44% on ARC-AGI-1 in 67 cents
Article URL: https://mvakde.github.io/blog/44-on-arc-1/ Comments URL: https://news.ycombinator.com/item?id=49519939 Points: 228 # Comments: 63
The author has trained a small transformer from scratch in just 1.5 hours on a dataset containing 5090 Beats, achieving performance comparable to top LLMs like TRM and HRM. The model scores 44% on ARC-AGI-1, which is an upgrade to the author's previous model and remains open source. The author trained the model on the non-overlapping tasks from ARC-2, carefully avoiding any data leakage issues.
The model's performance is attributed to good representations such as 3D RoPE and per-task embedding, and the code is now publicly available. The author believes achieving 65% on ARC-AGI-1 should be possible without significant modifications. The author is also exploring ways to reduce costs and remove data augmentation techniques, with the goal of making the model more accessible to the wider research community.
Written by urgent.news from Hacker News Best's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
This story
This is one outlet's version. Read the fullest account.
- 44% on ARC-AGI-1 in 67 cents mvakde.github.io
