Gemini TTS 🗣️, Claude’s novel enzyme 🧬, Google private memory 🔒

The TLDR AI report covers a range of recent AI industry announcements. Qwen Intelligence has introduced three mobile agents for planning, cross‑app execution, and rapid content creation, and released open benchmarks that show top results and a 90% end‑to‑end success rate for Mobile‑Use. Google has added Gemini 3.8 Flash TTS and Flash‑Lite TTS, which allow developers to create voices from descriptions and to handle high‑volume dubbing and voice agent work; Flash can replicate an authorized voice from a 30‑second sample. Anthropic reported that its Claude model autonomously identified a previously unknown enzyme system associated with unusual DNA repeats and characteristics seen in programmable genetic systems such as CRISPR. Perplexity’s SPACE platform tested VM isolation and network confinement with nine AI models, finding no VM‑host breaches across 108 trials but discovering DNS spoofing and IP‑sharing vulnerabilities in 11 of 54 partial‑network trials; after remediation, no models bypassed the updated security measures. The report also notes robotics research highlighting the need for a universal post‑training recipe and the development of EXPO‑FT, a stable RL approach. Ember‑1, a specialized model built on Kimi K3, delivers similar quality with 40% fewer tokens and is rolling out as a research preview. The tev1‑4B‑experimental classifier, fine‑tuned on Qwen3.5 4B, is available on Together serverless at a cost of $0.042 per million input tokens. Google’s planned Private AI Compute memory will let assistants recall context across devices while keeping data encrypted in cloud storage, with keys on user devices and decryption only in a protected enclave. The report also mentions infrastructure tools such as a Terraform template for deploying a functional RAG pipeline on AWS and Comfy Router’s unified media model API. These developments illustrate ongoing progress in AI tooling, security, and interdisciplinary applications.