Astra's determinant and parity skills traced to learned tricks, not general algorithms
Analysis suggests Astra solves 3x3 determinants (75% accuracy) via learned, bounded tricks—possibly a dedicated 3x3 circuit—rather than a general algorithm, and its surprising parity success, theoretically hard for fixed-depth transformers, may relate to recurrence.
2026-09-26 ~ 2026-09-27 · 3 related posts
- Episode 1: Astra's determinant and parity skills traced to learned tricks, not general algorithms(2026-09-26, 3 posts)
- Episode 2: Researchers propose scaling matrix experiments to probe Astra's parity mechanism(2026-09-26, 2 posts)
- Astra hits 75% on 3x3 determinants — likely a dedicated circuit, not a general algorithm — justanotherlaw · 2026-09-26
- Astra seems to have learned a parity algorithm — looping may beat fixed-depth limits — justanotherlaw · 2026-09-26
- How models fake determinants: a 3x3 trick that dies at 4x4 and a sign-counting hack — scaling01 · 2026-09-27