Lately, it seems everyone uses LLMs to write. People are justifiably annoyed at the 'default' ChatGPT style of writing. Having ChatGPT write an email for you, or fix your grammar is a low value use case though.
I think there is a lot of value in doing the opposite: doing more writing yourself, so you can then feed it to an LLM. Writing not as an end, but to generate (quality) content for AI. Writing down things so that AI can help you (or others) down the line.
Interestingly: people like Tyler Cowen and gwern go as far as to say that writing publicly for LLMs is functionally a way to gain immortality.
In a world where written knowledge is easily accessible by LLMs: there is a lot to gain when you pivot to a writing culture. Instead of making decisions face-to-face, switch to writing memos and jotting down your decisions explicitly. You can then use AI to reflect on your writing, give you deep feedback, get it to spot patterns and trends, and dig up the nuggets of insight that are often hard to find.
Here are some trends I've noticed recently:
-
I have noticed people who are clued-in have started to take more notes than before.
-
For example: apps like Granola let you record your meetings, take notes yourself, and then ask questions about your notes.
-
Even with Granola, the high value use cases come about when you jot down your own thoughts as well, not just depend on it to transcribe the meeting.
-
-
There’s a trend to make API documentation, user guides etc all available as plain text in an llms.txt file. This documentation can then be easily fed to LLMs so that the LLM can help users write code.
-
Writing evals is one of the highest value tasks you can do when building AI applications. People have been paid by AI labs to create high quality content and high quality evals for AI.
- Epoch AI for example hired 70 mathematicians to generate problems for their frontier math eval. Imagine being one of the top experts in your field, and the highest value use of your time isn’t doing research, but just writing tests for frontier AI models!
Now, fit this pattern into your organisation. Writing tests, background material, context for AI shouldn't be a low value task; and you should put your best people on it.
Every organisation has a lot of tacit knowledge, which is completely invisible to AI. It's not going to be possible to codify every little part of your business — but being deliberate about writing will help!
Building tools to reflect on your writing
For personal use, you don't need to be too fancy here. With LLMs having huge context lengths like Gemini, I can put every personal note I’ve ever written over the last decade into one LLM prompt.
As an organisation: you could start collating information across multiple internal systems (eg: PRDs, design documents, internal tickets, etc) and consolidating it in a central space, and adding AI tools on top. But before you do this, you need to pivot your org to a writing culture.
Start writing things down — not just for yourself or your co-workers right now, but the AI agents you'll soon be working with. We have started doing this at Clear.