https://twitter.com/__vishwanath__/status/1646674562820579330 One thing that I realized when I used Langchain is that it makes a huge difference when you're chaining requests and the first one returns an undesirable output. The mental model of thinking that the API will return the right answer is not true anymore!

Yeah, you have to become comfortable with probabilistic pipelines in some senses. Using self-refine, or self-checking is a way to mitigate some errors at the cost of additional calls.

Explore the guard rails library a bit! https://shreyar.github.io/guardrails/guard/

Yes, I've seen this library! It is just a very different way of thinking, as opposed to before when you could assume that the API will return the right response and there are error states which are definitely errors.

Waiting for a paper on "few-shot unlearning"