July 17, 2026
Secret AI or secret cat detector?
Homomorphically encrypted CIFAR-10 inference in 200ms
Privacy AI demo wows the crowd, but commenters smell a giant asterisk
TLDR: Belfort says its demo can classify images almost instantly while keeping them hidden, a big deal for privacy-sensitive fields. Commenters were torn between calling it a breakthrough and questioning whether the setup really protects users, while others roasted its habit of seeing cats in everything.
A startup says it can identify tiny pictures in under 200 milliseconds while keeping the image hidden from the server, which is the kind of promise that makes privacy fans sit up straight. In plain English: you send a scrambled version of a picture, their system guesses what it is, and supposedly never sees the actual image. That sounds like sci-fi for hospitals, banks, and governments. But the real fireworks happened in the comments, where the crowd instantly split into "game-changer" and "show me the receipts" camps.
The biggest mood? Impressed, but deeply suspicious. One commenter called fast encrypted computing a "holy grail" and basically said, amazing if real, but where are the details? Another went digging through the site and claimed there was a major catch: if the company is helping create the secret keys, then critics say the whole privacy flex gets a lot less magical. That turned the thread from "cool demo" into mini tech courtroom drama.
And then came the comedy. One user posted results showing the system was very sure a calculator was a cat, while another joked it must have trained only on dogs with floppy ears because it labeled a German Shepherd as a cat too. Add in worries that the data size allegedly balloons 341 times, and the vibe became: breakthrough or party trick? Either way, the community delivered the real entertainment: equal parts hype, skepticism, and people roasting an AI that apparently sees cats everywhere.
Key Points
- •Belfort showcases a demo for real-time homomorphically encrypted CIFAR-10 inference.
- •The demo is designed so users can classify images without the server seeing the image.
- •Users can select an image from a library or upload their own image, which is resized to 32x32.
- •The workflow is presented in three stages: pick an image, run encrypted compute, and decrypt and check the result.
- •Belfort says real-time encrypted ML could support privacy-sensitive applications in finance, healthcare, governments, and other domains.