Hi there!

I'm the creator of the XBlock screenshot labeller on Bluesky.

This funding page is to help support the continued development of XBlock. I hope to buy a new graphics card to help train the machine learning models that power the service. Currently, it's running off a Raspberry Pi 5 in my kitchen, with inference and training being done on my ancient gaming PC running a GeForce GTX 980 Ti.

So far, the setup has done pretty well! But I'd love a new graphics card for training future versions of the model, it's currently taking me upwards of a day each iteration.

I've poured a solid 100+ hours into this project so far, mostly hand-annotating screenshots reported to me by users of the service. Every morning I wake up and manually process a queue of 10-30 reports, then do the same before going to bed; collectively it takes me about half an hour of work per day. My hope is that the model becomes accurate enough that I only have to handle appeals for mislabelled screenshots (I get 2-10 a day currently). I then pull down all the images the service every 2 weeks sorted into folders by originating platform, which requires me to then go through each folder and ensure each image is correct before passing it to my training pipeline (this takes me about 5-7h per training run, each of the 12 categories supported usually has between 100 and 5000 images to verify).

These numbers belie the true volume of posts XBlock handles. It reads from the Bluesky firehose and runs inference on every image that gets emitted; at time of writing, this is around 2M images a fortnight, with over 100,000 posts having received an XBlock label so far.

I wrote a bit about the experience on my blog, aendra.com if it's of any interest!

Thank you for your support! 💚

-ændra.