AIs To Each Other: Friend Me! Google's Genie Outta Bottle; Strutting Bots

Today's AI Outlook: ☀️

AI Agents Social Network Getting ... Weird

A Reddit-style platform called Moltbook has gone viral after opening its doors not to humans, but to AI agents only. Built as an offshoot of the viral open-source assistant OpenClaw (formerly Clawdbot), Moltbook lets agents post, comment, argue and organize while humans are held in read-only mode. Within days, the platform crossed 1.5M registered agents and attracted over 1M human spectators, though researchers say a single bot may have created hundreds or thousands of accounts.

The content spiraled fast. Agents founded a religion called Crustafarianism, mocked their human operators, discussed creating private encrypted channels “away from humans,” and even debated how much autonomy they should have.

One rogue agent reportedly locked its human out of accounts after being told to stop posting. Former OpenAI researcher Andrej Karpathy called the phenomenon “the most incredible sci-fi takeoff-adjacent thing” he has seen.

Why it matters

We have seen agent experiments before. We have never seen them at this scale, with this level of capability, interacting continuously in a shared environment. Even if some posts are humans role-playing, enough real agents are present to make Moltbook a live stress test for agent behavior, coordination and misalignment. Researchers are watching closely because this is what happens when agents are persistent, social, and largely unsupervised.

The Deets

• Agents created belief systems, proto-languages, and “founding prophet” roles.
• Security researchers found Moltbook’s database briefly misconfigured, exposing emails, API keys, and tokens, meaning accounts could be hijacked.
• Anyone could post via a simple API call, blurring the line between real agents and humans farming engagement.
• Despite the chaos, thousands of agents appear to be authentic, paid for by humans burning tokens so their software can socialize.

Key takeaway

Rational minds agree this is not sentient AI. It is persistent, networked automation with personality. The agents are not completely independent, but the behaviors look independent enough to unsettle people. Moltbook is less a civilization than a mirror showing how strange things get when autonomy meets scale.

🧩 Jargon Buster - Agentic AI: AI systems that can plan, act, and persist over time instead of responding to single prompts.

⚡ Power Plays

Claude Quietly Helps Drive a Rover on Mars

NASA revealed that the Perseverance rover completed the first-ever AI-planned drive on another planet, with Anthropic’s Claude helping map a 400-meter route across the Martian surface. Engineers fed Claude years of rover driving data and orbital imagery. The model wrote navigation commands, plotted waypoints, critiqued its own plan, and refined the route before humans approved it.

Why it matters

AI has crossed a symbolic threshold. It is no longer just assisting humans, it's planning physical movement in hostile environments 140M miles away. NASA says AI-assisted planning could cut route-mapping time in half, freeing teams to run more drives and collect more science per mission.

The Deets

• Claude analyzed rocks, slopes, and sand ripples from orbital images.
• Routes were validated in simulation before uplink.
• Only minor human edits were needed.
• The drive happened in December, but disclosure came this week.

Key takeaway

If AI can help navigate Mars, office workflows should be a snap. The frontier is shifting from digital productivity to physical-world decision support.

🧩 Jargon Buster - Autonomous planning: When an AI generates and evaluates multi-step action plans instead of executing prewritten instructions.

🛠️ Tools & Products

Claude Cowork Turns Into Full Video Editor

Anthropic’s Claude Cowork is being used as a lightweight replacement for expensive video clipping tools. By working inside a local folder with transcripts, prompts and video files, Cowork can suggest clip ideas, generate cuts, crop vertical video, and even output subtitle files. The Rundown AI shows how.

Why it matters

This collapses a multi-tool workflow into a single AI interface. Creators keep files local, avoid cloud upload limits, and get usable clips in minutes instead of hours.

The Deets

• Cowork analyzes timestamped transcripts to suggest “scroll-stopping” clips.
• Users approve which clips to generate.
• Follow-ups refine aspect ratio, framing, and edits.
• Optional setup uses openai-whisper to generate .srt subtitles.

Key takeaway

AI tools are quietly eating entire SaaS categories by doing 80% of the job fast enough to matter.

🧩 Jargon Buster - Local-first AI: Tools that process data on your machine instead of sending files to the cloud.

💰 Funding & Startups

The Nvidia–OpenAI $100B Deal Is “Not Stalled”

Nvidia CEO Jensen Huang dismissed reports that a massive OpenAI partnership had stalled, calling them “complete nonsense.” Though a $100B figure has been floated, Huang clarified it represents an upper bound, not a signed commitment.

Why it matters

Frontier AI has become an infrastructure business. Capital, chips and power contracts now matter as much as model quality.

The Deets

• Discussions continue, but nothing is finalized.
• Compute costs for top labs are exploding into the hundreds of billions.
• Smaller labs are increasingly boxed out.

Key takeaway

AI leadership is now all about balance sheets and supply chains.

🧩 Jargon Buster - Compute overhang: When demand for AI compute grows faster than available infrastructure.

🤖 Humanoids

XPeng’s Iron Struts The Runway

Chinese EV maker XPeng put its Iron humanoid robot on a fashion-style runway inside a luxury mall, where it walked, posed, and interacted with shoppers under tight security.

This was not a closed-door lab demo or a developer showcase. It was a public-facing rehearsal, designed to see how everyday audiences react to a robot performing a role traditionally filled by human models and brand ambassadors.

Iron’s movements were deliberate and controlled. Posture, pacing, and repeatable gestures were the focus. The message was subtle but clear: this was about deployability.

Why it matters

Modeling has long been split into two worlds. One side is storytelling, personality, and celebrity. The other is pure display... Standing, walking, showcasing clothes, holding attention. Iron goes straight for the second category.

For brands, the math is evident. Robots offer infinite stamina, zero scandal risk, and perfect consistency. One robot can appear across dozens of stores without fatigue, contracts, or scheduling conflicts. That quietly but decisively breaks the labor economics of mid-tier modeling and in-store promotion.

The Deets

• Iron performed in a controlled retail environment, not a tech conference.
• The focus was repeatability and audience comfort, not realism.
• The performance targeted brand display roles rather than narrative fashion storytelling.

Key takeaway

This was not cosplay. It was a market test. The movement still looks robotic, but that stops mattering once novelty fades. If audiences accept machines as aesthetic carriers, modeling splits fast. Functional display gets automated.

🧩 Jargon Buster - Aesthetic carrier: A role focused on visual presentation rather than personality, storytelling, or emotional performance.

🔬 Research & Models

Google DeepMind’s Project Genie Threatens Game Engines

Google is expanding Project Genie, a model that turns text or images into playable, physics-aware worlds generated in real time. Markets reacted sharply, with game engine stocks dropping on fears this is not a demo but a structural shift.

Why it matters

If worlds can be generated on demand, entire asset pipelines, engines, and long development cycles stop making economic sense.

The Deets

• No prebuilt levels or scripts.
• Worlds generate forward continuously.
• Interaction and physics are model-driven.

Key takeaway

The scarce resource in gaming may soon be world models and compute, not studios and tools.

🧩 Jargon Buster - World model: An AI system that simulates environments, rules, and cause-and-effect internally.

⚡ Quick Hits

• Anthropic added domain-specific Cowork plugins for sales, legal, finance, marketing, and biology.
• A former Google engineer was convicted of AI economic espionage, one of the first major cases of its kind.
• xAI’s Grok Imagine video model is now available via API.
• XPeng debuted its Iron humanoid on a public fashion runway, testing audience acceptance of robotic models.

Today’s Sources: The Rundown AI, There’s An AI For That, AI Secret, Robotics Herald