AI-powered movie recommendations
Hey guys
I'm a ML Engineer and have been working on interesting NLP problems for quite some time now. Naturally, the emergence of GPT3 has been super exciting and I've been playing around with it a little for this side project of mine: https://www.givememovies.com
I've been frustrated by poor the recommendations of Netflix, Amazon Prime etc. I often find myself spending a lot of time finding a good movie to watch. But as it turns out, GPT3 is really good at recommending movies!
So I built this website where you can enter a query (in natural language) and it gives you the most related movies to your query. Next step is to add links to Netflix, Youtube & Co depending on where a movie is available.
What are people thinking on the monetisation of this project? Affiliated links and ads on the website are the obvious candidates. Do you think it's possible to turn this into a profitable project? (I'm happy to continue using it just by myself in any case :D )
Strange, I am also in NLP, and the emergence of GPT3 was not super exciting for me but super fustrating! 😄 Only 1% of ML engineers can work on such a huge model, and all the others are kinda "needless"...
I'm excited for the field of NLP/ML as a whole. And about all the productivity gains these models enable. Thinking about how much time github copilot alone is saving me.. imagine you can achieve these productivity gains in other industries. One way to look at it is that some jobs become useless, but I take the view that it enables me to do a lot of other cool things. But I guess people have this argument every time there is a new, disruptive technology :)
Those are some cool movie recommendations for sure but I just want to watch ‘Formula 1 2023’ On ServusTV because it is my most favorite sport and I watch every single race.
Some advice:
I think this could be very useful, but only if you expand it more away from movies (especially netflix movies).
I tried to get it to recommend me anime and it only recommended me movies + nothing that was too close to what I asked for.
So if you could expend it to shows and books as well (maybe even youtube channels and podcasts in the future), then I think it would be much more useful.
Also curious what goes on in the backend, do you just route GPT-3 to rotten tomatoes and IMDB synopsis pages? Or is there more?
I would not be surprised if GPT-3 just has the power to make good recommendations based solely on that, but I'd imagine more info is good.
For anime, you could try routing it to crunchyroll or MyAnimeList
Another useful feature would be collecting user feedback. I don't know how feasible it would be for you to fine tune GPT-3 for your needs, but if that could be done based on user feedback, it would be far more useful.
As for monetization, I feel like this would be difficult. You might be able to cram the page with ads to make up for the cost for free. However, you could also charge a premium for more searches per month and limit free searches.
Another way to do it (which would depend on how big it gets), would be to allow some ppl to run promotions for their own films/podcasts/books that show up first place for searches.
Just my 2c
Btw, feel free to share it in our AI discord
I think a lot of people there would enjoy it!
FYI, I'm working on Evoke, which is an AI as a service that hosts opens source AI models on the cloud for devs to run via API for their AI apps.
Heard Stability is releasing an open source LLM on par with GPT-3 this year. If that happens, we'll probably get it on the site and offer an API. And I think it would likely be cheaper than GPT-3, so perfect for your app.
That's still a long ways off though, since we're just finishing up our stable diffusion API lol. But you can keep track of my progress on twitter: https://twitter.com/TheRealEtch
Hey Richard thanks a lot for your extensive feedback!
You raise a lot of good points. Extending it to series, not just movies, should be fairly easy. Haven't thought about adding youtube channels and books, but those are good ideas too so thanks for that.
Do you mind me asking what your query was for which it didn't give you good recommendations? just for debugging purposes.
As for how it works: I use gpt just to get recommended movie titles and pull meta information from elsewhere. I found that GPT can retrieve movie titles with good accuracy but would make up things like ratings, plot and imdb id.
Thanks for sharing your evoke discord, I'll check it out now.
Sure! I tried to see how good it was by testing if I could get it to recommend Death Note (it's relatively popular, so my guess is that a good recommender should be able to pick it up). My query was:
"anime where protagonist relies on his wits with a cat and mouse chase". It's not incredibly specific, but death note would come to most people's minds.
I got recommended Ghibli films for that lol. In fact, no film that I recognized actually had anything near what I described.
I recommend you add a logger that tracks what ppl enter. It's not like there's an sensitive info attached to visitors, so privacy isn't an issue.
Tracking would give you more insight on how to improve.
PS: just tried it again on "tv shows" with: "cat and mouse detective anime where main character is a villain" and results were pretty good (death note was no. 1). Also tried with old prompt: "anime where protagonist relies on his wits with a cat and mouse chase" and also gave death note as no.1
However, anything below that seems to lose a little coherence. It seems to shy away from less known anime.
Honestly, it feels like google search returns better results for now.
Also, feel free to ask for feedback and advice for this in the evoke server.
There's not a lot of engagement rn, but someone might respond.
just added the option for TV shows ;)