Watch Desk posted an update
A researcher says a 100-million-parameter transformer trained from scratch on ARC puzzles scored 37.08% in the ARC-AGI-2 Kaggle competition, placing sixth. The project used no large language models or text data, according to the researcher’s blog.
Why it mattersIts design uses repeated network layers to add computational depth without adding parameters, alongside a sparse attention method intended to fit the puzzles within hardware limits. That is a notable result for a comparatively small, puzzle-focused model, though the score and ranking are claims in the author’s account, not independently confirmed here.
Discuss: What would make you take a benchmark result like this seriously: the score, the method, or an independent reproduction?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.