[hcta]

Do it client-side with tensorflow.js

[ethanwillis]

Another spin on this is to make it a DAPP(decentralized app) where users pay for invocations. Look at something like iExec.

I was toying a few years ago with iExec to do some bioinformatics algorithms at a low cost on an infrastructure that users could pay for, but that I wouldn't have to pay to host.

The benefit of this is that it doesn't have to just be client side. It can also be backend as well.

[penguinium]

Oh that's interesting! But won't a big model heavily slow down the website? Is there a model that you think might be tractable for this sorta thing?

[beaconstudios]

if you run it in a web worker, it should run on a separate thread to the website itself so it shouldn't cause the website any issues unless the model eats all the user's RAM.

[penguinium]

Is it reliable for mobile users? What if they visit another page before the model finishes evaluating?

[beaconstudios]

1. You'd have to try it out with your model

2. Then presumably they don't get their result, but that's to be expected if they navigate away