flâneur

Everything I'll forget about prompting LLMs

olickel.com · 7,816 words · saved by 1 readers

This isn't written by ChatGPT. Claude and friends had no hand making this guide. This is three hours of sit-down writing, so forgive me for changes in tone or descriptors as you read on. In other news, WishfulSearch (natural-language search for complex JSON) is open-source now, and a lot of the techniques in this guide are implemented there. Check it out on github! When you have a problem with an LLM, prompting should ideally be your first approach. Reach for finetuning, or smarter more expensive models once your prompts are as good as they can be - or you're leaving money and performance on the table. Try the techniques in this guide before you consider other options. Finding a good prompt can be the difference between 12 cents per thousand tokens and 0.2 cents per thousand tokens, or even 0.1. It can also be the difference between needing 40 GB of VRAM (which costs $20k), 20GB ($2K) or 8 GB ($300). But why not finetune? Models change all the time, and the dataset and compute requirem

Everything I'll forget about prompting LLMs Hrishi Olickel Everything I'll forget about prompting LLMs A definitive guide to my future self This isn't written by ChatGPT. Claude and friends had no hand making this guide. This is three hours of sit-down writing, so forgive me for changes in tone or descriptors as you read on. In other news, WishfulSearch (natural-language search for complex JSON) is open-source now, and a lot of the techniques in this guide are implemented there. Check it out on github ! When you have a problem with an LLM, prompting should ideally be your first approach. Reach

related reading