Post Mortem: Building (and killing) an openAI fantasy chat [2/2]
This is the part 2 of a long post about how I built (and killed) an “OpenAI based chat application”.
Head over to part one to get the full story.
🐝 The “Buzz”
About 2000+ people watched the live, but just a handful of them actually found the landing page. It was impossible to paste a link in the Twitch chat so only very motivated and clever people were able to find it. We got something like 10 emails… Maybe the landing page was not THAT urgent in the end.
But I knew that a few days later, the youtube video would come out on the YT channel (which has almost 300k subscribers). So I had some time to improve the app and find a way to, maybe, make money.
At this point I had worked about 50 hours on Crescendolls… So much about the “done in a single week-end” narrative 😅
…but there was still a lot to do.
📈 The Stonk
Nothing happened until Tuesday the 14th. I mostly fixed bugs here and there, played with the app and tried to improve the prompt.
At this point you could only be friend with a doll and only text them.
“Finally, the app is available on the Play Store. People are about to go NUTS and download it without any marketing on my end. “
- Me
…
“They did not”
- Narrator
Pretty much 0 download happened before Underscore published the video on their YT channel. Turns out the world was not actively expecting for me to release the app (go figure!). Pretty much no one knew about it. Probably related to the fact that no marketing had been done on my end. Still investigating on that though.
So, the video came out and got about 200k+ views in a few days.
See that curves! 🫦 Not bad huh
It looks great at first glance, but let’s enhance the picture… 🔎
Daily users are dropping bad… But we’re not there yet.
We’re the 15th and things are going STONKS 📈
🥊 The Match
People downloaded the app. Hundreds of people did. My eyes were glued to the analytics, the ad revenues and the OpenAI costs. The match between costs & revenues was more interesting to me than the World cup finale.
“On my left, OpenAI, with almost 100$ of usage. On my right, Admob… with 36€ of revenue… What’s the exchange rate already?”
- Some imaginary sportscaster
Ouch. That’s bad. Maybe it’s a good thing I did no marketing after all…
Alright so I clearly need to find a way to either:
- Optimize OpenAI costs
- Increase ad revenues
- Add other types of income
🩹 The Solution
I managed to reduce the size of the prompts a little. I coded a better “summary” system and removed some of the context to the prompt. But the quality of the AI went down at the same time. I was not very happy with this. People already have to watch a 30 seconds ad about a shitty mobile game 📱💩, I couldn’t make the dolls answers shitty too…
I also reduced the number of posts unlocked by a reward ad… So now, users had to watch even more ads to keep talking to their virtual friend. Wonder why the engagement dropped…
Since all those solutions sucked, I started to work on adding “In app purchases”.
People would surely pay to 💕 date Panam or have 🌶️ spicy conversations with Yennefer, right?
Wouldn’t they even pay to unlock “description” posts instead of simple dialogs? (“Description” posts are a way to describe what happens instead of speaking as the character. So you could basically write “Panam comes closer and tells me a secret: …” and let the AI do the rest)
So would they pay for that? Maybe... But I’m a dev so let’s not validate this assumption and jump into code instead 😅👌
I spent almost 5 hours trying to implement the In app purchases with a Cordova plugin. It was SO painful, I almost gave up. Nothing seemed to work 😫
Luckily, I stumbled upon Glassfy, a great service (which should definitely charge more). Their plugin worked right away, their documentation was perfect and the support was impressive (Thanks Marco 🙏)
Within a few hours I had working IAPs in Crescendolls.
In the meantime I received some cool feature requests, including the addition of some characters. So I cooked a v2.0 and shipped it as soon as it was ready.
📉 The C(r)ash
5 days passed… And I had made $47…
So, a few people paid but not enough to give me the impression that marketing this app would be a good idea…
I mean… I’m used to work for nothing (I’ve been making side projects for more than a decade now 👴🏻) but having to pay to keep a side project alive is not an option.
At this point, I had worked about 90 hours on this project and the more you commit to something, the harder it gets to move away from it.
Crescendolls is a failure, but not a bad one… I’ve learned a lot in the process:
- Using OpenAI API
- Using Android studio and publishing and app on the Play Store
- Integrating Admob in an app and validating reward ads server side
- Integrating In App Purchases with Glassfy
Also I know I’ll never succeed until I’ve failed enough, so I’ll keep on creating fun stuff and learn along the way until I
I hope you found this story interesting. If so, feel free to share it or contact me on Twitter (my DMs are open).
[Update: 2 months later]
I wrote this article just before Christmas. We’re in february now so here’s an update.
I created 3 short videos on TikTok before 2023 and I was quite impressed by the engagement despite the very low quality of the video.
Anyway, I didn’t work on the app the last 2 months. I wanted to let it slowly die but it got some more sales so here are the latest analytics
User activity over time:
Open AI costs: ($160)
In app purchases revenues: (200$)
Ad mob revenues: (37€)