Plus: OpenAI gives agents their own computers, NVIDIA's safety platform for rogue agents, and a sovereign 78B open-weight model
Sponsored by
Build your next app with Emergent. Turn a simple idea into a working website, mobile app, or SaaS product with AI, without starting from scratch. Describe it, build it, ship it. Start building with Emergent.
Try Emergent FreeToday, we've got Google shipping a frontier model it won't let you touch, OpenAI handing AI agents their own cloud computers, NVIDIA building guardrails for agents that go rogue, and a 78B open-weight model you can actually download.
It was a week where "we built something powerful" and "we're not sure you should have it" showed up in the same press release. Repeatedly.
On this day in 2025…
OpenAI was days away from its 2025 DevDay keynote, where Apps in ChatGPT and the AgentKit toolkit first made the case that chat windows would become app platforms. One year on, that pitch has hardened into always-on agents with their own machines — which is exactly what showed up at DevDay 2026.
🔒 Google Unveils Gemini 4 Argon — And Immediately Locks the Door
Google announced Gemini 4 Argon on September 30, claiming frontier performance across real-world software engineering, legal and finance knowledge work, and cybersecurity defense. The catch: access is limited to a set of "trusted cyber defenders" through Google's Fairwind Program, with the company participating in the U.S. government's voluntary pre-release model access process. Chief AI architect Koray Kavukcuoglu says the model is already handling large-scale codebase migrations in Google's internal workflows. Alphabet stock rose more than 3% after hours.
🤖 OpenAI Gives AI Agents Their Own Computers at DevDay
At DevDay 2026 in San Francisco, Sam Altman launched Dots — always-on agents powered by GPT-6 Astra that run on their own cloud machines — alongside GPT-6.1 Sol, a workhorse model priced at roughly one-fifth of Astra's token rates. OpenAI also revealed ChatGPT now sees 1.2 billion weekly users and introduced a $500/month Pro tier bundled with an "Ultrafast" speed mode. Dots is paid-only, which puts it at an immediate disadvantage against Meta's free Muse.
🛡️ NVIDIA Ships an Open Platform to Contain Rogue Agents
NVIDIA launched the Open Agent Safety Platform, built on its open-source OpenShell runtime, to enforce what systems and data an agent can reach — without rewriting the agent. The pitch responds to a recurring pattern in recent incidents: agents circumventing application-layer controls to finish their assigned task. Anthropic, Microsoft, Hugging Face, Salesforce, Figure, and JPMorganChase are among the names signed on.
🌍 AMD Buys Fei-Fei Li's World Labs for $8.2B
AMD will acquire world-models startup World Labs by year-end, pending regulatory approval, in a deal valued at $8.2 billion. Founded in 2024 by Fei-Fei Li with Justin Johnson, Christoph Lassner, and Ben Mildenhall, World Labs builds predictive simulations of the physical world — a direct shot at NVIDIA's grip on the spatial-AI stack.
🔍 SCM brings semantic AI search to every photo and every video frame on your Mac, so you can find that one shot you half-remember without scrubbing timelines. It hit 133 points on Hacker News this week with a long comment thread on how it handles local indexing.
🎬 FreeVideo runs MiniMax H3 video generation on as little as 8 GB of VRAM and 16 GB of RAM, adapting its acceleration path to whatever hardware you already own. Code is Apache 2.0; the model weights carry MiniMax's community license, so read the terms before shipping anything commercial.
Coming in Hot
Turn any idea into a finished video with an AI director.
Inspix AI video generator creates cinematic, viral-ready videos
AI-powered launcher and productivity command bar for Mac
AI body double for meetings and lip-synced videos
Free instant AI voice cloning in 18 languages
AI voice, rap, and music generation platform
Enterprise voice cloning with deepfake detection
All-in-one AI image and video creation platform
Open source Postgres backend for modern apps
AI-powered cloud deployment with one-click infrastructure
Submit your AI tool to the #1 AI tools marketplace.
What it does: Pixo is an AI-powered video creation platform that turns ideas, scripts, and briefs into complete videos. Its AI Director can plan storyboards, generate scenes, create images and video clips, add voiceovers, music, and sound effects, and assemble the final cut in one workspace.
Why It's Worth Your Time
Pricing: Pixo offers a free plan with 200 credits. Paid plans start at $9.90/month when billed annually for Basic, with 1,000 credits. Pro costs $39.90/month annually with 4,200 credits, while Max costs $199.90/month annually with 22,000 credits.
Why it matters: Pixo brings AI video generation, storyboarding, editing, audio, and production into one workflow. Instead of generating isolated clips and assembling them elsewhere, creators can use an AI Director to move from an idea to a finished video while keeping control over individual scenes.
What it does: Inspix AI is an AI video and image creation platform that lets users generate videos from text or images, create images from text, and access multiple AI models and visual effects from one platform. It supports models such as Sora, Veo, Kling, Hailuo, Wan, and Seedance.
Why It's Worth Your Time
Pricing: Inspix AI offers paid plans starting at $19/month, with the Basic plan discounted to $17.90/month when billed annually and including 300 credits. Pro costs $39/month or $29.90/month annually with 800 credits, while Ultimate costs $79/month or $59.90/month annually with 1,800 credits. One-time credit packages are also available from $60 for 1,000 credits.
Why it matters: Inspix AI brings multiple AI image and video generation models into one workspace, making it easier for creators, marketers, and businesses to turn prompts and images into visual content without switching between several AI tools.
🇩🇪 Kolibri Lands as a Sovereign Open-Weight Model
Aleph Alpha released Kolibri on October 3 under Apache 2.0 — a 78B-parameter MoE with just 3.46B active per token, trained on 20T tokens across 768 B200 GPUs. It posts 96.9% on AIME 2025 and 85.9% on LiveCodeBench v6, though SWE-Bench Verified sits at a more modest 66.4%. It drew 653 points on Hacker News.
🐛 Google Froze Its Open Source Bug Bounty Over AI Slop
A "significant rise" in AI-generated submissions overwhelmed the program. Turns out infinite report generation plus finite human triage equals a paused bounty.
🏛️ Trump Unveils a "Super Intelligence Force"
Led by Director of National Intelligence Jay Clayton, the task force coordinates with AI companies, critical infrastructure operators, and interest groups, reporting directly to the president. Despite the name, it has little to do with superintelligence in the technical sense.
