Most privacy discussions focus on encrypting data, but very few people talk about metadata.
Even if nobody can read your request, they may still know who sent it, where it came from, and when it was made.
That's why I found OpenGradient's Oblivious HTTP Relay architecture interesting.
The relay can see the user's IP but not the content. The gateway can process the content but doesn't know who originally sent it. Instead of putting complete trust in one party, the information is split between different participants.
I like this approach because it doesn't rely on users blindly trusting a platform's privacy policy.
As AI becomes a bigger part of everyday applications, the challenge isn't only about building smarter models. It's also about making sure users don't reveal more information than they need to.
That's what makes this design interesting to me. Privacy isn't being added later as a feature - it is part of how the system works from the start.