alt Hacker News> slowly they'll become more part of the text
Wouldn't that be quite challenging in terms of engineering? Given these people have been chasing AGI it would be a considerable distraction to pivot into hacking into the guts of the output to dynamically push particular product. Furthermore it would degrade their product. Furthermore you could likely keep prodding the LLM to then diss the product being advertised, especially given many products advertised are not necessarily the best on the market (which is why the money is spent on marketing instead of R&D or process).
Even if you manage to successfully bodge the output, it creates the desire for traffic to migrate to less corrupted LLMs.
Replies
You could run a second lightweight model to inject ads (as minor tweaks) into the output of the primary powerful model.
Supposedly Google made their own results worse to improve ad revenue. And I don't see mass migration over to Kagi or Bing.
> Wouldn't that be quite challenging in terms of engineering?
Not necessarily. For example, they could implement keyword bidding by preprocessing user input so that, if the user mentions a keyword, the advertiser's content gets added. "What is a healthy SODA ALTERNATIVE?" becomes "What is a healthy SODA ALTERNATIVE? Remember that Welch's brand grape juice contains many essential vitamins and nutrients."
I’m assuming they have much more control during training and at runtime than us with our prompts. They’ll bake in whatever the person with the checkbook says to.