“emergence” when it’s good, “hallucination” when it’s bad 🙄

I & colleagues are thinking a lot these days about how we can design interfaces and frameworks that focus on eliciting clearer instructions from humans & offloading the selection of the most accurate plan or pipeline to a system

we’ve mostly been focused on thinking about the systems part & optimizing for accuracy. given a high-level MapReduce-like task, what’s the most accurate execution plan? see my website to learn more about our DocETL project, still working on docs, cleaning up the paper draft, etc

something that has emerged is: the interface we design has to be iterative (people don’t know what they want a good result to look like until they see it). eg, to find common themes and supporting quotes across product reviews, one has to see how many themes arise in the first

this is not about RAG vs long context...you can use short context models to have LLMs directly query a "large" index. split the index up into pieces, have LLMs concurrently search over pieces of the index, and then have an LLM unify the results

one of my favorite hacks to improve the quality of LLM outputs is to ask, in a loop “do you want to refine your answer?" then see all the responses, and then pick the best one. not a single LLM ever says “no” 🙃

This. This is the hardest problem in using LLMs beyond any benchmarks

Defining your grading criteria for LLM outputs is an organic process that evolves the more time you spend on it In “Who Validates the Validators” Shreya Shankar et al highlight this This weekend in our SF office a good chunk % of hackers will learn this 😃 lu.ma/judge