Which explains why I've been getting ads for brides' maid dresses.
However, it also seems as if a tuned LLM could look at the corpus of my on-line profile (mail, FB, HN) and produce a model of my attention and commentary (positive/negative). Then that model (or several models) could be tested against various ads to determine which are more likely to produce positive responses. Right now I think that would be too expensive except for very high value ads/targets, but perhaps pricing will come down and non-keyword based advertising might appear.
Absolutely not.
Advertising is applied sociology, not psychology.
Ads don't care about your personal wants or desires, they only care about Bayesian probabilities across broad target audience groups.
(E.g. "dog lovers are X more likely to Y than cat lovers".)
That makes sense if we're looking at broad-scope advertising.
I always figured there are some verticals where the value per acquisition was high enough, and the product narrowly-scoped enough, where it might make sense to microtarget individual customers-- profiling them individually and building custom messages for each of them.
The obvious target would be the narrow-scope prescription drugs-- all those ads that say "If you have XYZ-positive, PDQ-negative Exploding Sphincter Syndrome Type XVI with Sprinkles and Cheese Sauce". Presumably there are only a few hundred people with that condition in the country, but you're spending a lot of money to reach everyone else to try to catch them in your net.
Super-premium luxury goods might also fit there, but I wonder if the brand risk would be too high. Rolls Royce probably knows the 200 customers most likely to buy their new model, but it might not fit their brand to actually put together a campaign that seemed like they were actually TRYING to sell cars.
If the cost is reasonable enough, it doesn't even have to be crafted for the individual, simply target them with one of multiple ads, which are modeled to have the highest (positive) effect... If you're not going to interact with it showing it to you has minimal effect. If one is modeled to have 5x higher interactivity, that's the one you want to show.
I think there are high value customers for whom this probably makes sense, but only if those doing the targeting make money from it. Right now you're the product, but the supplier is the one doing the targeting (or not) and the customer doesn't know enough about their potential customers to target. The suppliers don't really make money by effective targeting.
However, it also seems as if a tuned LLM could look at the corpus of my on-line profile (mail, FB, HN) and produce a model of my attention and commentary (positive/negative). Then that model (or several models) could be tested against various ads to determine which are more likely to produce positive responses. Right now I think that would be too expensive except for very high value ads/targets, but perhaps pricing will come down and non-keyword based advertising might appear.