Tyler Cowen suggested an interesting perspective regarding the future of AI mediated knowledge. If the primary audience for ones writing is an LLM instead of a human googling, how should we think about the nature and style of our writing?
Just like SEO and algorithmic feeds warped the nature of internet writing, i'd expect some major evolution. How does the experience of asking ChatGPT differ from a Google Search? The intent is likely to be very different. Asking an LLM "Tell me about Tyler Cowen" has less specificity and focus than "Tyler Cowen writing for ai".
There is a difference between making it onto the first page of Google vs making it into the LLM's training dataset. Superficially that might seem the same, they're both crawling and indexing the web for knowledge and documents but when you consider the optimisation functions, I'd argue theres a subtle but significant change.
The quality of a Google search hinges on its ability to find or guess the exact piece of content being sought. The quality and intelligence of an LLM query hinges more on its breadth and depth on knowledge, combined with the models curation and summarisation abilities. Furthermore, the LLM inference might be part of a greater agentic workflow.
The Bad Place
The dystopian reading of the shift from Search Engine to LLM as the prime mediator of knowledge work is that the entrenched companies get even more entrenched as they scale and build insurmountable data monopolies. Summarization cannabalises the ad revenue that has kept independent web sites afloat. The free and open web of the 90s and 2000s retreats ever more behind the paywalled gardens and algorithmic feeds.
Free from our SEO Listicle prison
My optimistic take is that if breadth and depth become more important than SEO minmaxing on long tail ahrefs keywords that means that we can all worry less about the specific formatting of content. LLM knowledge training data will have a nearly infinite learning capacity to the point that running out of training data is real concern.
Maybe the best way to "write for the ai" is just to write. About anything and everything, to give the LLMs the largest possible dataset to understand us. Long form, short form, maybe even audio or video to capture the ineffable vibe of our personalities to the machine god. Utopia would be everyone free to write or vlog on open personal websites with infinite flexibility of format and taste. Geocities but with a billion users. And the LLM seeing eveything that we make and it is good.
Writing is enjoyable, writing for distribution is a chore. Maybe the death of the search engine will free us from that box. Interestingly, after writing this I was tempted to proof read it with an LLM. But I wonder if in the future the missed capitialisation, missing word or inconsistent British/American spelling becomes a high signal of organic human thought.