More information and a demo in the article.
After a year in beta, Ente’s machine learning is live for all.
You can search for people, places and a lot more; while retaining absolute privacy of your memories.
Multiple years of research and development have gone into delivering what we believe is a novel approach to running computer vision models on the edge. Along the way, we’ve overcome quite a few challenges.
We’ve published a whitepaper documenting the underlying novelties at ente.io/ML.
Monumental. Congratulations to those who burnt the midnight oil to pull this off (the blog post says, it took 2 years!)
For Photos, Google’s technological moat has always been ML and Ente’s strong v1 shows they are more than up for it.
all PII is zero access enrypted?
Yes, details are available @ https://ente.io/ml/#implementation-details
I’m not exactly versed in the technicalities to appreciate it for what it is, but the general concept seems very elegant and neat. Amazing what you’ve been able to achieve without a massive dataset. Serves to show that all this ML stuff can be done without it, and that training AI is not a universal justification for collecting everything, especially private pictures.
yeah you can just use transfer learning, also LLM doesn’t require pixels, as text is now as equivalent as images for Machine commands when AI is used.
Also things like neural engine on macbooks are available that have their own secure enclaves afaik, so you can do “edge computing”.
Nothing PII needs to be sent to servers even for “AI”.