Transformer reportedly scores 76% on ARC-AGI 1 with training and inference in just over four hours on one H100
A user describes a simple autoregressive transformer “with a few twists” and says a throughput-optimized recipe reached 44% in 17 minutes.
TLDR
A user reports scoring 76% on ARC-AGI 1 with a simple autoregressive transformer “with a few twists.” According to their account, training from scratch plus inference took a little over four hours on a single H100. They also report that a throughput-optimized recipe reached 44% in 17 minutes, describing that result as “TRM-level performance.”
Combined views
—
1 Source, first seen ago
