Source record

Published by Hacker News · Show HN (50+ points) (T2 source). The original is at https://github.com/trycua/cua.

Pipeline notes

The summary and note below are generated by the signal pipeline — they are Beyond Desk’s reading, not quotations from the source.

The Cua team released CUA-S1, a 706k-parameter 'System 1' model for computer-use form tasks. Inspired by Kahneman's dual-process framing, it scores given choices (CHECK, CLICK, SKIP, USE) instead of generating tokens. The argument: many agent decisions are local and don't require large reasoning models.

Concrete example of splitting agent work into a planning layer and a tiny specialist decision layer, with hard numbers on size and training time. Useful architectural reference for anyone building agent loops.

Cited by

No citations on record.