Apple is putting Alibaba's Qwen inside Apple Intelligence in China, with a November 4 developer beta and a December launch on iOS 27.3
AspireAI community · Apple Newsroom (China)
Apple's China newsroom announced that the first wave of Apple Intelligence features will open for testing with the iOS developer beta on November 4 and reach Chinese users with iOS 27.3 in December. The first wave covers Live Translation in Messages, FaceTime and Phone, Writing Tools, Visual Intelligence, the Photos app's Remove tool, Genmoji, natural-language photo search, Memories movies, Priority Mail, Mail summaries, Smart Reply, Notification summaries, Priority Notifications, Reduce Interruptions, audio transcription summaries in Notes, and call transcript summaries. The material change is the Alibaba partnership, which adds Qwen to Siri, Writing Tools and Visual Intelligence for text and image understanding: users can ask Siri about their photos and about documents including PDFs and presentations, and can let Qwen create original content from scratch in the system-wide Writing Tools. Access is by Qwen account or Sign in with Apple, and Apple says the rest of Apple Intelligence continues to run on its own generative models. The honest limitation is scope: the release states that Siri AI and the other next-generation features announced at WWDC 2026 are not included, and support for other platforms follows only after further approvals.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.