Hacker News Reader

1. In-context learning - you can teach the model things without retraining the underlying model parameters.

2. Actual model updates (it's not like we'll be stuck with the 2023 version forever).

I don't think you can do 1, can you demonstrate that is possible?

I don't argue you can't add to its trainingset, of course you can.

Ah I wouldn't call that teaching it something new (like a concept), but giving it prompts and responses. It is really cool, though!