Forecast
State direction, magnitude, assumptions, confidence, and likely failure modes.
Chess GPT learning lab · Field notebook 00
Predict before
you train.
Learn machine learning by forecasting what each lever will do, testing it on a real chess model, and explaining the evidence.
The learning contract
Understanding has observable evidence.
Discriminate
Choose the best implementation and explain why the alternatives fail.
Measure
Run one controlled change against a trustworthy baseline.
Update
Explain surprises from memory and revise the mental model.
The complete path
Ten causal questions.
Ten chapters form the core. A four-part reinforcement-learning extension follows once the supervised model and evaluation protocol are trustworthy.
→
01What does it mean to learn?
Separate reusable patterns from memorized games.
□
02How can a number learn?
Turn prediction errors into useful parameter changes.
□
03How do moves become probabilities?
Generalize beyond histories seen word-for-word.
□
04Did it learn or memorize?
Know whether improvement will survive unseen games.
□
05How fast should it learn?
Use the budget quickly without destabilizing training.
□
06How can any earlier move matter?
Reach relevant context anywhere in the game history.
□
07What makes a Transformer trainable?
Assemble a stable model within the size limit.
□
08How should we spend the budget?
Allocate 100 MB and one exaFLOP deliberately.
□
09Does lower loss mean stronger chess?
Select for wins, not attractive training curves.
□
10How does an idea become evidence?
Make creative experiments credible and useful.
Phase II · optional competitive extension
Reinforcement learning
RL asks a different question: not “what move did a human play?” but “which action improves expected game result?” Self-play only reaches the tournament model after the shared-data rule is clarified.
□
R1When is prediction the wrong objective?
Separate imitating human moves from maximizing game result.
□
R2How does a win teach earlier moves?
Assign credit when the clearest signal arrives at checkmate.
□
R3Should we learn values, actions, or both?
Choose an algorithm family that fits chess and the budget.
□
R4Can self-play improve us honestly?
Control opponent drift, variance, compute, and evaluation bias.
Placement diagnostic · 10–15 minutes
Show me how you reason today.
No preparation and no pass mark. Answer without searching, record your confidence, then submit once. Your result determines the starting path for Chapter 1.
Important: a confident wrong answer is more useful for teaching than a disguised guess.
What is the final value of best?
scores = [2, 5, 1]
best = scores[0]
for score in scores:
if score > best:
best = scoreThe played move was Nc6. Model A assigned it probability 0.8; Model B assigned 0.2. Which model has lower negative log-likelihood?
For L(w) = (w − 3)² at w = 1, the derivative is −4. A small gradient-descent step moves w in which direction?
The learning rate becomes 100× larger and updates repeatedly overshoot the minimum. What is the most likely training behaviour?
A fixed dataset changes from 10% to 40% validation data. What trade-off should you predict?
Model width increases substantially while data and training stay fixed. What is the safest initial prediction?
A batch has 32 games, 16 moves of context, and 64 values per move embedding. What is the resulting tensor shape?
Which workflow preserves an honest final estimate while still allowing model selection?
0 answers · 0 confidence ratings8 decisions remain
The source shelf
Primary sources, assigned with purpose.
Neural Networks: Zero to HeroAndrej Karpathy · build-from-scratch spine↗Understanding Deep LearningSimon J. D. Prince · visual textbook spine↗Speech and Language ProcessingJurafsky & Martin · language-model theory↗Mathematics for Machine LearningDeisenroth, Faisal & Ong · selective mathematics↗Deep LearningGoodfellow, Bengio & Courville · selective theory↗Spinning Up in Deep RLOpenAI · conceptual RL extension↗