In partnership with
Hello from The AI Night,
Today in AI:
OpenAI Launches Codex App For Windows
Google Introduces Cinematic Video Overviews inside NotebookLLM
Perplexity Computer Now Has Voice Mode
Image Source: Web
Here's the deal: OpenAI released a Windows version of its Codex app, giving Windows developers access to the full Codex experience without needing WSL or virtual machines. The release includes a purpose built, Windows-native agent sandbox.
The Breakdown:
The sandbox uses OS-level controls including restricted tokens, filesystem ACLs, and dedicated sandbox users to let agents run safely inside real Windows environments like PowerShell.
Users can run multiple agents in parallel, manage long running tasks and review diffs in one place.
No need to switch to WSL or VMs. Agents operate directly in existing Windows developer setups.
The Windows sandbox implementation is open source, available on GitHub under the codex-rs repository.
The bigger picture: A large share of enterprise and individual developers still work in Windows-native environments. Removing the WSL or VM requirement lowers friction significantly for that base. The open-source sandbox also gives teams a reference for building secure agent execution on Windows.
Image Source: Google Tweet
Here's the deal: Google launched Cinematic Video Overviews in NotebookLM, upgrading its existing narrated-slides Video Overviews into fully animated, AI-generated videos. The feature is available now for Google AI Ultra subscribers in English.
The Breakdown:
NotebookLM's tweet indicated Pro users are on the roadmap, though no timeline was given.
The bigger picture: This is Google stacking its model suite (LLM + image + video) into a single product workflow. It turns NotebookLM from a text synthesis tool into a multimodal content engine which raises the bar for how AI research assistants deliver outputs. Builders working on document to media pipelines should pay attention to this model orchestration approach.
Perplexity
Perplexity Computer Now Has Voice Mode
Image Source: Reddit Post
Here's the deal: Perplexity announced Voice Mode for Perplexity Computer, its agentic orchestration system that routes tasks across 19 AI models. Users can now speak instructions instead of typing to trigger end-to-end workflows.
The Breakdown:
Perplexity Computer is described as a general-purpose digital worker that operates the same interfaces users do, capable of executing workflows that run for hours or even months.
The computer uses Opus as its core reasoning engine to match each task to the best-suited model, running agents in parallel across 19 models.
The upgraded voice mode runs on OpenAI's GPT Realtime 1.5 model with over 25% improvement in tool call reliability and better voice expressiveness.
Currently available to Max subscribers via the web, with Pro and Enterprise access coming soon.
Max members receive 10,000 credits monthly, plus a one time 20,000 credit launch bonus valid for 30 days.
The bigger picture: Voice input to a multi-model agent system changes the interaction model from "chat with AI" to "delegate to AI by talking." If the tool-call reliability holds, this moves Perplexity Computer closer to a voice first operating layer for complex workflows, something no major competitor currently ships at this scope.
What do these names have in common?
Arnold Schwarzenegger
Codie Sanchez
Scott Galloway
Colin & Samir
Shaan Puri
Jay Shetty
They all run their businesses on beehiiv. Newsletters, websites, digital products, and more. beehiiv is the only platform you need to take your content business to the next level.
🚨Limited time offer: Get 30% off your first 3 months on beehiiv. Just use code JOIN30 at checkout.
What else you need to know:
Google expanded Canvas in AI Mode to all U.S. users, adding creative writing and coding support that lets people build interactive tools and dashboards directly within Search.
Cursor launched support for JetBrains IDEs including IntelliJ IDEA, PyCharm, and WebStorm through the Agent Client Protocol (ACP), giving developers access to frontier AI models without leaving their preferred editor.
Claude surpassed ChatGPT atop the U.S. App Store with over 500,000 downloads last week, driven by its Cowork desktop agent, model upgrades, and backlash over OpenAI's Pentagon deal.
Google released gws, a Rust-based CLI tool installed via npm that dynamically pulls commands from Workspace APIs, enabling automation of Sheets, Docs, Drive and Gmail with built-in AI agent support.
OpenAI with Stanford and the University of Tartu, developed a Learning Outcomes Measurement Suite to track how AI tools affect student learning over time, now being validated with 20,000 Estonian students.
That’s it for today’s edition of The AI Night.
Our goal is to cut through the noise, surface what actually changed, and explain why it matters.
If this was useful, you’ll get the same signal here tomorrow.