CPU-Safe Rules For NanoGPT Experiments

These rules are strict on purpose.

Allowed

• Tiny text files
• Small context windows
• Small batch sizes
• Short runs meant for learning, not for leaderboard chasing
• Comparing one variable at a time

Not Allowed

• Large datasets
• Long overnight training runs
• Huge vocab experiments without a reason
• Multi-hour runs “just to see what happens”
• Anything that overheats the laptop or makes the rest of the machine unusable

Recommended Boundaries

• Context window: keep it small
• Batch size: tiny
• Number of steps: short enough to finish in minutes, not hours
• Model size: as small as possible while still teaching the concept

Decision Rule

If an experiment will not help you learn a specific idea, do not run it.

Co-Author Prompt

If a rule here feels too strict or not strict enough, propose a better rule and explain why.