Exploratory work — notebooks written to learn machine learning concepts by building them from scratch rather than reading about them.
Nothing here is a library or a product. It's a sandbox.
-
research-notebooks/01-chain-of-thought-prompting.ipynb— an experimental replication lab for Wei et al. (2022): run direct, chain-of-thought, and ablation prompts on a deterministic GSM8K sample with an open Qwen model; measure accuracy, uncertainty, tokens, latency, GPU memory, and errors. Designed for a free Colab T4; no API key required. See the research-notebooks index. -
math_of_ml_screensaver.ipynb— the math behind ML, worked up from dot products and gradient descent to a tiny MLP trained by hand (no autograd). The learned scalar field becomes a divergence-free flow, and particles running through it produce a generative screensaver. A final section scales the same idea to a GPU CPPN in PyTorch. -
epistemic_vs_aleatoric.ipynb— the two kinds of uncertainty in a prediction: what the model doesn't know yet, versus what nothing could know.
The .mp4 and .png files are rendered output from the screensaver notebook.
python -m venv .venv && source .venv/bin/activate
pip install numpy matplotlib jupyterlab
jupyter labThe first research-paper lab installs its pinned-range Hugging Face dependencies in Colab and requires an NVIDIA GPU for the default 3B-model experiment.
The GPU section additionally needs torch and an NVIDIA GPU. MP4 export uses
ffmpeg if it's on PATH, and falls back to GIF if not.