise resolve into pictures.

    [ stack · PyTorch ]
    • Forward noise schedule
    • Small U-Net denoiser
    • Train on a 10k-image set
    • Sampling notebook with intermediate steps
    theory 04 · 06 · 07 · 09project 09

> 10 · final boss · sealed

????????

??? ??? ??? ??? ??? ??? ??? ??? ??? ???. Unlocks only after the ninth capstone. We will not say more.

[ stack · ??? ]capstone · capstone · capstonetheory · ??

> frontier

Machine learning is the live route of the Intelligence galaxy, one of five on the map. The moment you understand it from the ground up, jump points open. The same structures keep reappearing elsewhere on the map.

Linear algebra→deep learning, dynamical systems, statistical physics, quantum mechanics.

Backpropagation→automatic differentiation, optimization, control, computational graphs.

Attention→information retrieval, kernel methods, graph neural networks, cognition.

Generative models→thermodynamics, information theory, statistical mechanics, language.

One galaxy at a time.
The map keeps expanding.

enter the mapback to the universe