The Day in AI

• Google lets US users aged 18 and over build an Android app in AI Studio and install it privately through a Personal Library in Google Play.
• Alibaba's Qwen team reports its QwenGyre framework trains long-running coding agents up to 1.85 times faster than the Colocate setup.
• Alibaba opened applications for Qwen Go!, an agent in the Qwen app that files flight-delay claims and requests price-protection refunds for users.
• Meta launched Muse Wonder and Muse Wonder Full, and Digital Trends reports that Artificial Analysis ranks the Full model second, behind Claude Opus 5.5.
• SpaceXSI merged X Premium, Grok and Cursor into The X Subscription, which has five tiers from free to $200 per month.

The user describes the app in plain language in AI Studio, the model writes the code, and the finished native Android app goes to a Personal Library in Google Play. Only its creator can see the app there, and it skips Google's review. Until now, running a self-built Android app meant publishing it on Google Play or installing it from a computer. ZDNET reports the feature is free. Google says sharing with a small group of friends and family will come in the coming months, along with a path to public release after trust and safety review.

Alibaba's Qwen team published QwenGyre on 12 October. It is a reinforcement learning framework for coding agents that run for hours inside unmodified harnesses such as Claude Code, with rollouts of close to 1 million tokens. An elastic scheduler moves idle GPUs between generating rollouts and training while the agents keep running. A trajectory processor merges branching agent histories into trees that share their common prefixes.

Alibaba reports a best case of 1.85 times faster than Colocate, on TerminalBench with Qwen 3.6 122B. On the 2.4T Qwen 3.8 model the gain over Colocate was 1.21 times, because many queries ran close to the timeout. On that model, 48 training steps took 75.42 hours, against 134.55 hours for an asynchronous setup. Over those 48 steps the NL2RepoBench pass rate rose from 52.48% to 58.54%.

Alibaba also opened applications on 13 October for Qwen Go!, a personal agent in the Qwen app, with access granted in phases. It works on outside platforms the user has authorized. It can compare options and contact customer service on the user's behalf, and its actions show on screen so the user can stop it. In one tester case described by Leiphone, the hotline was blocked, so the agent moved to online chat and secured a 70-yuan flight-delay payout.

• Meta launched Muse Wonder and Muse Wonder Full on 12 October with Active Memory, which keeps user preferences across chats, and Active Focus, which keeps trying other approaches when a task fails.
• Liquid AI released LFM2.5-24B-A2B, an open-weight mixture-of-experts reasoning model with about 2 billion of its 24 billion parameters active, at $0.05 per million input tokens and $0.70 per million output tokens on OpenRouter.
• Meta brought its Your Algorithm feed controls from Instagram to Facebook Feed and Reels, so users can describe in their own words what they want to see.
• Z.ai announced GLM-Legal, a legal agent model built on GLM-5.3-Flash, and reports a 19.2% all-pass rate on Harvey LAB.
• Zhipu began a limited test of GLM-5.3 HighSpeed in its ZCode coding tool, and one tester reported output above 490 tokens per second, according to ZheDongXi.
• Wilson Vu, who ran AI and GPU platform hardware architecture at Microsoft, joined OpenAI to work on compute architecture.
• Character.AI filed for voluntary Chapter 11 bankruptcy, and Ars Technica reports that Disney has offered $20 million for its chatbot technology.
• Applied Intuition has received investment offers that would value it at $30 billion or more, up from $15 billion in June, according to The Information.
• Jack Clark, Anthropic co-founder, wrote in Import AI that full recursive self-improvement now looks more likely to become possible in 2027, a year earlier than his previous forecast.