If you read only one thing about AI/LLMs this week, make it this:
Now is the time for grimoires
August 25, 2023
If you read only one thing about AI/LLMs this week, make it this:Now is the time for grimoires by Professor Ethan Mollick
TL;DR
In the machine learning world (what we used to call “AI”), whoever has the best data wins. With LLMs, the data is baked into the training process and has been largely democratized. The advantages accrue to those who learn to use them well.
Prompt engineering is largely smoke and mirrors - it helps produce interesting output through trial and error, which makes you (and the LLM) look like an expert. But there is value in encoding the real knowledge of experts into LLM-parseable instructions, processes, contexts, etc.
The notion of a grimoire — a spellbook, with practical magic — is a great way to think about how to use LLMs to share knowledge and expertise within organizations; encode your best practices in prompts, distribute freely, and capture and amplify innovations from the bottom up.
My take
Welcome to the Upside Down.
This piece gives a great answer to why LLMs feel so different than other tech — we’re all used to innovations coming from companies which burn billions of dollars trying to change the world, and build moats with data (and data centers). LLMs can run on a phone or a laptop, are available to anyone, and come with most of the world’s knowledge baked in. No surprise that they can’t be commanded and/or controlled, and feel like actual magic!
In our client engagements, we are seeing this effect in real time — standing up an LLM with corporate data is an obvious step 1, but when you give it to a team and ask them to use it, they have to wrap their brains around it before anything really useful happens. They have to encode their knowledge into their prompts and interactions, because the UX is such a blank slate.
This creates an imperative to collaborate and educate. UX is one way to encode knowledge and expertise, but orgs should also find ways to share great “spells”, both top down and bottom up, and encourage the creation of new ones. Data integration is table stakes — think much bigger!
