Google’s private AI pitch meets the verification problem
Google is arguing that homomorphic encryption can let AI systems process protected data, but HN readers focused on a less glamorous question: how can a customer verify the service did exactly what it promised?
Homomorphic encryption keeps data encrypted while computation happens, offering a route to cloud AI that does not expose plain inputs and outputs to the provider.
The thread accepted the cryptographic appeal while questioning overhead, commercial viability, and whether privacy alone proves the cloud executed the intended model.
Encrypted inference is not the same as verified inference
Google’s security blog presents homomorphic encryption as a way to make private AI practical: data can remain encrypted while a service computes on it.
HN’s most useful response separated confidentiality from correctness. Encryption can conceal inputs and outputs, but a customer may still need a way to know the provider ran the requested computation rather than a substituted one.
"FHE guarantees only that you need the key to see the inputs or outputs, but not necessarily that the computation is the one you want." — an HN commenter
That distinction is where remote attestation, reproducibility, and service contracts enter the picture. They are different problems from encryption, even if a product presentation compresses them into one promise of privacy.
Overhead still decides the market
A commenter working on privacy-preserving machine learning said inference overheads can be around three orders of magnitude. The exact cost depends on workload and implementation, but the commercial test is clear: a private inference system must be economical enough that users will choose it over keeping the job on-device.
Get breakdowns like this every day
The top Hacker News stories, unpacked — by email, Telegram, or RSS.
RISC-V’s critics say openness did not prevent old mistakes
A detailed critique of RISC-V argues that its designers repeated problems computer architects already knew how to avoid. HN readers mostly accepted the trade-offs while resisting the verdict.