Pretty cool. Shameless plug: My team at Anthropic is hiring people that can write accelerator kernels. Please reach out to <my username>@gmail.com if you want to make state of the art models faster :)
Very interesting. I've implemented something similar[1] using a pure Collaborative Filtering approach[2][3], that I think works better for me, but it's unable to recommend unpopular repositories.
The New York Times recommender system uses a hybrid approach (Content Based + Collaborative Filtering) called Collaborative Topic Modeling on top of LDA[4]. It would be interesting to try that.
Here at The New York Times we are using it to power some of our recommendation algorithms. We are actually training the models with Python and serving them with Go using gonum.
Our library was just open sourced (and still in my personal account, until we add more documentation): https://github.com/jbochi/facts
MJPEG is indeed much better for this. It uses less bandwidth (GIFs have no compression) and less browser memory (old frames can be discarded). When I was at globo.com, we developed this to serve animated thumbnails for live video streams: https://github.com/jbochi/live_thumb
I am glad you liked. I am not the author of the blog post, but I was a member of the team.
1: We just repackage it.
2: The problem is that EvoStream stores the manifests and chunks locally and we needed high availability. That's why we use an external data storage. We have had up to 30 simultaneous streams, with 7 bitrates each and 2 hours of DVR.