Sang Ha Park

Civilization advances by extending the number of important operations which we can perform without thinking of them.

Alfred North Whitehead

Several people including Gwern (1), Tyler Cowen (2, 3), and Astral Codex Ten (4) have recently talked about what it means to write in the age of AI and specifically for them¹. They each take slightly different perspectives: Gwern on how to optimize content for AI, Cowen on why you need to consider writing for AI in the first place, and Scott Alexander on implications for writing for AI. Yet perhaps because it’s too well understood among their readers, like discerning that you are swimming inside the water, they don’t really state the obvious: AI becomes the mediator, synthesizer, mouthpiece and everything in between a reader and content.

Today I want to talk a little about that water we are swimming towards.

There are two main modes of content consumption for AI. One is at pre-training as source of training data, and the other is at inference in production where the AI uses search result to generate a response to a given user query. To me, the latter is getting more interesting because it means that we are entering a new era of consuming information. As we expanded our circle of knowledge from oral knowledge from ancestors to books & libraries to indexed web content at our finger tips, we’ve always maintained close proximity to source of truth and agency to interpret them independently, albeit sometimes wrongly. After all, Google only made finding of relevant content possible (and judging the weight of relevance, which some may argue is an important skill and agency that we have).

However now with AI, for the first time, we are giving priority to AI’s interpretation of said search results. This means that by default we would be seeing not just AI’s curated content, but also its synthesis of the content as well. Immediate consequence is that we would be now a few clicks more away from the original content. In reality, these several clicks will change the default mode of content consumption for billions of people to not consume the original and rely mostly on the AI-mediated output.

And yes perhaps that may be ultimately overall better because AI is actually smarter, fairer, and has less bias than most of us. On a similar note, Gwern interestingly mentions that quality of web scraped data has actually gone up post-ChatGPT. We can also imagine AI exhaustively going through vast amount of search results including obscure and long-tail ones and finding meaningful results from previously ignored data OR making previously unfounded interdisciplinary connections (creativity?).

Yet it does leave a bit of bad aftertaste on how this would influence general human cognitive capabilities & generations of people to come post-LLM. Whitehead wouldn’t have thought that we may even outsource our cognitive capabilities to interpret information.

Furthermore, it’s not just that we’re seeing this inversion of audience only within the consumer space. We are seeing it inside companies too. Recently our marketing team nonchalantly mentioned that they are now creating sales content not for sales teams, though they are always welcome to peruse the original content, but for custom GPTs. These GPTs will in turn interface with the sales team. For now, it seems to be enormously valuable at contextualizing the static marketing decks to relevant scenarios that befit prospects and providing sales with right talk tracks (though I worry about model collapsing here). I don’t have numbers to prove that this new process is working, but the team is at least lot more confident.

However I feel, I’m confident that this shift to LLM-as-an-audience is happening inside every company right now and ensuing transformation of existing trove of data will be one of the biggest opportunities in B2B AI. It’s so deceptively simple yet unimaginably powerful. Of course, there are many many challenges here to be solved such as turning company’s internal multimodal data like PowerPoints and PDFs into LLM-aware structure.

If we follow this stream of thoughts, one of the defining pillars of AI-native companies will be whether you have well-oiled information engine that can seamlessly churn out content for AI, serve them in a readily manner, and repurpose conversations back to meaningful content. Future of work may be totally different to what we have imagined before, but totally restructured around working with (more like for?) AIs.

1. Unironically, I’ve re-discovered them recently through ChatGPT. And all the outgoing links have “chatgpt” attached in the utm_source (ex: https://gwern.net/llm-writing?utm_source=chatgpt.com), so every blog owner / platforms will be seeing the impact of AI search engines. ↩︎