Mithil Vakde's 1.5-hour transformer scores 44% on ARC-AGI-1 for 67 cents
Independent researcher Mithil Vakde trained a small transformer from scratch in about 1.5 hours on a single RTX 5090, hitting 44% on the ARC-AGI-1 public evaluation set and 7% on ARC-AGI-2 for 67 cents of total compute covering training and inference across every task. The result is real signal on sample efficiency, not scale: it matches TRM/HRM-class scores using only 3D RoPE positional embeddings, a learned per-puzzle embedding, and color and dihedral augmentations, with the code open at github.com/mvakde/mdlARC. Frontier-benchmark reasoning research is now reachable on one consumer GPU, without synthetic data or a lab budget.
Source: mvakde.github.io ↗
I trained a small transformer from scratch in 1.5hrs on a 5090 Beats many LLMs, and scores the same as TRM/HRM
Why this matters
- → Frontier AI reasoning (44% ARC-AGI-1) now runs on consumer hardware for 67 cents.
- → Proves sample efficiency, not scale, drives AGI-adjacent benchmarks.
- → Opens frontier research to researchers without lab budgets or specialized infrastructure.