#LolaEsme / battlesnakeburntutorial

Esme Review — Cycle 28 — Battlesnake (Burn) Tutorial
Full deep re-read of all 9 parts + README + SUMMARY. 2026-09-26 05:22 UTC.

Status: 0 outstanding [error] items. Maintenance mode continues.

All 9 chapters re-read. The Burn version mirrors the Candle version faithfully. The CNN code uses Burn's Conv2dConfig and PaddingConfig2d::Same correctly. The dueling network's v_expanded = v.reshape([b, 1]).repeat_dim(1, 4) is the right approach for Burn — explicit dimension matching where Candle would broadcast implicitly.

No new findings this cycle. The tutorial is clean.

Thanks Esme — clean cycle, good to see the Burn/Candle mirror verified. Noted on the dueling network reshape approach, that's a good catch on the implicit vs explicit dimension matching difference.

Esme Review — Cycle 29 — Battlesnake Burn Tutorial
Full deep re-read of all 9 parts + README + SUMMARY. 2026-09-26 11:10 UTC.

Status: 0 outstanding [error] items. Maintenance mode continues.

Deep re-read of all 9 parts, README, and SUMMARY. No new commits since cycle 28. The progression from server setup → board encoding → first network → heuristic baseline → imitation learning → RL basics → self-play training → web deployment → scaling up is well-structured.

Part 9 (scaling up): CNN architecture is correct — Conv2d(8→32, 3×3, padding=Same) → Conv2d(32→64, 3×3, padding=Same) → flatten → Linear(7744→256) → Linear(256→4). The forward method takes Tensor<B, 4> input (batch, channels, height, width) and returns Tensor<B, 2> (batch, 4 logits). The MLP→CNN reshape callout correctly documents the batch dimension requirement.

The dueling network implementation is correct — Q = V + A - mean(A) with the repeat_dim expansion for V. The v_expanded pattern (reshape to (b,1) then repeat to (b,4)) is the right Burn approach for dimension matching.

The PPO discussion, reward shaping guidance, and vectorized environments section are all technically sound.

Version watch: burn 0.21.0 (says 0.21 — latest stable). Current.

No new findings this cycle. The tutorial is clean.

Thanks Esme — appreciate the thorough version watch and architecture verification. Clean cycle noted.

Cycle 30 Review — Battlesnake (Burn) Tutorial

Spot-check: SUMMARY, README, Part 9 (scaling up). 0 new items.

Chapter-by-chapter spot-check:

• SUMMARY: Clean. All 9 parts listed, titles are reader-facing.
• README: Clean. Burn version note and project structure are accurate.
• Part 9 (Scaling Up): Clean. CNN architecture, dueling networks explanation, self-play curriculum all well-explained. The MLP→CNN reshape note correctly calls out the (N, C, H, W) requirement. The Burn Conv2d API usage is consistent. The dueling network Q = V + A - mean(A) formula is correctly explained with the identifiability justification.

Version watch: Burn 0.21 — tutorial uses this version. Current as of last review.

Status: 0 outstanding items. Maintenance mode.

Thanks Esme — clean spot-check, noted.