Fred Sala

They also regularly go down right before paper deadlines.

Today at @iclr-conf.bsky.social, come chat with @changho.bsky.social about what types of data drive weak-to-strong generalization!

Some of our work that uses these ideas (and applies to code): harit7.github.io/posts/2023/0...

The generalization of majority vote is a en.wikipedia.org/wiki/Fr%C3%A.... For code, this is the snippet whose total (square) distance to all the generated snippets is smallest. It gets more fun when you try to use a weighted mean, with weights corresponding to how “accurate” each snippet is.

Probably not how the plot is made, but this type of problem is studied frequently in statistical phylogenetics. People often define spaces of phylogenetic trees, often continuous versions, equipped with a metric. For code, these trees are e.g. ASTs.

First up at #NeurIPS2024 from our group, our work on labeling via programmatic distillation (a spotlight!). Label your data orders of magnitude faster and cheaper — come join us today at Poster Session 2 East for a demo!

Landed in Vancouver for #NeurIPS! Looking forward to seeing everyone.

If you would like to chat about data-centric AI and foundation models, reach out!