Pareto frontier literally off the charts

Update:
Wow this blew up. Pressure is on.
Please bear with me as I want to do careful ablations.

• upload the run-script.ipynb file to google colab or modal
  • (optional) if you want to save results, mount your drive/volume. If you don't, then comment out the 2 cells that save save a checkpoint to drive.
• choose A100
• Hit run all

Every DL approach on ARC today trains a supervised algorithm (other than compressARC)

I think this is suboptimal.
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

Implementation details: New pareto frontier on ARC-AGI For why I chose these specific implementations, read my blog on Why all ARC solvers fail today

Performance - 27.5% on ARC-1 public eval Total Compute cost - $1.8

• ~127min on 40GB A100 for training (1.2$)
• ~49min on 80GB A100 for inference (0.6$)

This is early performance. I was too GPU poor to do hyperparameter sweeps.

I should be able to push to 35% with just basic sweeps

I expect to hit 50% with a few obvious research ideas