LLMs will soon disrupt algorithmic media feeds

A viral post on LessWrong by user lsusr predicts that LLMs (Large Language Models) are poised to disrupt the algorithmic media feeds of giants like YouTube, Facebook, Instagram, and X/Twitter. The central argument is that these platforms are "misaligned"—their recommendation algorithms are finely tuned to maximize corporate metrics like clickthrough rate and watch time, often at the expense of user well-being by promoting 'ragebait' and 'clickbait.' This creates a media environment the author compares to 'crack cocaine,' where user values are secondary to engagement.

The author contends that LLMs now enable a new paradigm: media curation based on a user's explicitly declared preferences. For example, a user could instruct a feed, "I don't want to see any more videos about Donald Trump news," and have that obeyed. They predict this innovation will be pioneered by startups, beginning with long-form written content like blogs, where personalized feeds are currently scarce. Established platforms have a conflict of interest, as their current 'cocaine' model is lucrative, but they will be forced to adopt similar LLM-driven features once a startup proves the market. The long-term vision is a world where LLM-curated media is the 'good for you' option, relegating today's attention-optimized feeds to the status of a vice, much like candy or soda.