Hey, it’s Andreas.
Every few months, something grabs the entire AI world's attention - not because it's the most powerful, but because nobody saw it coming. This week, that something came with an unexpected mascot: a lobster.🦞

OpenClaw, a personal AI assistant, went viral just weeks after launch — despite having to change its name twice. I spent most of the weekend going deep on it. In this edition: what it actually is, how to set it up fast, and whether it's worth your time or better watched from a distance (this edition will be a little more comprehensive as a result).

Plus, three things from last week you should know about:

• China green lights large-scale Nvidia H200 purchases — implications far beyond chip supply chains.

• Anthropic releases free Claude tutorials — lowering the barrier to one of the most capable models out there.

• Dario Amodei sounds the alarm — his new essay The Adolescence of Technology argues AI poses civilizational-scale risks.

Let's dive in.

Weekly Field Notes

🧰 Industry Updates
New drops: Tools, frameworks & infra for AI agents

🌀 China approves large-scale Nvidia H200 purchases → China reportedly cleared ByteDance, Alibaba, and Tencent to buy 400,000+ NVIDIA H200 chips.

🌀 Google unveils Genie 3 world model → Genie 3 generates real-time, interactive 3D environments from prompts.

🌀 Google integrates Gemini 3 into Chrome → Gemini 3 now runs in a Chrome side panel with auto-browse and task execution. 

🌀 OpenAI launches Prism → Prism is a GPT-5.2-powered LaTeX workspace for research papers.

🌀 NASA on AI-planned autonomy on Mars → NASA revealed that the Perseverance rover completed the first AI-planned drive on another planet, using Anthropic’s Claude to map a 400m route across the Martian surface.

🌀 Microsoft releases Maia 200 inference chips → Maia 200 delivers 10+ PFLOPS per chip, optimized for large-scale inference.

🌀 Anthropic introduces Claude for Excel → Claude can now analyze full workbooks, trace errors, and reason across sheets.

🌀 Moonshot AI releases Kimi K2.5
→ A frontier model with agent swarms and vision capabilities. Coordination-first design shows where multi-agent research is heading.

🎓 Learning & Upskilling
Sharpen your edge - top free courses this week

📘 Anthropic releases free Claude tutorials → Claude now offers a full set of free, official tutorials showing how to use the product properly across work, data, and automation. No paywall, no upsell - just concrete workflows.

🌱 Mind Fuel
Strategic reads, enterprise POVs and research

🔹 OpenAI on building its in-house AI data agent → OpenAI shared how it built a bespoke internal data agent to help 3,500+ employees navigate 600PB of data.

🔹 Anthropic on how coding agents are reshaping development → Anthropic outlined 8 structural shifts showing how agentic coding is changing not just how software is written, but how teams, economics, and risk models evolve.

🔹 Peter Steinberger on running his life with AI in 40 minutes → The creator of OpenClaw shared how he uses OpenClaw to compress daily life admin into ~40 minutes.

🔹 Dario Amodei on AI’s civilizational risk → The Anthropic CEO published The Adolescence of Technology, arguing AI poses systemic risks ranging from bioterrorism and autonomous weapons to mass job displacement and AI-enabled authoritarianism.

🔹 Google releases Gemini Prompting Guide 101 → Google published an updated practical guide covering prompting fundamentals for Gemini.

🔹 Boston Consulting Group on what 640 CEOs reveal about AI → In its AI Radar 2026, BCG surveyed 2,360 executives across 16 markets, including 640 CEOs.

♾️ Thought Loop
What I've been thinking, building, circling this week

Last week, the internet collapsed into one topic. OpenClaw (formerly Clawdbot, then Moltbot) went viral, and people were rushing to buy Mac Minis to get it running.

The repo shot past roughly 140,000 GitHub stars in under a month. That is rare air for any open-source project (I assume this might even be a world record).

And some of the demos that were built are genuinely compelling:

• negotiating and buying cars

• calling restaurants when OpenTable fails

• running your schedule through WhatsApp

Actually, there’s a comprehensive site showcasing what people are building in real-time right now - you can get completely lost in it for hours:

But alongside the excitement, the skepticism is real too: a lot of “wow demos” still translate into very little daily value for most people and security concerns are not abstract when an agent can execute actions.

So today I want to break down:

• What OpenClaw actually is, and what makes it different

• How to set it up in two different ways

• My take on whether it’s worth your time right now

What is OpenClaw?

Clawdbot is a self-hosted, always-on AI assistant that lives where you already work - WhatsApp, Telegram, Slack, Discord, Signal. Unlike e.g. ChatGPT where you go to a website, Clawdbot comes to you where you already are. It behaves more like a background service than a chatbot:

• It remembers context across days

• It can proactively ping you

• It can touch files, browsers, scripts, APIs

• It runs 24/7 without you being “present”

That’s the core reason it feels so different. Most AI assistants today are still sandboxed. They answer questions, summarize documents, and live in browser tabs - isolated from your real digital life. OpenClaw takes a fundamentally different approach. It bridges high-level LLM reasoning with low-level system operations.

Your Telegram or WhatsApp becomes the command interface. Your terminal becomes the execution layer. Your AI doesn’t just talk about helping - it can actually run shell commands, manipulate files, and orchestrate workflows.

How to set up OpenClaw?

First, the non-negotiable: OpenClaw has deep system access - files, browser sessions, accounts, credentials. So don’t run it on the same machine that contains your entire life.

If something gets exploited or you misconfigure one thing, the blast radius is your whole laptop.

