FOMO, Search, and Large Language Models

Favorite Tweet of the Week

Morgan Housel @morganhousel

“Lack of FOMO” has to be one of the most important investing skills.

The benefits of compounding in any discipline come with time and FOMO causes you to interrupt that compounding.

What I Learned This Week

I’ve been experimenting with and deploying semantic search models for a product I’m building. I came across this excellent tutorial called Retrieve & Re-Rank from the sentence transformers library. Before I was only performing the Retrieving step where you use vector search to rapidly go from a million possible search results to something smaller such as a 100. The tutorial describes how to use cross-encoders to rerank those 100 promising results so the top 5 are hyper-relevant to your search query.

Boring and Remarkable Applications of Large Language Models

There’s no shortage of demos from LLMs like GPT-3 that generate incredible and exciting artifacts such as stories, thought experiments, and jokes. These are types of use cases that captured the public’s imagination and are also a big source of debate about replacing jobs and having baked in biases.

However, what excites me the most is applying LLMs to boring, well-defined, and narrow use cases where they perform remarkably well such as data extraction and back-office type tasks.

Amjad Masad ⠕ @amasad

Very soon Large Language Models will be a basic cloud primitive, like say compute or storage. I honestly can’t think of a feature or a domain where applying this tech would *not* be useful.

I’m convinced the act of teaching the LLM to do these tasks or “prompt engineering” is going to breed new types of jobs and engineers. The art of asking the LLM what you exactly want it to do is going to be a superpower.

Here’s an example from OpenAI’s playground where you convert a paragraph into a structured table that can be imported into software like Excel.

A few years ago this would’ve taken an army of NLP researchers, engineers, hundreds of thousand dollars, and many hours of data labeling.

GPT-3 accomplishes this when you merely ask it to “summarize the fruits from Goocrux”

Still not convinced?

Here’s an example I drummed up in 5 minutes. Imagine you’re a customer success manager and you want to score customer reviews on a spectrum. I described my review spectrum, gave the model some specific examples to learn from, and it outputted scores highlighted in green for a couple of reviews.

Not bad for 5 minutes of work.

It’s too hard to exactly predict how much impact any technology will have on society. But I’m sure we’re going to see amazing companies, products, and jobs built on the capabilities of LLMs.