News for hackers. Stuff that matters.
Re:Continuing our work to improve recommendations on YouTube by maxxxxx on Friday January 25, 2019 @04:00PM (#19002444)
All the AI stuff is really still incredibly dumb. Same for Amazon: Buy a mixer and suddenly you will be haunted everywhere by ads offering more mixers.
Re:Continuing our work to improve recommendations on YouTube by jlebar on Saturday January 26, 2019 @01:22AM (#19005116)
All the AI stuff is really still incredibly dumb. Same for Amazon: Buy a mixer and suddenly you will be haunted everywhere by ads offering more mixers

Do we know it's actually dumb?

Like, given that you just bought a mixer, perhaps in an absolute sense, you're unlikely to buy a second mixer. But relative to everything else that you might buy, perhaps you are very likely to buy a second mixer!

There are huge financial incentives to get these ads right. I'm more inclined to guess that Amazon is doing something right than that they're leaving money on the table.

Re:Continuing our work to improve recommendations on YouTube by ddebernardy on Saturday January 26, 2019 @03:22AM (#19005445)
Just spitballing here, but if you just bought a mixer, your odds of buying another mixer is lower than that of buying a cook book or some other related stuff -- which Amazon could totally figure out mining the other stuff people bought at the same time or shortly after a mixer.
Re:Continuing our work to improve recommendations on YouTube by edanm on Saturday January 26, 2019 @02:45PM (#19008029)
It's still not necessarily the better idea for them.

Some percentage of people who just bought a mixer will get a product they aren't satisfied with, for whatever reason. That population has a much higher change of buying another mixer than other people. And mixers are probably far more expensive products.

I have no idea if, in an absolute sense, there's more chance you'll buy a cookbook rather than another mixer, but it wouldn't surprise me either way - on the one hand, obviously most people don't need two mixers. On the other hand, plenty of people who buy a mixer will never buy a cookbook in their life, but literally 100% of people who bought a mixer have proven that they were at some recent point in the market for a mixer. I'm not sure which effect wins out here.

Re:Continuing our work to improve recommendations on YouTube by tjoff on Saturday January 26, 2019 @05:39AM (#19005741)
There are huge financial incentives to get these ads right.

Not really. All you have to do is convince companies that machine targeted ads are worth it. Apparently that is orders of magnitudes easier than actually recommending anything relevant.

Which isn't hard to imagine either, there aren't much competition after all. So the primary purpose of targeted ads is that it functions as an imaginary hammer you can use to squash new competition with.

Re:Continuing our work to improve recommendations on YouTube by simongr3dal on Saturday January 26, 2019 @03:54AM (#19005529)
Alternativly there are many matching and complementing products for a mixer (or any other products for that matter)

Just bought a mixer, how about: A new set of mixing bowls, alternative whisks and hooks, perhaps a set of measuring cups, a kitchen scale?

Just bought a new smartphone, how about: an external battery, screen protector (that fits the specific model), a fancy charging dock?

It really doesn't seems like it would take much intelligence to create these sets of complementing products and then use them for further ad targeting in the future, you could even use AI and ML to create the sets in the first place so you can fill out your buzzword qouta.

Re:Continuing our work to improve recommendations on YouTube by paulddraper on Friday January 25, 2019 @05:46PM (#19003351)
You think that's dumb, but I bet the marketing data behind that says otherwise.

You think "I already have a mixer, idiot!" but in reality the chance that you are interested in buying a mixer just went from 1% to 2%. (Because you want another one, because you want to gift one, because you returned yours, etc.)

Re:Continuing our work to improve recommendations on YouTube by thaumasiotes on Friday January 25, 2019 @11:11PM (#19004725)
This has been discussed on HN before.

The last time I saw an Amazon employee weigh in, he said the ads are in fact not effective but the ad team isn't measuring it well. In particular, he claimed that he showed them that after adjusting for returns, the order-again rate for a particular expensive item was effectively zero, they asked him to repeat his study for another product, and he wandered off and found work to do for his own team.

Re:Continuing our work to improve recommendations on YouTube by paulddraper on Saturday January 26, 2019 @03:30AM (#19005470)
My skepticism level remains high. This is billions of dollars, with easy access to tons of data.

I'm not sure if put more stock in the government successfully hiding a fake moon landing, or analysts in one of history's most successful companies not running the numbers.

Re:Continuing our work to improve recommendations on YouTube by elihu on Friday January 25, 2019 @07:09PM (#19003818)
I think the problem here is that "recommendations that maximize expected marketing revenue" and "recommendations that maximize user experience" are not the same. Most of us would prefer Google to present us the latter, but they have considerable financial incentives to present the former instead.
Re:Continuing our work to improve recommendations on YouTube by paulddraper on Friday January 25, 2019 @08:02PM (#19004021)
> "recommendations that maximize expected marketing revenue" and "recommendations that maximize user experience" are not the same

That, I think, hits the nail on the head.

I think it explains the Amazon case quite well. I'm not so sure sure about the YouTube idiocy.

Re:Continuing our work to improve recommendations on YouTube by maxxxxx on Friday January 25, 2019 @06:01PM (#19003455)
The mixer probability maybe went from 1% to 2% but there must be 100s of other products they could show that have much higher probability. Maybe a mixing bowl or a cookbook?
Re:Continuing our work to improve recommendations on YouTube by nradov on Friday January 25, 2019 @08:37PM (#19004176)
I'll take that bet. Show us the marketing data.

So far it looks just as dumb to me.

Re:Continuing our work to improve recommendations on YouTube by kolinko on Friday January 25, 2019 @06:51PM (#19003725)
Some other online store, but related :)

I once bought crab meat in a can from an online store (hard to get it any other way in Poland).

A banner for buying more canned crab meat kept haunting me for the next two years, in various internet locations :)

Re:Continuing our work to improve recommendations on YouTube by Scoundreller on Friday January 25, 2019 @08:06PM (#19004038)
I guess that’s proof that consumers overpay for it.
Re:Continuing our work to improve recommendations on YouTube by rasz on Saturday January 26, 2019 @05:01PM (#19008790)
> All the AI stuff is really still incredibly dumb.

and scarily brilliant at the same time. I was looking up Newtek some time ago of the Amiga Video Toaster fame. Googled Wil Wheaton promo video (Will worked for Newtek :o), Penn Jillette (also promoted it) and finally Kiki Stockhammer (spokesperson). Here is where things get interesting. Google spewed 1997 "How To Have Cybersex On The Internet" https://www.youtube.com/watch?v=PBDKEn-TeQg as the result for Kiki haha, why would it do that? Turns out this "instructional video" was made using Video Toaster and features one of the iconic special effect transitions

https://youtu.be/PBDKEn-TeQg?t=9s

here on Toaster demo reel https://www.youtube.com/watch?v=K1OVWfmynPw&feature=youtu.be...

so what does it have to do with Kiki Stockhammer? This transition is Kiki doing poses on the green screen. Google search somehow connected this information together and was able to recognize Kiki Stockhammers silhouette on random youtube meme video :o !

This isnt the first time I got image search results that could only happen if Google was analyzing actual image content. Microchip part searches often return pictures of random boards featuring that particular IC despite no bom list in the html, Google is OCRing images and recognizes part numbers. Same with faces, cars etc.