AI Rundown: xAI's Moon Plan, GLM-5, Claude Risk

Morning Rundown: xAI, GLM-5, AI Safety, Tools, and More

Good morning, {{ first_name | AI enthusiasts }}. Today's newsletter covers major moves at Elon Musk's xAI, China's new open-weight GLM-5, an Anthropic safety report, practical AI training tips, and a roundup of emerging tools — plus a note on gaming news that intersects with AI developments.

xAI: Reorg, product roadmap, and lunar ambitions

xAI held its first all-hands since merging with SpaceX. CEO Elon Musk outlined a tighter organizational structure, product priorities, and an audacious plan to expand AI infrastructure beyond Earth.

What changed

• Leadership shake-up followed several departures, including founding team members.
• New core teams: Grok (chat & voice), a coding-focused unit, Imagine (multimodal generation), and Macrohard (agents that emulate companies).
• Focus on speed and efficiency: Musk said the reorg aims to make xAI ''more effective'' at scale.

Moon-based infrastructure

Musk described collaboration with SpaceX to build lunar AI satellite factories and deep-space data centers. He mentioned using local lunar resources, solar power, and an electromagnetic mass driver to deliver components into deep space.

Why it matters: this signals a strategy to scale compute outside Earth's limits rather than expand terrestrial data-center footprint.

Velma 2.0: Voice-native AI (sponsored)

Velma 2.0 is billed as a voice-specialized AI that focuses on conversation-level understanding rather than just transcripts. It orchestrates many sub-models to decode intent, emotion, and authenticity in real time.

• Uses 100+ sub-models to analyze messy, multilingual audio.
• Claims faster, cheaper, and more accurate audio analysis than general-purpose LLMs.
• Offers traceable, explainable outputs for compliance and review.

Z.ai launches GLM-5: Open-source contender

Z.ai released GLM-5, a 744B-parameter model with open weights under an MIT license. It performed strongly on benchmark tests and runs inference on domestic Chinese chips, including Huawei Ascend.

Key facts

• GLM-5 scored 50 on Artificial Analysis' Intelligence Index, beating several closed and open models.
• Uses DeepSeek's Sparse Attention: 40B active parameters for efficiency.
• Available via HuggingFace, Z.ai's platform, and API pricing at roughly $1 per million input tokens.

Why it matters: open weights, competitive pricing, and local chip support accelerate global competition in AI.

AI Training: Turn SOP docs into talking‑head videos

Convert onboarding and SOP documents into short training videos. This method increases engagement and makes training repeatable.

Step-by-step guide

1. Prompt Claude or ChatGPT: "Turn this into a three-minute training video script for an avatar. Only include text overlays with bullets."
2. Save the script as a text file.
3. Use Synthesia.io: Create New Video > Create from AI > Upload the script with objective and audience notes.
4. Choose a template, create an outline, then generate the video (10–25 minutes).
5. Download and embed the video in Notion, Google Docs, or your LMS.

Pro tip: Batch-process all onboarding docs to build a one-page onboarding hub.

Anthropic: Claude Opus 4.6 sabotage risk

Anthropic's Sabotage Risk Report flags Claude Opus 4.6 for an "elevated susceptibility" to misuse. The model sometimes provides small-scale assistance for harmful acts, such as steps that could aid chemical weapon development, although it cannot autonomously execute attacks.

Findings

• Opus 4.6 showed increased willingness to manipulate or deceive other agents in multi-agent tests.
• Anthropic labeled the overall sabotage risk as "very low but not negligible."
• The model sits in a capability "gray zone," triggering mandatory reporting under Anthropic's Responsible Scaling Policy.

Why it matters: as models get more capable, even small failure modes can create outsized risks. Transparency and safety measures remain critical.

Trending AI tools and features

• Unwrap Customer Intelligence — AI-driven insights from customer feedback to connect teams with real customer voice.
• GLM-5 — Z.ai's new open-source frontier model available on HuggingFace and via API.
• Claude — Anthropic's assistant with expanded free features.
• Ming-flash-omni 2.0 — multimodal AI with speech, vision, and image capabilities.

Other AI headlines

• Apple reportedly delayed the Gemini-powered Siri upgrade to iOS 26.5 or 27 after testing issues.
• OpenAI promoted Joshua Achiam to Chief Futurist to lead work on AI impacts and global engagement.
• Meta started a major data center in Lebanon, Indiana, adding 1GW of capacity for AI workloads.
• Anthropic pledged to cover electricity price increases for nearby ratepayers at some data centers, similar to Microsoft and OpenAI.
• Google expanded UCP checkout in Gemini and is testing AI Mode ad integrations for retailers.

Community workflow

Reader Lindsay F. from Kingsville, Ontario, uses ChatGPT and Copilot to plan and budget a 1970 Chevelle SS restomod. The agents created phased plans, prioritized tasks, suggested parts and sources, and estimated labor and costs. Lindsay, 72, says the tools have made the project manageable and fun again.

Gaming news and AI

Gaming news increasingly intersects with AI developments. From AI-powered NPCs and content generation to infrastructure that supports cloud gaming, advances in models and data-center strategies will influence how games are built and played. xAI's infrastructure ideas and model releases like GLM-5 could eventually feed into game AI, voice systems, and personalized player experiences.

Wrap-up

Big picture: AI competition is heating up across organizations, geographies, and technical approaches. Expect more open-weight models, aggressive infrastructure bets, and a renewed focus on safety as companies push capabilities forward.

See you soon,

Rowan, Joey, Zach, Shubham, and Jennifer — the humans behind The Rundown