The server never sees it

Homomorphic encryption has been mathematically available for a decade and practically unavailable for almost as long. The missing piece was never the maths. It was a compiler.

Google has open-sourced HEIR, a compiler toolchain that converts pretrained models to run inference on encrypted inputs. The server computes on ciphertext and never sees the data.

Why this was stuck

Not for a mathematical reason. Homomorphic schemes have existed and worked for over a decade. The obstacle was that writing an efficient homomorphic version of a model by hand is specialist work almost nobody can do, which put it permanently out of reach of the teams who needed it.

A compiler moves it from a cryptography problem to a build step. That is a smaller-sounding change than it is: the history of any technology becoming ordinary runs through the moment somebody wrote the compiler.

The caveat is not small

Homomorphic inference is orders of magnitude slower than plaintext. A compiler makes it accessible without making it fast, and this is not a drop-in replacement for hosted inference on anything latency-sensitive.

Slow-but-permitted beats impossible.

Where it changes the calculation

In the class of workloads currently blocked outright. Medical records, financial data, anything under a residency rule that forbids plaintext leaving the perimeter — those are not slow today, they are impossible. A technique that is unusably slow for a chatbot is transformative for a workload that has no option at all.

Two answers to one question

Read it against the 3B vision model tuned for edge hardware from the same week. Running a small model locally and running a large model on encrypted data are two answers to the same question — where is sensitive data allowed to go — with entirely different cost profiles.

Local means you accept a smaller model. Encrypted means you accept a slower one. Which trade you make depends on whether your constraint is capability or latency, and for the first time both options are actually available to a normal engineering team.

The interface is where tools are competing now

Elsewhere in tooling, Cursor 3 shipped a Design Mode that lets you annotate UI elements in the browser to give the agent precise visual targets. It solves the specific ambiguity that has made agentic frontend work frustrating: backend instructions have named referents, and "the padding on the card" is three questions rather than an instruction.

Both stories point the same way. The models underneath these products are increasingly interchangeable and increasingly cheap. What differentiates is the surrounding engineering — how precisely you can express intent, and where your data is permitted to be while the work happens. Interfaces and compilers are harder to copy than weights.

Google — The latest AI news we announced → · AI Weekly — AI News Today, August 15 — Top AI Stories & Live Updates → · SitePoint — Claude Code vs Cursor vs Copilot: The 2026 Developer Comparison →