大爱 AI 日报 AI Daily 2026/10/14: Gemini 4 Argon Officially Released, The Lancet: AI Extinction Risk Exceeds Nuclear Weapons

Today’s Summary

Gemini 4 Argon is now live, available to API and Pro/Ultra users, and callable in AI Studio.
The Lancet ranks malicious use of AI as a top extinction-level risk alongside nuclear conflict, even as developer toolchains continue to spread rapidly — both forces are accelerating in parallel.
Start today by trying Gemini 4 Argon in AI Studio, then verify whether Codex has removed the change-summary popup.

🔥 Today’s Top 10 Highlights

1. Gemini 4 Argon Officially Released

A new model has arrived. A release summary posted by 歸藏 on Jike shows that Gemini 4 Argon is now available to Gemini API users as well as Pro and Ultra subscribers on the Gemini app. It is also accessible via Google AI Ultra membership’s Antigravity feature and AI Studio. Developers already on Gemini Pro/Ultra can head to AI Studio to start using it now.

2. The Lancet: AI Extinction Risk Exceeds Nuclear Weapons

A leading medical journal issues its highest-level warning. 36Kr reports The Lancet: AI Extinction Risk Exceeds Nuclear Weapons . In a commission report published on October 11, The Lancet ranked malicious use of AI alongside nuclear conflict as one of the most likely causes of human extinction before 2100, out of 17 global threats assessed. The report was written by 35 leading experts and covers 204 countries. It notes that the two risks could amplify each other, and further states that the potential damage from AI exceeds that of nuclear conflict.

3. Muse Two-Week Retention Falls to 22.5%

A retention curve sparks industry debate. A Sensor Tower data chart analyzed by 莫唯书 on Jike shows Muse achieved a day-one retention of 73.3%, which fell to 22.5% by day 14, prompting Wall Street to question whether growth has peaked. The author counters with the Threads example: usage dropped by nearly 80% in the first month, yet monthly active users eventually reached 500 million. The real product cycle test for Muse is only just beginning.

4. Oneira Lets Generated Worlds Remember What You’ve Moved

Pan the camera back — your changes are still there. A team from Oxford, Monash, and other universities has proposed Oneira, which makes objects in generated scenes interactive and persistent . It uses a Coding Agent to maintain a world-state table that records object positions and changes, then passes that information to a video model to render the scene. Newly introduced objects can also be added to the table. The project demonstrates targeted object interaction and multi-step operations — developers working on world models or interactive video should check the demo.

5. 歸藏 Tests: GrokBot + HTML Layout Skill Combination

Long videos become readable long-form articles. A hands-on record shared by 歸藏 on Jike demonstrates an unpublished skill: formatting complex content as an HTML page and using it together with GrokBot 📄. His use case was turning an extremely long thread covering an entire 3D dragon modeling session into a well-formatted long-form article readable on a phone. For users who regularly need to digest large volumes of long content, this combination is worth trying once.

6. BaoCut 3.5 Integrates Qwen Image 2.1 Turbo for Local Image Generation

Local image generation on Mac no longer depends on a server. An introduction from the AI探索指南 channel shows that BaoCut 3.5 has integrated Qwen Image 2.1 Turbo, with image generation running entirely on-device 🖼️. On a Mac Mini M2, generating a 1080p image takes roughly 3–5 minutes; 2K resolution takes approximately 20 minutes. The project is fully open source and free, and agents can also call the local image generation service.

7. Character.ai Founders Have Returned to Google

The AI companion landscape is shifting. A post by Jike user rosicky311, Character.ai Founders Have Returned to Google , notes that c.ai’s original founders have gone back to Google, making the project’s direction largely settled. The post also points out that this space saw a major revenue surge in 2026, with growth coming primarily from several Chinese companies. The competitive center of AI companion apps is moving from the US toward Chinese teams.

