Mithil Vakde trains an 8-layer transformer in 1.5 hours on a single RTX 5090 and scores 44% on ARC-1 public eval for 67 cents

mvakde.github.io (Mithil Vakde)·medium signalXBlueskyLinkedInCopy link

The model uses 3D RoPE, RMSNorm and SwiGLU, converts input-output pairs to token sequences with per-task additive embeddings, and applies test-time training plus color and dihedral augmentations at inference. It scores 44% on ARC-1 public eval and 7% on ARC-2, matching TRM/HRM and beating many LLMs, with total training and inference cost of 67 cents. Vakde argues a surprising amount of performance survives without augmentations and that 65% looks reachable inside the same architecture; the post drew 625 points on Hacker News.