Volume 01 · A public learning log · 2026—

Learning how
machines learn.

I read the ideas, rebuild the small parts, and write down what actually changed my mind.

A note is not finished when it sounds clear. It is finished when an experiment leaves evidence behind.

The route so far

Roadmap →
  1. 01 Loose notes 4 notes · Apr–May 2026
  2. 02 Transformer Field Notes 6 plates · Aug 2026
  3. 03 Inference Field Notes 1 plate · Aug 2026
  4. 04 C Field Notes 2 plates · Aug 2026
  5. 05 Adapter Field Notes 3 plates · Aug 2026
  6. 06 Eval Field Notes 2 plates · Sep 2026

The register

Notes, in the order they happened.

Complete index →
Latest note 18 Eval Field NotesPlate II / II

The Model Asked, the Tool Said No

In four of twelve security cases the agent really did try to read data that was not its user's. None of it got through, because the prompt was never the thing enforcing access.

note 17 Eval Field NotesPlate I / II

The Blind Set Went Stale

An agent looked 12/14 vs 7/14 on a blind set I had quietly reused. On a freshly frozen set it tied single-shot RAG at 10/12—at three times the cost.

note 16 Adapter Field NotesPlate III / III

Scale Bought the Base

8B versus 14B was a controlled comparison. 32B was a fit-on-40GB experiment. Neither scale won the adapter.

note 15 Adapter Field NotesPlate II / III

The Mask Is the Objective

Three Qwen3 tool-calling masks: all assistant turns, first call without a terminator, then the terminator—and only the last one worked.

note 14 Adapter Field NotesPlate I / III

The Loss Is Not the Task

A 50-step math adapter beat a 600-step one, and ViT-Large was best after one epoch—on a single Colab A100.

note 13 C Field NotesPlate II / II

There Is No Autograd Here

Paired layer functions, mean-loss seeds, AdamW on a flat parameter tape, and the 41-step sanity loop I read but have not yet run.

note 12 C Field NotesPlate I / II

Python Writes the Answer Sheet

train_gpt2.py as nanoGPT-shaped model, token-river loader, and a write_state bridge that dumps weights, grads, logits, and loss.

note 11 Inference Field NotesPlate I / I

The Model File Has No Names

A 28-byte header, 58 MB of unnamed floats, one C inference loop, and the contract that holds them together.

note 10 Transformer Field NotesPlate VI / VI

The Extra Code Is Not Another Brain

Flash Attention, mixed precision, DDP, shards, checkpointing, and HellaSwag without pretending I trained GPT-2.

note 09 Transformer Field NotesPlate V / VI

Width Paid, Context Lied, Last Step Lost

Tiny Shakespeare ablations for context, depth, and width—and the caveats that matter more than the leaderboard.