The curse of movie recommendations
Increasingly, I see people online unable to find something to watch. Or perhaps more accurately: people unable to choose something to watch.
Choice paralysis
In a previous job I worked on recommendation systems and one thing that often came up during testing is that adding more would lead to more engagement. More clicks, everyone's favourite optimisation. Distracting? Maybe! But people are clicking it so it must be good! As users we know this isn't true. It sucks to see 40 options and then be unable to decide; more isn't always more! The feeling of being paralysed by choice is not good. This is the problem faced by anyone choosing a film to watch, one amplified by the streaming wars. Attempts are occasionally made by offering a "shuffle" button to users, but it never works because who would use that? I want to have agency in what I watch!
It can't be one extreme or the other if you want to build a satisfying flow (or just avoid spending 30 minutes hunting and then giving up and reading a book ... or scrolling).
3000 is too many to choose between. But 2 isn't.
Recently, I've recently taken to using LLMs as a proof-of-concept Head-to-Head Bracket recommendation system, starting off with a prompt like this:
Hi, I'm looking for a movie recommendation - not sure what mood I'm in though. Let's approach it more like a twenty questions - you suggest two moves and I choose which one I prefer, then based on that you suggest another two and I pick which one I prefer and from there we narrow it down. Don't zoom in too fast, just because I chose a comedy over a horror doesn't mean I want to watch a comedy over a drama. If I say neither then zoom back out a bit for more serendipity.
You can add other constraints, for instance once you've nailed down the vibe insist on academy award winners or foreign-language, it's an innately flexible system. It probably goes a bit too niche too fast some times, but overall it's a really fun way to get out of the choicehole.
A downside of this approach is that you have to pay per movie if your eventual choice is not available on one of your subscribed streamers (justwatch can be great for checking). This is fine for me - who watches about 1 film a week - but might be a turnoff for others. I'd love to see a version of this built as a data-science app using the justwatch database so that it automatically integrated filtering to only films available on specific platforms. Maybe using thinking/web search modes you can get part way there. On the plus side, due to the structure of residual payments for sales/rentals of films, I'm pretty sure the people actually making the films get a lot more money from renting like this.
Here's a brief transcript of how it works, which led me to watching Big Fish (2003).
LLMs as proof of concept
I love being able to quickly game out an idea like this; previously, I'd have had to build a front end, find a movies database, embed all of the data somehow, then create a system of using some random sampling to triangulate the user's mood in embedding space. That would still be fun to do! I would love to see someone create that, and might eventually have a go myself, but for now just doing it this way is enough to see if it actually works. That's a workflow that didn't exist as recently as 2022, and I think it's under-utilised by people who treat LLMs purely as a replacement for search engines. Most importantly, it's fun, and leads to watching more challenging things. IMO that's a win-win.