A tiny from-scratch transformer hit 44% on ARC-AGI-1 for 67 cents

“Someone just scored 44% on the famous AGI puzzle benchmark with a tiny model that cost 67 cents to train.”

The Story

An independent researcher trained a small autoregressive transformer from scratch (not an LLM) that scores 44% on ARC-AGI-1 at a total cost of 67 cents. Previous strong scores on the benchmark came from giant LLMs or expensive finetunes. The point is extreme sample efficiency under a hard compute cap.

Why It Matters

Builder utility + quiet future-shock. The “AGI test” that was supposed to need frontier models is falling to a cheap, tiny, non-LLM architecture. Changes how people think about what is actually hard.

Evidence

Author blog + HN (110+ pts, fresh). Primary URL: https://mvakde.github.io/blog/44-on-arc-1/

Sources

Daily scan: 2026-09-01