Thanks a bunch! For launch/beta, we're paying for the GPU time ourselves. :) Quoting Matt below:
"We weren't really sure how to price it, so we're using the beta period for now to figure out what mix of models people are using and trying to figure out reasonable pricing based on that, and also ironing out various bugs and sharp edges. Then we'll start charging for it; personally I'd prefer to have it be usage-based pricing rather than the monthly subscriptions that ChatGPT and Claude use, so that you can treat it more like API access for those companies and don't have to worry about message caps."
Appreciate the feedback! We currently use fly.io as our cloud GPU provider, but we're actively investigating other providers due to various limitations (like NVLink support).
It means that we can spin up a single model server and use it for multiple people, effectively splitting the cost. Whereas if you try to rent the GPUs yourself on something like Runpod, you'll end up paying much more since you're the only person using the model.
We're working on fleshing out ToS, privacy policy, and company specifics, but just to answer your first question, I'm Billy Cao, an ex-Google eng, and Matt Baker is ex-Airbnb, ex-Meta.
Re: concerns, our infra will scale relatively well (several qps per model, probably), but we're still in the stages of fleshing things out and getting feedback. :)
Feel free to drop us a line at [email protected] if you wanted to chat specifics!
Great point. Right now we don't log or store any chat messages for the API (only what models people are choosing to run). We do store messages for the web UI chat history and only share it with inference providers (currently together.ai) per request for popular models, but I know some hand-waved details from an HN comment doesn't suffice.
I might subscribe to the oft-repeated mantra to not fix what isn't broken, especially considering that users cannot view other users' calendars, so mediocrity or spam is not of a major concern to you.
That said, awesome app. It's certainly a problem I'm sure many are constantly faced with. (I know I am) My first impressions though include a lack of detailed info (how does it work?) compounded by the inconvenience of making an account for the site. (I still haven't registered)
Edit: As for a small change you can do right away, I feel like users would consider "John Doe" <[email protected]> more intuitive than "Invite Name" <Invite Email>. (Consider making the name optional altogether or just having a more intuitive input method, like multiple <input> prompts)
I agree, I highly doubt that the percentage of storage that is non-unique is enough to increase Dropbox's profit margins significantly.
I'm certain they use this and several other techniques such that each user reflects a far less impact on Dropbox's storage than the 50GB bought, but if that's printing money than Amazon would be a first world country by now.
I'm wondering, how do you strike a balance between having a long runway and not falling into a sense of well-funded sloppiness?
Many suggest that the lack of funds forces a startup to be lean and mean, but I can see how appropriating funds has the obvious benefit of allowing you to estimate and plan your risks better.
Have you worked with any Node.js projects before? I'd actually say this is a relatively sparse list of dependencies for a user-facing tool.