8. Li Jigang Illustrates the Transformer QKV Mechanism

A complex concept rendered as a single diagram. This illustration posted by Li Jigang on Jike uses the /ljg-explain command to visualize the QKV attention mechanism in Transformer models. Developers currently teaching themselves the fundamentals of large language models will find this diagram worth saving as a reference.

9. Codex No Longer Shows a Change Summary After Completing Edits

A silent change to the workflow. V2EX user frankxzw reports in Codex No Longer Shows a Change Summary After Completing Edits that OpenAI Codex previously displayed a popup with change information after completing code modifications, but that popup has disappeared. Developers who relied on this summary to confirm changes now need to manually diff the output.

10. AI Cracks Android to Run Steam Games Directly

Head out with a controller and play Steam on the go. A hands-on video shared by 歸藏 on Jike shows that the DroidDeck project enables Android phones to run Steam games directly, with indie titles like Hades running smoothly 🎮. Paired with a tablet that supports DP-in and a small gamepad, you can take it anywhere. Project link: github.com/Droid-Deck/DroidDeck.

⌘ Top Open Source Projects

vercel-labs/skills: Install and Manage Skills for Agents

Skills now have a command-line tool. Vercel’s skills tool supports finding, installing, and managing agent skills via npx skills, written in TypeScript. It topped the GitHub trending chart today, gaining 190 new Stars for a cumulative total of 34,508 Stars. If you’re already using an AI coding assistant, try picking one skill relevant to your current task. And before installing, don’t forget to read its description and script.

◉ Social Media Picks

Designing a Furniture Leg — AI Even Weighs In on How Many Balls to Stack

This time it’s not about designing a whole house — just one leg. Google Gemini retweeted Josh Woodward’s custom furniture leg design demo , showcasing a small use case for Gemini 4 Argon. The furniture leg is 96 millimeters wide, and the caption jokes that the hardest decision was how many balls to stack on it. Compared to sweeping, visually impressive demos, this kind of small part with specific dimensions is more worth opening for anyone doing product prototyping — though keep in mind a video demo is not a complete manufacturing tutorial.

Humans Throw Beanbags, Robots Watch and Build a Training Environment

To teach a robot, start by showing it a video of a person. Google AI shared an experiment in automatically building training simulations from human action videos : Gemini 4 Argon analyzes footage of people playing a beanbag-toss game to help generate a training environment, and a robot then practices in that simulation. The key point is not “watch once and immediately know” — it’s about reducing the manual work of writing simulations. The original post includes a research team description and an actual demo, making it worth a look for anyone working on robot training.

😄 AI Fun

AI Built a Japanese Temple in Blender from a Single Prompt — and Added Bell Sounds the Next Day

Someone gave Gemini 4 Argon a single prompt, and the next morning they opened their computer to find that a multi-agent team had already built an entire Japanese temple complex in Blender — complete with corridors and a pagoda. The full pipeline showing the orchestrator agent coordinating and the critic agent reviewing render results frame by frame against reference images reveals that the system first generated blueprints, then had Nano Banana produce four reference images, with multiple agents dividing up the modeling work and iterating repeatedly. When the demonstrator came back the next day and liked what they saw, they casually added “add bell sounds” — and a procedurally generated temple bell audio track appeared. The “single prompt” framing conceals a complete engineering pipeline; it’s just packaged as one line.

❓ Related Questions

Does Using a Bot to Organize Long Videos in Grok Require Re-explaining the Layout Format Every Time?

You don’t have to repeat it every time — the formatting method just hasn’t been saved as a Skill yet. Today’s skill thread suggests an approach: lock in your preference for “convert to a well-formatted HTML long-form article” as a Skill, then hand it off to GrokBot without having to re-explain your style each time. Try it on a video type you watch regularly, confirm the formatting matches your preferences, and then reuse it. Once your toolset is stable, you can think about building a longer-term routine. If you have an existing account that needs renewal, check out 大爱·AI账号店’s Grok renewal — the renewal covers account access; you’ll still need to configure the layout Skill yourself.