Using LLMs for Search

Before getting to the content of this week’s newsletter, I’m offering a quick reminder that I’m teaching an “ETA 101” course on Maven starting Wed Nov 1st. There’s an info session on Thursday Oct 25th at 5:30pm PT on LinkedIn - reply to this email if you’d like to join.

As a person doing a search fund, I have a regular need to get smart on an industry quickly. Sometimes this is because I’m proactively investigating an industry, and sometimes it’s because I’m looking at a company for sale by an intermediary and am trying to get up to speed on industry dynamics. The end goal in either case is a modicum of credibility when talking to a business owner in the space.

Sometimes this is easy: a lot of SaaS products are pretty easy to grasp from poking around the website. And sometimes it’s really hard. The first company I was really interested in was a genome sequencing & interpretation lab and software service. That experience forced me to find quick ways to learn, and I found Large Language Models (LLMs) like Chat GPT and Claude more effective than Google in getting up to speed.

General Advice on Using LLMs for Search

Think of an LLM like an intern - the results you get will be as good as the effort you put into training the intern. I found this article helpful in getting me in the mindset of “managing” the LLM. Not crazy about the results? Ask the LLM to do it over again and provide feedback on why the answer was not helpful. Make it a conversation and don’t be afraid to have the intern re-do work multiple times until you’re happy with it. Also, don’t be afraid to ask for an unreasonable amount of iterations or versions of things. My favorite quote from Ethan’s article above:

I mostly use ChatGPT-4. The gap between GPT-3.5 and GPT-4 is significant and worth the $20 / month. As a recent article from Casey Newton stated, “You can fast forward through time by spending money,”

ChatGPT also offers “Custom Instructions.” In the past, answers were constructed with no knowledge of your query history or information about you - similar to a google search in Incognito mode. With custom instructions, you can provide some guidance that will influence your query results.

I use Claude for analyzing PDFs.

Accelerated Industry Learning with LLMs

Below are a few use cases and associated prompts that I’ve found useful. Prompts are listed up top in italics.

Explain a concept to me

Imagine that you are a who also has a knack for teaching complicated subjects in an easy to digest way. Give me an overview of and describe it to me first like I'm a fifth grader, then like I'm a senior in high school.

This has been helpful when I’m learning about a new concept or industry that’s complicated and not well-understood by me. I get a high level understanding and often get great metaphors that can be used with investors and others who don’t know the concept well.

Teach a concept to me

Imagine that you are a who also has a knack for teaching complicated subjects in an easy to digest way. Teach me how works:

1. Ask questions about my level of understanding of the core concepts.

2. With each response, fill in the gaps using simple words.

3. Recursively ask me more questions to check my understanding.

I use this when I know I need to understand a concept in a deeper way. I’ve found a lot of the magic with LLMs to happen in the sparring and back-and-forth - this prompt helps you do that.

Generate ideas

This one is harder to share as a concise, pre-written prompt. The more thought you put into writing a good prompt, the better the output will be. Some general principles:

• Tell GPT who it should be (usually someone with deep knowledge of a subject)

• Tell GPT who you are and what your general needs are

• Tell GPT the specific needs to solve for in the query

Here’s an example prompt for generating industry ideas in pharma services.

Summarizing PDFs and webinars

Sometimes I’ll find lengthy PDFs containing industry analysis or trends and will ask the LLM to summarize the key insights that are relevant to me. As with the prior example, it’s hard to write a general prompt; be clear about your needs and the role the LLM should play. Note that Chat GPT does not analyze PDFs - you’ll need to use Claude instead.

Same with webinars - convert to text / transcript and run through an LLM. This gives me a sense of whether all or parts of the webinar are worth my time, and I can ask it to look out for trends that are of interest to me.

Future Newsletters

Thanks to those of you who have written with suggestions on future topics. If there’s anything in the ETA world you’d like to see covered in more depth, drop me a line and let me know.