Thermostats

In machine learning, there’s the concept of “temperature”: a parameter which controls the amount of randomness that’s fed in to the model’s predictions. When interacting with LLMs through their web interface, you can’t usually set the model temperature yourself, but when making API calls (or using the OpenAI playground) you can.¹ I’d like social media sites and other platforms using recommender systems to include “thermostats” which allow consumers to adjust the amount of personalisation they receive.

I’m not a big social media user, but one place I’m very keen on having this feature is Spotify. Personalisation is meant to increase user engagement so that platforms can make more money from advertising (or keep users paying for premium services, I guess). But in my case, Spotify’s suggestions make the experience significantly worse. I’m constantly recommended the same songs when I’d much rather discover new music, and often I actively dislike the ones it persistently pushes me towards.²

For example, I find Billie Eilish’s song LUNCH pretty boring and repetitive, so I skip it whenever it plays. Yet Spotify seems determined to have me listen to it! It shows up on multiple Daily Mixes, and is the second song on a strangely high number of the “Made for you” artist radio stations I have.

It seems implausible to me that LUNCH sounds like the music of all those singers and is being suggested everywhere because of sonic similarities. Possibly Spotify is being paid to promote the song, either directly or via lower royalties? If so, and the song isn’t explicitly marked as sponsored, I think that would be quite a misleading and borderline manipulative thing for them to be doing.

What I really want is Spotify to have a dislike button. At the moment, you can block specific songs, but not albums or artists, and I’m sceptical that they use the information for personalisation anyway. There’s a similar flaw on dating apps: when you reject a profile you don’t select a specific thing as being the main contributor to that assessment, whereas when you sending someone a like – on Hinge at least – it’s on a particular image or prompt response. Again, I think the app is leaving personalisation data on the table, though it would make the UI a bit more complicated and require slightly more effort from users.

Even better, you could have the chance to choose between multiple algorithms. Twitter has open-sourced its recommender algorithm, but there’s no way for users to opt in to an alternative one – yet you could very easily imagine an App Store of algorithms that people switch between, with developers contributing their own improved approaches. (I’m sure I read an article or Tweet about this but looks like I didn’t save it and I can’t find it now; let me know if you know the original source.)

For me at least, being able to raise the sampling temperature of my recommendations might well increase the amount of time I spend on Twitter. I’m wary of getting stuck inside a bubble of opinions, and so I set limits on how long to use the platform for. If I could periodically switch to a less filtered feed and reassure myself that I’m still seeing a relatively balanced mix of content (albeit of higher average quality) using my normal algorithm, I would be a lot less concerned about letting my consumption be at the mercy of the recommender system.

I also don’t think that adding this as an option would significantly reduce the engagement of the standard user. Probably they’d just stick with the vanilla algorithm and get polarising content as they currently do – and so the attention business model wouldn’t be adversely affected.

Possibly there’s another commercial or strategic motive that I’m missing out on which would lead platforms (including Spotify) to want to have users (including paying ones) trapped inside unsatisfying echo chambers? I can’t really think what it would be though.

So maybe there’s something here about enshittification? Because they’ve got more content available than anyone else, Spotify can afford to be rubbish in other ways?³ Perhaps a little like how, because of its network size, Facebook can afford to be unbelievably difficult for developers.