Anthropic Begs the AI Industry to Slow Down

That's the twist. The company behind Claude put out a post arguing the world should be able to slow or pause frontier AI, so safety research and the rest of society can catch up. Their worry is that models are inching toward improving themselves with no human in the loop.

Highly doubtful. Anthropic argues a pause only works if the big players in the US and China all agree at once, with rules everyone can verify, since whoever stops first just hands the lead to everyone else. Now they want governments and rival labs in a room to figure out how that could even work.

Hermes is an open-source agent from Nous Research that gets better the more you use it, saving each new trick as a reusable skill it can pull up later. The catch was always that you had to run the whole thing from a terminal, but not anymore.

It finally has a real deskptop app for Mac, Windows, and Linux, with no command line in sight. It is free and MIT licensed, with one memory that follows you across Telegram, Slack, email, and a dozen other places.

💻 Google's Gemma 4 runs on your laptop

📚 Wikipedia editors threaten to walk off the job

🧠 Microsoft launches seven of its own AI models

🛡️ An AI agent pulled off a hack alone

🔮 GPT 5.6 leaks early, and it's all about efficiency

🧰 Mistral folds Le Chat into one work agent

Hermes Desktop just dropped, so this week you get your own agent up and running, no terminal required. Warning: API costs can get expensive, so start with a low cost model like DeepSeek V4 or Gemini Flash.

Would you hand a self-improving agent the keys to your desktop? And is the industry ever going to actually slow down, or just race faster?