The Day in AI

• ByteDance ended the 30-day notice for its SAEP protocol on 14 October, so the Doubao phone agent will gradually start operating apps that have not opted out.
• SpaceXSI released Grok Voice 3.0, which scores 86.9 and ranks first on the Artificial Analysis speech-to-speech index, at $0.08 per minute.
• ModelBest released MiniCPM-V 4.7 with open weights, including a 35B-A3B mixture-of-experts version it says runs in about 3GB on an iPhone 17 Pro Max.
• Sachin Katti, OpenAI's head of compute strategy, is reportedly leaving on 15 November after about a year at the company.
• StepFun launched the STEPX Neo, its first phone, at 3,999 yuan ($597), running its own Step AOS 1.0.

The agent reads the screen and simulates taps and typing. That means it can operate an app without the developer building an integration. Any app whose developer has not declared a refusal becomes available to the agent in stages. Android developers can set limits for the whole app, for one page or for one business intent, and they can opt out at any time. Payments and account authentication can require the user to confirm, and the agent's decisions are logged for audit. National Business Daily counted 151 apps on the refusal list, 65 from Tencent's ecosystem and 23 from Alibaba's.

SpaceXSI released Grok Voice 3.0, a speech-to-speech model that replaces Grok Voice 2.0 at the same price of $0.08 per minute. Artificial Analysis gives it 86.9 on its speech-to-speech index, which puts it first. Artificial Analysis also measured 97% on Full Duplex Bench and 71.6% on τ-Voice. SpaceXSI says that τ-Voice score is comparable with GPT-Live-1 at 40% lower latency. ServiceNow rated it highest on EVA-Bench enterprise scenarios at 74%. SpaceXSI reports that it resolves about 70% of the roughly 20,000 daily Starlink support calls without a human. Users on grok-voice-latest get the upgrade automatically.

ModelBest's MiniCPM-V 4.7 is the first mixture-of-experts model in the MiniCPM-V series. A mixture-of-experts model sends each token through a small subset of its weights, so the 35B-A3B version uses about 3 billion parameters per token. ModelBest says a new inference method, lookahead expert scheduling, decodes at over 13 tokens per second in about 3GB of memory on an iPhone 17 Pro Max. It reports 91.2 on OCRBench. ModelBest also open-sourced Self-Replay Fine-Tuning, a recipe for adding capabilities while limiting forgetting. ModelBest hasn't said exactly how lookahead expert scheduling predicts which experts to load.

• Sachin Katti announced on X that he is leaving OpenAI and is expected to stay on as an adviser, and he said his team secured double-digit gigawatts of compute.
• Uday Ruddarraju, a former xAI executive who joined OpenAI last year, will lead OpenAI's infrastructure and compute businesses.
• StepFun sells the STEPX Neo with dual screens and the Amoo assistant, and it uses Model Context Protocol to reach apps such as Meituan.
• Alibaba launched the Qwen N1 and N1 Pro glasses in China on 13 October from RMB 2,899, with batteries that swap out of the temples.
• Fish Audio released Drama 3, an open-weights speech model with 35B total and 3B active parameters, which it says decodes 271 frames per second on one H200.
• Google DeepMind and GovTech Singapore fine-tuned a Gemma model to detect local scams on the phone, and rollout inside apps has not started.
• Meta launched Muse Wonder and Muse Wonder Full on 12 October, and Artificial Analysis ranks Muse Wonder Full second, behind Claude Opus 5.5.
• Nous Research raised $90 million to bring its open-source assistant to enterprises, according to the Wall Street Journal.
• Anthropic struck a licensing and acqui-hire deal with an AI security startup, The Information reports.
• China approved Alibaba models for Apple handsets, according to the South China Morning Post.
• Character.ai filed for Chapter 11 bankruptcy, citing costs and litigation, The Information reports.