Update 13 dec: 27.5% now. Looking at the output grids, I expect 50% with scaling. Also gonna run this on ARC-2 today Update: 20% right now (In progress, I expect improvements till 30%)

10% on ARC-1 for less than a dollar using a 1M transformer

This already beats the pareto frontier btw

Self supervised compression on ARC

Every DL approach on ARC today trains a supervised algorithm[1]

This is dumb.
A self-supervised compression step will obviously perform better:

• There is new information in the input grids and private puzzles that is currently uncompressed
• Test grids have distribution shifts. Compression will push these grids into distribution

For more reasoning behind the approach, read My Blog

Details

Performance - 10% on ARC-1 public eval
Total compute cost - $0.709

• 52m on A100 for training (0.7$)
• 40s on A100 for inference (0.009$)

This is early performance. Haven't run all ablations yet

I should be able to push to 30% on ARC-1 and 8% on ARC-2

[1]: CompressARC is an exception, but that compresses each task individually. Mine jointly compresses all tasks together