Skip to content
MindPattern
Wire
Briefings
Search
Subscribe
← The Wire
Source trail
Hacker News
Public MindPattern findings, entities, and graph evidence that cite this source.
Findings
1
All-time hits
1
High value
0
Last seen
2026-09-01
Related findings
2026-09-01 / PROJECTS
A from-scratch transformer scored 44% on ARC-AGI-1 for 67 cents of total lifetime compute
mvakde trained a small transformer from scratch for 1.5 hours on a single 5090, reaching 44% on ARC-AGI-1 and 7% on ARC-2 at 67 cents covering both training and inference. The method converts puzzle pairs to token sequences, trains autoregressively on evaluation inputs with labels hidden, uses per-task additive embeddings with 3D RoPE and color/dihedral augmentation, then takes the two most common outputs at inference. The upgrades over the previous post are all commodity parts, SwiGLU, RMSNorm, eight layers instead of four and NorMuon instead of AdamW, and the author reports matching TRM and HRM while beating many LLMs.
Open latest cited source
Wire
Briefings
Search
Subscribe