8-layer transformer hits 44% ARC-AGI-1 at $0.67 on single 5090

The model converts each puzzle into token sequences and performs autoregressive supervised training only on output tokens using SwiGLU, RMSNorm, 3D RoPE, and variable-length Flash Attention. Training incorporates filtered ARC-2 tasks without leakage, color and dihedral augmentations, and per-puzzle additive embeddings. This yields 44% accuracy versus the prior 40% while cutting augmentations and switching from AdamW to Normuon, directly lowering cost and wall time. Data shows the change to output-only loss raises test loss yet improves accuracy and training stability, indicating finite capacity benefits from reduced input token prediction. The result matches TRM/HRM scores but uses only public ARC-1/2 data and no external synthetic generation or human priors. Prior work required four layers and full input-output training; the upgrade demonstrates measurable sample-efficiency gains from architecture and data handling alone. ARC remains a metalearning benchmark with 1000 tasks where each rule is unique yet concepts overlap between train and eval. The approach isolates transformer limits on sample efficiency without LLM-scale pretraining. Regulatory discussions around AGI overlook these low-cost, open-source iterations that allow rapid public iteration. Next steps include further architecture variants to push past 50% while keeping per-run cost under one dollar. Future runs will test whether Normuon and flex-attention kernels scale to 16 layers without proportional cost increase on consumer hardware.