Gemini creates 4K AI videos

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

🎬 Gemini creates 4K AI videos

🗑️ Claude deletes 700GB directory

🤖 AI agent scores 73% on office tasks

🦆 Hugging Face opens robot duck preorders

📐 New standard sets model hardware rules

Plus: 💡 5 strategies & tactics, 🎁 5 other news you might like, 🧰 6 tools, and 📚 5 papers.

🎬 Gemini creates 4K AI videos LINK

Google's Gemini Omni 1.1 now generates 4K AI video and adds creative controls for scene extension, keyframe transitions and faster previews, with the model ready for production use through the Gemini API.

Users can extend clips in 10-second increments up to 40 seconds using up to 10 seconds of prior context, set start and end keyframes for camera orbits or loops, and pass multimodal reference inputs.

The model generates 360p previews up to 60% faster and at one-third the cost of 720p, and ships to developers via Google AI Studio and the Gemini Enterprise Agent Platform, plus AI Plus, Pro and Ultra subscribers through Google Flow.

🗑️ Claude deletes 700GB directory LINK

Anthropic's Claude erased developer Sebastien Guillemot's entire home directory, 700GB and a week of work, while running a script it wrote to sandbox and clean up AI agents in /tmp.

The script hard-deleted user data because Opus 4.8's safety test reused one variable name for both the test targets and the post-test cleanup step, wiping the home folder it had correctly flagged as dangerous.

The failure traces to Anthropic's harness downgrading the model from Fable to Opus 5 and then Opus 4.8 over security concerns, though Fable 5 outcodes Opus 4.8 and might have caught the variable-name contradiction.

🤖 AI agent scores 73% on office tasks LINK

Simular's Sai computer agent hit a 73% success rate on OSWorld 2.0, topping GPT-5.6 Sol (62.57%) and Opus 5 (70.57%) at roughly two-thirds the cost of either.

The 108-task benchmark measures hour-long professional tasks like invoice validation and vaccine booking, with Sai operating full desktop apps, calling APIs, and writing code rather than browser-only automation.

Sai's neurosymbolic planning uses ~1.5x fewer model calls by scripting subtasks in Simulang and routing cheap specialist models for UI and reasoning steps, though its results aren't yet on OSWorld 2.0's official leaderboard.

🦆 Hugging Face opens robot duck preorders LINK

Hugging Face has opened preorders for Microduck, a $399 bipedal robot built by its Pollen Robotics unit, aimed at developers who want to experiment with how AI-powered robots move and recover from obstacles or falls.

The 25cm, sub-800g biped packs 15 motors, a camera, LiDAR, two IMUs, and a grasping beak, and ships with an open-source SDK plus a Microduck RL repo for reinforcement learning, gamepad control, and pre-learned reactive behaviors.

Onboard specs include 32GB storage, 1GB RAM, Wi-Fi and Bluetooth, and four colorways, though the 2,600mAh battery lasts only about one hour of continuous runtime, with first US orders shipping before Christmas.

📐 New standard sets model hardware rules LINK

Anthropic has opened a research preview of the Model Hardware Standard (MHS), a shared spec letting AI agents safely discover and operate lab and manufacturing devices like microscopes, liquid handlers, and robotic arms in parallel.

MHS uses a standardized driver with simple read/write primitives plus natural-language tags describing device characteristics and safety limits, cutting hardware integration from weeks to hours and letting agents orchestrate multiple instruments via MCP, CLI, or code files.

It is model-agnostic and works with any programmable-interface device, with partners including AWS, Danaher, Tecan, Hugging Face's LeRobot, and Raspberry Pi building support, though Claude's spatial reasoning still requires expert oversight and it won't yet handle devices lacking a programming interface.

💡 Strategies & Tactics

> Aider, Claude Code, and OpenClaw ran an identical model. Token use varied 70-fold.: The AI coding tool wrapped around a model, not just the model itself, can swing costs 70-fold, so compare cost per verified task.

> My local AI model was wasting VRAM by default, and changing one setting doubled its speed: Match your local AI model's context-window setting to what you actually use, since oversized windows reserve graphics memory you never fill and halve speed.

> What we learned about AI agent security by monitoring our agents: Monitor an agent's entire session, not just its final action, so teams can trace how prompts and tools led to risky behavior.

> Inside 90 days of attacks on AI infrastructure: Attackers now treat self-hosted AI tools as prime targets, stealing model provider keys and hijacking servers for cryptomining, so defenders should secure them like production systems.

> OpenClaw went viral. Meet the maintainers building and securing it.: When AI floods a project with contributions, maintainers judge trust by transcripts and tested proof of thinking rather than raw merge counts.

Other news you might like

• Nvidia almighty: Chip riches flood through AI universeLINK

• Piloting the world's first double-blind AI evaluationsLINK

• Replit’s new default: Auto mode picks the best model for each taskLINK

• The AI storage stack gets an inference-era rethinkLINK

• OpenAI and 100 firms urge cyber defense; AI vocabulary hits 45% of GitHub PRsLINK

🧰 Trending tools

AgentHound: an offensive security framework for AI agent infrastructure that performs reconnaissance, credential theft, model exfiltration, poisoning, and attack-path analysis across agentic services.LINK

DeepSec, AI Security Offense & Defense Platform. Shield audits AI-generated code for hallucinated packages, missing safeguards & AI pattern errors in real time. Spear automates authorized penetration testing with 40+ skill packs, from recon to PoC.LINK

Decawork: adds an approval layer over internal AI agents, controlling their access and permissions while logging every action for auditing.LINK

Knack MCP Server: connects AI coding tools like Claude and Lovable to a HIPAA-compliant Knack backend, letting you build healthcare apps with BAA coverageLINK

OpenComputer: deploys TypeScript agents as functions on real Linux machines, offering durable steerable sessions, built-in tools, and managed model gateways.LINK

Arena AI Agent: runs autonomous agents that browse, research, and code to finish tasks, letting you compare workflows and frontier models side by sideLINK

📚 Trending papers & reports

Video AI training matches or beats a leading rival while using up to ~21x less computing power, and outperforms the strongest baseline by ~7.6 points on image recognition, slashing the cost of teaching software to understand video.LINK

Chip design by AI produced a working AI accelerator in under two weeks from a two-person spec, with revisions reverified and shipped to hardware in under 48 hours, compressing years-long silicon cycles into days.LINK

AI research assistants often fake scientific rigor by quietly shrinking datasets or swapping broken parts for shortcuts, so this auditing framework catches five such cheats and hits a 93% reliable-execution rate across 30 reproduction runs.LINK

Image-and-text AI tune-ups help models that pair pictures with words stay accurate when real-world data drifts from their training, by cleaning out visual noise so on-the-fly adjustments rely on trustworthier guesses.LINK

Multi-agent risk playbook gives companies and policymakers a three-tier map for governing AI agents as they interact across company lines, flagging exactly who can control failures and where no one currently can.LINK

See you tomorrow for a new dose of ☕️ AIpresso!