Challenge #3: Sparse parity on the grid
Solve sparse parity using only about nine instructions on the Dally 2D grid.
The grid-model version of Challenge #1: solve sparse parity using only about nine instructions on the Dally 2D grid, with a restricted op set and no abstraction.
Where it lives
sutro-problems, the sparse-parity/ directory. The scorer (sparse_parity.py) is locked.
Cost metric
The Dally 2D-grid cost model.
State
Launched May 8. Precomputing intermediate XORs plus bit-packing works well; tiling does not help at this problem size. A 50%-accuracy target variant has been added. The tiny op set makes it hard to cheat, and no agent-cheating results have shown up with the IR approach.
Open questions
Whether continuous floating-point ops are needed at all, or whether integer and 8-bit instruction sets are better, since fewer ops means fewer hidden free optimizations for agents to exploit. See the May recap.