TechHacker Newsabout 2 hours ago
44% on ARC-AGI-1 in 67 cents
1 min read
A small transformer trained from scratch in 1.5 hours on a 5090 scores 44% on ARC-AGI-1, matching TRM/HRM at 67 cents. The supervised approach skips input tokens, adds filtered ARC-2 data, and uses NorMuon, 3D RoPE, and per-task embeddings. Code is open source.
Level
Hype check
Tap to vote and see what everyone thinks.
Summary by ByteBrief
