Code
Implement the object with tests, shape checks, seed control, and enough structure that another person can run it.
Practice
Practice and Projects
Last updated: April 22, 2026
The standard
| Question | Output |
|---|---|
| State the assumptions | Written assumptions, limits, and conditions |
| Build or derive it | Code, derivation, or worked calculation |
| Compare with a baseline | Metric, sanity check, or baseline model |
| Break it once | Failure case, ablation, or counterexample |
| Write down the result | Short note with results and limitations |
Paper study
Identify the claim, assumptions, comparison point, result, limitations, and one small reproduction or ablation when possible.
Systems
Measure memory, throughput, latency, data movement, or failure modes. Make system behavior visible with measurements.
Explanations
Turn a concept into a diagram, derivation, misconception check, or teaching note that exposes where the idea can fail.
Language models from first principles
The first flagship project path connects mathematical foundations, implementation, evaluation, and systems in a small language-model build that can be run and inspected.
| Stage | Output |
|---|---|
| 1. Tokenizer | Tests + benchmark |
| 2. Bigram baseline | Baseline loss/perplexity |
| 3. Transformer | Shape/mask/attention checks |
| 4. Training loop | Checkpointing + sampling |
| 5. Data pipeline | Filtering + dedup notes |
| 6. Evaluation | Baselines + uncertainty |
| 7. Ablation | One controlled change |
| 8. Systems profile | Memory/throughput/KV cache |
| 9. Report | Results, limits, next experiment |