Looking for complex patterns is hard. It usually makes more sense to imagine where a family of complex patterns might be, then look there. Can we do that for LLMs by factoring out (1) things very similar to an individual training datum and (2) fuzzying from context, then see what's left that LLMs are still good at?
If you are inspired by this idea, you can reach out to the authors for collaboration or cite it:
@misc{holtzman-looking-for-complex-2026,
author = {Holtzman, Ari},
title = {Looking for complex patterns in LLMs},
year = {2026},
url = {https://hypogenic.ai/ideahub/idea/oHdUjZOCO70Vqk72eGP4}
}Please sign in to comment on this idea.
No comments yet. Be the first to share your thoughts!