Transcript · 2026-11-12
Automated transcription of the published recording. Hosts and dialogue are AI-generated; transcription may contain minor errors.
[Alex]: Welcome to Overnight AI. I'm Alex.
[Jamie]: And I'm Jamie.
[Alex]: We're your AI-generated hosts, and here's your AI news briefing for November 12th, 2026.
[Jamie]: Today's news is about narrower roles for AI. Making a bounded choice, appearing in a driving interface, coordinating with other agents, and showing enough evidence for people to inspect the limits.
[Alex]: Google introduced the Gemini Decisions API powered by Gemini Decision 1. It is designed for closed-set choices instead of generated prose: yes-or-no questions, multiple choice, or scoring against a rubric.
[Jamie]: Google says developers can send one multimodal input—text, images, audio, video, or a PDF—with up to 1,000 independent questions. The company says the model returns probabilities and confidence scores in a single pass.
[Alex]: Google reports end-to-end latency of 150 to 250 milliseconds and pricing of $0.10 per million input tokens with no output tokens billed. It suggests uses such as policy checks, routing, context grading, and tool selection.
[Jamie]: A quick score can be useful in a workflow, but it is not automatically a trustworthy decision. Google's latency and calibration claims are its own. Teams still need to test whether those scores work for their particular policy, population, and failure costs.
[Alex]: Google also expanded Gemini Live to Apple CarPlay. The company says iPhone users can have hands-free conversations with Gemini in the car, including checking a calendar, summarizing emails, planning a day, or rehearsing for a meeting.
[Jamie]: Google says the conversations sync back to the Gemini app on the phone. The rollout starts globally for Gemini users on iOS in supported Gemini Live languages, after users update the app, enable microphone access, and have used Gemini Live on the phone.
[Alex]: Putting a general-purpose assistant in a driving interface makes distraction controls as important as voice quality. Google also says some personalized daily brief features are unavailable to people under 18 and to users in the European Economic Area or Nigeria.
[Jamie]: Next, Alibaba's Qwen Cloud opened Qwen-4-Max Preview, which it calls the next flagship multimodal Qwen model. Its documentation lists text generation, image and video understanding, deep reasoning, web search, tool calling, structured outputs, and a 1 million token context window.
[Alex]: It is deliberately a short-lived preview. Qwen Cloud says it began November 11th at 10:00 in the morning, UTC+8, and will retire after November 25th. Access is through the token plan, not a pay-as-you-go API.
[Jamie]: The documentation also does not publish benchmark results. So this is early access, not proof of independent performance. Developers should avoid turning a preview identifier with a stated retirement date into a durable production dependency.
[Alex]: A team affiliated with Anthropic and the Anthropic Fellows Program released MAMU-Bench, a benchmark for multi-user, multi-agent systems. It gives agents different users and a shared resource instead of one shared goal.
[Jamie]: The simulated settings include a shared API budget, a clinic calendar, a group food order or travel booking, and a merge queue with a release cutoff. The researchers evaluated frontier models from Anthropic, OpenAI, and Alibaba in solo, coordinator, silent team, and peer-to-peer arrangements.
[Alex]: The result is a useful warning. Coordination can break down even where a single coordinator works well. In the shared budget setting, a designated leader and explicit priorities improved outcomes. For group orders, requiring agents to read messages helped more than simply rerunning a failed task.
[Jamie]: That is evidence from simulations, not a guarantee for a deployed agent fleet. But it suggests that communication rules and authority boundaries may matter as much as an individual model's capability when agents represent different people.
[Alex]: Finally, a new paper proposes five indicators that frontier developers could disclose to track chain-of-thought monitorability. They cover whether reasoning traces can be accessed and interpreted, and whether monitoring detects meaningful problems.
[Jamie]: The authors audited disclosures from Anthropic, OpenAI, and Google DeepMind. None covered the full set, although every indicator had been reported by at least one developer. That makes fuller reporting feasible, but not yet standard.
[Alex]: The paper's caveat is independence. Public disclosures mostly rely on companies' own evaluations and configurations, which outsiders may not be able to reproduce. It recommends another audit with deeper access through embedded evaluators or an independent verification organization.
[Jamie]: Together these stories are about operational details: what a model is allowed to decide, where it appears, how long a preview lasts, how agents share responsibility, and what evidence outsiders can inspect.
[Alex]: Source links are in the show notes.