Google's HEIR compiler runs AI inference on encrypted data
Servers compute on data they can't read — plus Nvidia's router that cuts model bills 74% and Copilot's Sept 1 cull of six models.
Copy markdown
Inference without ever decrypting
HEIR compiles pre-trained models to run on homomorphically encrypted inputs, so a server returns a result without ever seeing the plaintext. Google shipped four working demos — recommendations, card-fraud detection, network-threat detection, and audio hotword spotting — all open source on GitHub, with a stated goal of one-click encrypted inference for non-experts.
The catch: ~1,000x overhead
Homomorphic encryption still costs roughly 1,000x normal compute, and Google's latency numbers are single-threaded CPU. Hardware partners (Belfort, Niobium, Cornami, Optalysys) are chasing acceleration — so prototype privacy-preserving inference now, but keep it off your hot path until the silicon catches up.
Nvidia's open router cuts model bills 74%
NeMo Switchyard is a fully open-source router that sends agent calls across models while staying OpenAI-, Anthropic-, and Responses-API compatible. A LangChain benchmark cut cost 74% by routing just 7% of calls to a frontier model; Cognition's Devin ran about 28% cheaper than Opus 5 while staying within 2.8 points of accuracy.
Copilot drops six models on Sept 1
GitHub Copilot retires Gemini 3.1 Pro, Claude Opus 4.5 and 4.6, Sonnet 4.5 and 4.6, and Raptor Mini across every experience on September 1. Pin replacements now — Opus 4.7/4.8/5, Sonnet 5, Gemini 3.6 Flash, MAI-Code-1-Flash — or your agents silently reroute. Sonnet 4.6 survives only for annual individual subscribers.
Elsewhere: an open agent that proves math in Lean
MathCode turns plain-language problems into Lean 4 theorems and proves them — a persistent Lean REPL with ~0.4s checks, Mathlib lemma search, and tree-of-subgoals parallel proving. The latest build adds custom tools and skills, so you can wire formal verification into a coding agent this weekend.