MLPL Literate Walkthroughs

Org-mode documents whose MLPL source blocks are live: each was run top-to-bottom and its real output baked in. The 1:1 companions to the playground demos — especially the connect-only ones the public browser demo can't run.

• Tic-Tac-Toe, end to end, in MLPL CUDA · NVIDIA GPU

  The whole game in the array language — winner as a matmul, alpha-beta minimax, a self-play training set — then a board → move policy LoRA-fine-tuned on an NVIDIA GPU via device("cuda"). Proven by playing: a before/after waffle. No domain-specific builtins.

• CUDA LoRA fine-tune (true-GPU walkthrough) CUDA · NVIDIA GPU

  LoRA-fine-tune a tiny language model on an NVIDIA GPU via candle — forward, backward, and Adam all on-device — with the loss curve falling, parity-tested against the CPU path within fp32 tolerance.

• Tic-Tac-Toe, end to end, in MLPL MLX · Apple GPU

  The whole game in the array language — winner as a matmul, alpha-beta minimax, a self-play training set — then a board → move policy LoRA-fine-tuned on the Apple GPU. Proven by playing: a before/after waffle (30 losses → 2). No domain-specific builtins.

• MLX LoRA fine-tune (true-GPU walkthrough) MLX · Apple GPU

  LoRA-fine-tune a tiny language model on the Apple GPU — forward, backward, and Adam all on-device — with a concrete before/after perplexity drop.

• Basics: a literate REPL tour CPU · runs anywhere

  The playground's Basics demo as a literate program: arrays, broadcasting, and visualization across nine shared-session source blocks.