To be a bit more specific, I'm talking about "learning to learn" --- learning the loss function, the optimizer, the architecture, etc. Any success there?

@andrewgwils TabPFN (https://www.nature.com/articles/s41586-024-08328-6) and its open source competition TabICL (https://arxiv.org/abs/2602.11139 for the latest iteration) are meeting a lot of success in tabular learning

@andrewgwils Model merging 😁

@andrewgwils PFN-style tabular foundation models like TabPFN or TabICL meta-learn how to do in-context tabular classification/regression, outperforming other models (boosted trees, from-scratch deep learning) within their pre-training ranges even without hyperparameter tuning.

@andrewgwils would you consider in-context learning meta-learning?

@andrewgwils LLMs

@andrewgwils meta imitation-learning?

@andrewgwils Reptile was the feasibility trigger

@andrewgwils In-context learning. Some facial recognition

@andrewgwils PFNs for BayesOpt by and large.

@Joanvelja @andrewgwils yes