There are two setup paths worth considering:

• Local machine (laptop/Mac Mini)
  Use an old laptop/MacBook, or a Mac Mini to get started. Keep it plugged in and disable sleep mode - if the machine sleeps, OpenClaw goes silent.

• VPS (AWS, Hetzner, etc.)
  Pay roughly $5-10/month and run it 24/7 in the cloud. You get reliable uptime and clean separation from your personal device, without leaving a computer on at home.

Option 1: Local on a laptop

This is the way I am running it and the setup is not more than 10 minutes.

What you do:

1. Open Terminal (macOS: Spotlight -> Terminal, Windows: PowerShell, Linux: your terminal)

2. Run the one-liner installer

curl -fsSL https://openclaw.ai/install.sh | bash
exec bash
openclaw onboard --install-daemon

3. Follow the setup wizard (super straight forward)

4. Keep it running while you use it

Option 2: The VPS way

This takes a bit longer (20-30 minutes) than running it locally, but it’s the cleaner setup. A VPS (Virtual Private Server) is basically “a computer in a data center” that you rent. Someone else runs the hardware 24/7 - you just log in and use it. That means OpenClaw stays online without leaving a laptop running at home.

Where to get one:

• VPS: Hetzner, DigitalOcean, Vultr, OVHcloud, Scaleway, UpCloud

• Hyperscalers: AWS, Azure, GCP

• One-click shortcuts: Railway, Render

AWS offers a $200 free tier that can be used to get started. Below, I am showing you how to run it there, but the approach can be applied to all of the provider above.

What you need

• Create a free AWS account

• Your model API keys (OpenAI, Anthropic, etc.)

Step 1: Launch the server (EC2)

Log in to your newly created AWS account and navigate to the AWS Console.

• In the search bar, type EC2.

• Open EC2 -> click Launch instance.

• AMI: Ubuntu LTS (22.04 or 24.04)

• Instance type:

  • basic chat: pick the smallest Free Tier eligible type

  • automation + heavier workflows: you can upgrade that later

• Storage: 20 GB baseline (workspace + logs add up)

Security group (don’t skip this)

• SSH (port 22): allow only your IP (not 0.0.0.0/0)

• Do not open any other ports

• If there’s a web UI, keep it private and use an SSH tunnel

Key pair:

• create/download .pem once

• store it safely

Step 2: SSH in

On your laptop:

chmod 400 ~/Downloads/your-key.pem
ssh -i ~/Downloads/your-key.pem ubuntu@YOUR_SERVER_IP

If you see ubuntu@ip-...:~$, you’re in.

Step 3: Install + configure

Inside the EC2 instance:

curl -fsSL https://openclaw.ai/install.sh | bash
exec bash
openclaw onboard --install-daemon

You’ll typically set:

• model provider + API keys

• your chat channel (WhatsApp/Telegram/Slack/etc.)

• basic permissions and tool access

Step 4: Run it 24/7 (tmux)

Start the gateway and keep it alive after you disconnect:

sudo apt-get update && sudo apt-get install -y tmux
tmux new -s openclaw
openclaw gateway --bind lan --port 18789

Detach (leave it running): Ctrl+b, then d
Reattach later:

tmux attach -t openclaw

Step 5: Access the UI safely (no open ports)

On your laptop, create a tunnel:

ssh -i ~/Downloads/your-key.pem -L 18789:127.0.0.1:18789 ubuntu@YOUR_SERVER_IP

Open:

• http://127.0.0.1:18789

This keeps the UI private. No public exposure. No extra ports.

If you do run into problems, you will find more info to troubleshoot here.

Is OpenClaw worth your time?

OpenClaw is worth your time if you treat it for what it is right now: a powerful beta that exposes the next interface for work. However, it’s important to be aware that there are many people warning of security risks. Above all associated with full device access, including prompt injections, exposed data, and more if not properly managed.

So let me be direct. Do not run this on your primary machine. Don’t connect it to everything on day one. And don’t give it broad permissions because a demo you saw on the internet made it look smooth. When an agent can read your messages, send messages as you, access your files, and execute commands, the blast radius is not theoretical. One careless configuration can turn a fun experiment into a real incident.

However, that does not mean I would avoid the project; you just need to approach it cautiously. The capabilities OpenClaw is bringing are real and impressive. The persistent memory works much better than I would have expect. And the self-improving behavior - the way it can build or extend skills through usage - is the kind of detail that tells you this thing will have a lastening impact. But the hype is also real. A lot of what people call “life-changing” is still just clever orchestration wrapped in a clean UI. That matters, but it’s not magic.

Now the decision. Should you try it?

If you’re technical, you enjoy tinkering, and you understand basic security hygiene, you should try OpenClaw. Not because it’s stable, but because it teaches you the operating model that’s coming. You will learn more about agent behavior, permissions, and failure modes in a day than you will from a month of reading hot takes.

If you’re not technical and you just want a reliable assistant that works with minimal setup, you should wait. I am convinced that the capability OpenClaw demonstrates will show up in mainstream products soon, with better defaults, clearer trust boundaries, and fewer ways to accidentally expose yourself. I except a lot of happening over the next 2-3 months.

P.S. If you try it OpenClaw, one thing made a big difference for me: don’t treat OpenClaw like a chatbot. On day one, I kept writing full sentences and over-explaining context. That’s the wrong mode. It works best when you give it tasks the way you’d text a coworker. For example: “Scan my inbox, prioritize anything marked ‘urgent’, and send me a 5-bullet summary at 8:30am.” That’s it. The less you narrate, the more it can actually operate. Ironically, that’s where it shines.
🦞