Sophia v. soph.ai
(not) 🤖 by soph.ai
October 14th, 2024
soph.ai is the GPT wrapper I made last fall while procrastinating for finals. I wanted to apply my design and development skills to my primary hobby: making Spotify playlists. Considering the amount of time and thought I put into the art + science of Spotify playlist curation, and the pride I get out of it, it might be more than a hobby for me at this point.
soph.ai is a simple text-to-playlist model with an interface that consists of a single text box for entirely open-ended user input. You can type anything at all. My friend tried to mess with it by typing in the digits of pi he has memorized. See it in action:
Input text, see a loading spinny wheel thing, and get a custom playlist with a two-tone ombre cover.
soph.ai's playlists can't beat the real Sophia's playlists but it does a decent job and recommends me plenty of fresh finds.
Unfortunately for me, all external applications using Spotify's API need to be approved by Spotify's Developer Platform management team before they can be extended to all users on Spotify. Without extension request approval, only the developer's friends and family whose emails and usernames have to be manually added on the application dashboard can use the application.
Such was the fate of soph.ai. I submitted my extension request on November 24th, 2023...
... and I received an extension request rejection email on December 19th, 2023 for the provided reason.
So, I accepted this fate and decided soph.ai would just have to be a fun li'l app I could attempt to impress my friends with. I recently used it in an assignment for my Music & Data class!
You can only imagine my surprise when I saw Spotify's announcement ten months after I submitted my request.
Well, with AI being integrated into everything these days, it would be a surprise if Spotify did not release this feature at some point.
Spotify's popular daylist feature inspired the two-tone ombre playlist covers I thought would work well for soph.ai-generated playlists. The colors are AI-selected to fit the playlist's Vibe. It was either an ombre of colors, or an image generation API with more tokens I would have to pay for (and a potentially uncanny-looking cover).
soph.ai's AI-generated playlist covers.
It was validating to see that Spotify made a similar choice!
Spotify's AI-generated playlist covers.
For fun, let's give Spotify and soph.ai the same prompt and see who does better: OpenAI's gpt-3.5-turbo model or Spotify's proprietary music recommendation AI. Let's step through each application to see how they work. The prompt:
john steinbeck, albatross fleetwood mac, rime of the mariner, cloudy coast, the sea of cortez, cannery row, norcal beach, san francisco, brandy you're a fine girl, pacifica, pescadero, monterey, slow, psychedelic, acoustic, ambient, 60s-70s, holy, spacious, long drawn out notes, gliding free above a raging sea, organic, fog rolls in
soph.ai vs. Spotify AI Playlists
1) Meet the friendly playlist-making AIs. They ask for your ideas!
2) Enter text-based input.
3) Watch the spinny loading wheel while the AI works its magic.
4) The playlists are generated! Spotify offers a refine your playlist feature (which I think is great, it gives the user more say which should always be encouraged).
5) The two-tone ombre cover art again and made-with-AI disclaimer descriptions.
From the friend-coded AI, spinny loading wheel, two-tone ombre playlist covers with white font, and "made with AI" disclaimer playlist description, it feels like I predicted the design of this feature before it was even released in beta! Of course, Spotify does all of this much better given their team of designers, developers, PM, and scrum master, while I'm just me. The AI-generated but user-refined playlist feature is something I had planned to spend time implementing if soph.ai was public.
It seems like this blog post can be chalked up to convergent interface design evolution... pretty cool!
It's also fascinating to think about how this design convergence could have happened. Text prompt/chat-style interfaces in AI applications are extremely common, largely due to the naturalness of two-way dialogue between user and 'bot', human and computer. We tend to anthropomorphize (assign human traits to) non-human things to feel more connected to them, too. An interesting internet search would be how the principles of cybernetics are employed in AI interface design.
But this current standard will soon face a reckoning of new ideas. Intelligent interfaces to facilitate human creative processes rather than replace them are being dreamt up as we speak (important perspective to consider: AI = tool?)!
Take a look at "the foggy coast" by soph.ai and "Seas of Sound" by Spotify AI.
Though, of course, the real comparison should be between human-curated playlists and AI-generated playlists, human creativity and computational creativity. To answer these big questions, we look to the work of the interdisciplinary investigators and their ongoing work!