Demos and Learning Paths

Demos and Learning Paths

The demos are executable documentation. In the browser, each is paired with an introduction, progress narration where needed, and a takeaway. Learning paths sequence concepts; the glossary provides searchable definitions tied back to MLPL constructs.

Representative demo families

Family	Examples and concepts
Data forge	Rejection sampling / best-of-N, graph multi-hop task generation with held-out regions, arithmetic curriculum with the interference lesson
Experiment quality	Robustness suite (one model, five conditions), scaffold dependence (train with hints, test without), Pareto frontier with the pareto_plot staircase
Language foundations	Basics, computation, matrix operations, APL2 structure, loop-avoidance (Thinking in Arrays: data loops vanish, time recurrences survive, scans absorb the associative borderline), errors, lenses
Classical ML	Logistic/softmax classification, k-means, PCA, kNN, Gaussian Naive Bayes, linear SVM + kernel trick, decision stump, voting ensemble, metrics playground
Neural networks	Tiny MLP, moons/circles MLP, gradient flow, architecture diagrams
Attention and language	Attention, transformer blocks, tiny LM training/generation, tokenization
Fine-tuning	CPU LoRA, MLX LoRA, CUDA LoRA, tic-tac-toe policy fine-tuning
Vision	Patchification, ViT attention patterns, multi-head comparisons, prediction galleries
Embeddings	PCA, t-SNE, UMAP comparisons, 2-D/3-D embedding views
Research ideas	Neural Thickets, perturbation, Engram conditional memory, feasibility estimation
Simulation and games	Tic-tac-toe, minimax/self-play, Game of Life, pattern zoo, glider gun
Visualization	Scatter, line, bar, heatmap, boundaries, 3-D, animated frames, model IR
Structural diagrams	Dataflow graphs with groups, weighted/highlighted edges, recurrence back-edges, and an SPM pipeline example
Integration	Remote MLX, connected GPU demos, LLM tool calls, analysis demo

Suggested learning route

flowchart LR
    Arrays[Arrays and shape] --> Algebra[Linear algebra]
    Algebra --> Classical[Classical ML]
    Classical --> Grad[Autograd]
    Grad --> Models[Model DSL]
    Models --> Train[Optimizers and training]
    Train --> Attention[Attention and tiny LM]
    Attention --> Tune[LoRA and quantization concepts]
    Models --> Vision[Convolution and vision]
    Train --> Research[Thickets, Engram, feasibility]

Loading

Running repository demos

From the CLI component workspace:

cd sw-mlpl/components/cli
cargo run -p mlpl-repl -- -f ../../demos/basics.mlpl
cargo run --release -p mlpl-repl -- -f ../../demos/tiny_lm.mlpl

Device demos require the appropriate feature, platform, and potentially a connected server. The web dropdown can mark a demo visible but disabled when the connected peer lacks its device.

Literate pages and generated assets

The repository also publishes literate HTML for selected demonstrations and contains a large set of architecture/flow diagrams under pages/diagrams. These complement, rather than replace, the interactive demo registry.

Browse the repository's demos directory, demo scripts guide, examples/literate, and published literate index.

The Engram Hash, Learnable Phrase Memory, and Tiny LM + Engram demos cover addressing, trainable memory, and integration with an attention model. The Thinking in Arrays demo contrasts removable element-by-element data loops with legitimate time or recurrence loops, then shows how associative cases become running_sum or running_product expressions.

For larger curricula and applications maintained outside the core repository, see Companion Repositories.