B01 — Godot-RL neural-foundations bridge
← Diagnostic & learning path · F01 — Python & NumPy → · Glossary
Only if you completed the Godot RL course; otherwise go to F01 — Python & NumPy.
Purpose
This compact page recognizes demonstrated Godot-RL skills without treating a course certificate as proof. Complete the challenges below; revisit only the foundation blocks where evidence is missing.
Transfer map
| Godot-RL foundation | LLM foundation | Verify or learn now |
|---|---|---|
| weighted sum, bias, activation | neuron and MLP | verify F04 shapes and one gradient |
| layers and forward pass | token representation pipeline | add the [B,T,C] time axis in F02 |
| loss and backpropagation | next-token cross-entropy | learn vocabulary softmax and F03 loss |
| policy logits | next-token logits | both normalize scores, but pretraining has labelled next tokens |
| PyTorch training | PyTorch LM training | verify gradient clearing and causal target shapes |
| failure diagnosis | experiment diagnosis | retain trace, baseline, seed, and repair evidence |
| ONNX/native inference | optimized LM inference | later add autoregressive decoding and KV cache |
Twenty-minute evidence check
- Infer
[4,8,64]from activations[4,8,32]and weight[32,64]. - Calculate softmax and both target losses for logits
[2,0]. - Explain why reward is not used in ordinary next-token pretraining.
- Obtain gradient 28 for
(2x+1)²atx=3by chain rule and central difference. - Run one PyTorch
[4,3] → [4,2]update with seed 7 and finite gradients.
| Result | Route |
|---|---|
| all five independent | go directly to the F06 common checkpoint |
| one supported answer | review the corresponding F02–F05 section, then F06 |
| any critical shape, loss, or gradient error | complete that full block, then F06 |
Do not skip F03 merely because policy networks also emit logits: vocabulary targets, sequence axes, causal masking, and likelihood training are new.
What’s next
Take the shared F06 checkpoint. It is the same gate for both entry paths.
← Diagnostic & learning path · F01 — Python & NumPy → · Glossary