[MagnumPIG]

Part of this inaccuracy is on purpose.

People can really freak the fuck out when you can predict what they need accurately. For example the real case of a store advertising specific baby items to someone who didn't know she was pregnant.

This is why "suggested items" often have one wildcard thrown in: They don't want you to realize just how much they really know you.

Of course this doesn't explain failures, but I wouldn't rule out smart ads just yet.

[smnrchrds]

I remember reading that article too (about the pregnancy thing). I was blown away at the time, but seeing how much AI has progressed in the past decade yet how far away we still are from seeing that in practice, I must say I no longer believe it. I believe like all AI hype of the past decade, the achievements were embellished beyond the point of truthfulness. Perhaps they sent the ad to 100 people, 20 of them pregnant and knowing it, 1 pregnant and not knowing it, and 79 not pregnant, and chose to only report the 1. Perhaps they just made up the story. Perhaps they just got lucky and by pure chance they got one prediction right, and have been trying for the past decade to recreate that magical moment. Remember Google Flu Trends? What ever did happen to it? Surely they didn't discontinue it because it was predicting flu too accurately.

[derefr]

I wish this was true. I would love to be able to go to e.g. Amazon and just find something I didn't know I wanted recommended to me on the front page. Instead, the only things on there are advertisements for Amazon Prime affiliate services; the list of things in my wishlist/saved items/recently browsed; and some objectively (i.e. un-customized) "hot" items in categories it knows I browse.

Meanwhile, the thing I might want, if only I knew about it? It's not on the front page; it's not on the hot or new pages; and it's not anywhere near the first page of results for any search I do. These are SEO death-zones, where I just see 1000 optimized contenders for the one boring highest-profit-margin product.

Instead, to find genuinely-interesting new products, I have to go into particular micro-categories, and then browse through ~20 pages of irrelevant same-y things to get past the micro-category's own SEO death-zone. (Even within e.g. the "Scientific Instruments" category, the first ten pages are all either N95 masks or brewing equipment, rather than, y'know, beakers and test strips and CO2 monitoring equipment and such. I know why—they don't re-rank per the browsing habits of the other people who've viewed a given category, but instead reuse the item's global rank in all categories it appears in—but it's still ridiculous.)

I mean, maybe I'm an outlier; I watch YouTube reviews of life-hack tools, kitchen gadgets, etc., so most "novelties" aren't all that novel to me. When I'm looking for "something I don't know about", I more mean "something that would excite me, but which nobody within my filter-bubble is excited about yet."

But surely an AI could deduce a ranking algorithm that would show people like me what they want to see, right? I feed it plenty of training data in terms of what I do and don't bookmark/save on the site. It just needs to think one level up from "tags" / "similar users."

And the weird thing is, I feed plenty of data on exact products I've been interested in in the past to every service I use. Like I said, I watch YouTube videos about e.g. knock-off portable game consoles; I search Google for those products, say things about them in Facebook Messenger, etc. I know I'm getting my privacy invaded by these services—the least they can do is to actually use that information to get me a "recommended" product listing for the thing I'm considering buying!

[tylerhou]

She knew she was pregnant, and had made purchasing decisions based on that. Her father did not, and called Target when the house received ads for baby-related items (cribs/diapers).

[dchichkov]

Hmm. Maybe this story was simply marketing of "perfect" advertisement services? Or an outlier?

[JAlexoid]

I would have thought that HN readers would at least be somewhat tech savvy, but alas.

Advertisement systems are way less advanced and much more stereotype driven, than privacy freaks care to acknowledge. Think of this - advertisers have less than 100ms to decide the best ad to show you... often multiple times per website. Do you really think that any ads are actually personalised?

I remember having to differentiate between Bike Helmet for Barbie(a toy) and Barbie Bike Helmet(safety device for children)... Or the fact that if you let a learning algo run through the categories, they can place dildos(adult toys) right next to water guns(kids toys).

There are services that target you with offers that are highly tailored*, but online ads are not one of those services.

(Amazon's "You May also like", food delivery services suggestions, and similar things that can calculate for a long time)

[andrepd]

?? 100ms? Why, is precomputing outlawed or anything and I didn't hear it? And "privacy freaks", really? I thought after Snowden the notion that only crazies care about privacy no longer flies as it did.

[JAlexoid]

I'm not interested in educating you of all the constraints that each display ad tech company has to operate with. Suffice to say - no one is calculating or storing a list of "perfect ads" for you, nor does anyone particularly care that you like tentacle porn. You're really not that valuable.

Yes. Privacy freaks are overreacting on how deep ad tech cares to know about you. It's way less than what NSA or spying agencies(hello Facebook) are able to collect and use.

[rightbyte]

100ms? I mean Google et al. can precompile a list to serve you. There is no need for a hurry.

[JAlexoid]

They can't. The lists of viable ads are constructed, built and updated real time. There are shortcuts, but then you have to account for things such as quality scores, probability of you clicking on the ad if you recently have seen them, etc.... and do that trillion times per second.

Even light speed is not infinite.