Email or username:

Password:

Forgot your password?
3 posts total
Siderea, Sibylla Bostoniensis

I've noticed a thing about the YouTube recommendation algorithm.

I had heard about the new Beyoncé single Texas Hold 'Em, so I searched for it on YouTube so I can listen to it. Subsequently, it started proposing to me various dance videos that people had made to that song. Perfectly reasonable. I watched several of those too. Then I got bored and wandered off.

After that, I noticed that the YouTube algorithm was absolutely convinced I wanted to see completely random videos that happened to use Beyoncé's Texas Hold 'Em as background music.

Which, yes, is the way that machine learning often fails: it gloms on to some detail, fails to note any sort of context bc it doesn't *do* context, and concludes that's the relevant detail to optimize for.

It was a pretty striking example, and suddenly I realized: this is not the first time. Until this happened, I hadn't noticed, but YouTube's recommendation algorithm will often show me video after video with the same *background music*.

I've noticed a thing about the YouTube recommendation algorithm.

I had heard about the new Beyoncé single Texas Hold 'Em, so I searched for it on YouTube so I can listen to it. Subsequently, it started proposing to me various dance videos that people had made to that song. Perfectly reasonable. I watched several of those too. Then I got bored and wandered off.

Siderea, Sibylla Bostoniensis

For instance, I noticed a while back that umpteen zillion random shorts wound up in my feed that all had Jain's Makeba as the soundtrack. I had just thought it was wildly popular, because it is indeed a super catchy bop.

In retrospect I think maybe what happened was that the YouTube algorithm noticed I – entirely coincidentally – watched all the way through two random videos which happened to use the same soundtrack, and concluded what I like to see is videos that use that song.

Same thing happened with Paris Paloma's Labour. And the Longest Johns' Wellerman. And some group of basses singing Hoist the Colors (which, delightfully appropriately, resulted in a stream full of 100% unadulterated thalassophobic naval architecture porn.)

For instance, I noticed a while back that umpteen zillion random shorts wound up in my feed that all had Jain's Makeba as the soundtrack. I had just thought it was wildly popular, because it is indeed a super catchy bop.

In retrospect I think maybe what happened was that the YouTube algorithm noticed I – entirely coincidentally – watched all the way through two random videos which happened to use the same soundtrack, and concluded what I like to see is videos that use that song.

Siderea, Sibylla Bostoniensis

Huh. Google Maps on Android is no longer plotting walking paths when returning public transit directions.

It's doing the following instead.

(To generate this example, I requested public transit directions from Park Street Station to the Harvard Science Center.)

Show previous comments
Sova

@siderea I noticed that today as well and it appears that it does it on desktop as well now.

Aljoscha Rittner (beandev)

@siderea
I checked it for Germany and here it's working (Android 14, Pixel 6pro) with public transportation routing.

anufea

@siderea I noticed this as well when I tried to find out whether there was a pedestrian crossing at a particular spot in a junction the other day. Completely useless.

Siderea, Sibylla Bostoniensis

Honestly, the most alarming thing about AI isn't so much about AI itself, but about how utterly hell-bent humans are to use it for things that it does a bad job at. H sapiens is bound and determined to use this chisel as a screwdriver.

Case in point: the recent news story about a lawyer (or pair of lawyers - finger pointing is underway) who submitted a filing in federal court which had actually been written by ChatGPT.

What is the one thing one can assume everyone has heard about LLMs? That they make completely bogus shit up, including inventing nonexistent citations.

It would be hard to overstate how unacceptable to a court it is for a lawyer to submit a legal argument which cites nonexistent case law. That's the kind of shit that can get a lawyer disbarred. It's a, uh, *career-limiting* move.

But apparently some lawyer actually did it: he took the output of a computer program famous for fabricating false citations and piped it directly into a court.

1/?

Honestly, the most alarming thing about AI isn't so much about AI itself, but about how utterly hell-bent humans are to use it for things that it does a bad job at. H sapiens is bound and determined to use this chisel as a screwdriver.

Case in point: the recent news story about a lawyer (or pair of lawyers - finger pointing is underway) who submitted a filing in federal court which had actually been written by ChatGPT.

Siderea, Sibylla Bostoniensis

This is working out about as well as one would expect.

Now, in this case, the fool is hoist by his own petard, but, alas, the general tendency to use AI for things it is bad at is already racking up examples of it hurting third parties. The really obvious example of this is using AI in policing, where racist policing practice provides the training data for predictive models that successfully capture racism in the model.

I have the strong suspicion that there's actually a whole bunch of undetected examples in the wild, already fucking things up for everyone.

For instance...

2/?

This is working out about as well as one would expect.

Now, in this case, the fool is hoist by his own petard, but, alas, the general tendency to use AI for things it is bad at is already racking up examples of it hurting third parties. The really obvious example of this is using AI in policing, where racist policing practice provides the training data for predictive models that successfully capture racism in the model.

Go Up