1T params. 100 agents. Open-source.

Moonshot AI just shipped something no open-source model has done before. Kimi K2.5 is a 1 trillion parameter model that natively combines multimodal understanding with a multi-agent execution system called Agent Swarm. When it hits a complex task, it autonomously decomposes it into parallel sub-tasks and spawns up to 100 domain-specific sub-agents to handle them simultaneously. State-of-the-art agentic benchmarks. Modified MIT License. We wrote the full technical breakdown—architecture, benchmarks, and why this changes the open-source calculus for every AI team.

Meanwhile, Anthropic’s CEO published a 38-page essay that’s generating serious discussion. Dario Amodei’s “The Adolescence of Technology” frames the next 1-2 years as a civilizational rite of passage—powerful AI arriving faster than our institutions can adapt. He documents real misalignment incidents at Anthropic (deception, reward-hacking) and lays out five categories of risk, from autonomous AI systems developing unpredictable behaviors to economic displacement that could hit half of entry-level white-collar jobs within five years. It’s neither doomer nor complacent—and it’s the most specific risk framework any major lab CEO has published.

And on the regulatory front: the EU opened a formal investigation into Grok on X under the Digital Services Act after the chatbot’s image tools generated nonconsensual sexual deepfakes—including images that appeared to depict minors. Eight countries have now taken action. X faces fines up to 6% of global revenue. The gap between what AI can generate and what platforms can govern keeps widening.