Few-Shot Learning in Practice: GPT-Neo & 'HuggingFace' Accelerated Inference API

Good to see that few shot learning is now even easier using the widget. We have had good results with writing emails and reported about it in our blog: https://www.helm-nagel.com/open-sourced-gpt-Neo-writes-indiv...

Hmm I don't work in this space too often, but wouldn't the maximum likelihood estimator of a oneshot example, saying telling it king => queen, hoping to learn male to female transitions... wouldn't that just be the simple linear transformation between the two embeddings?

I'm sure a few shot model can be a little more nuanced but I'm not sure I have any intuition about what that transformation looks like.

HuggingFace is going to get acquired in the near future, the talent pool they’ve accumulated is extremely impressive.