Send us Fan Mail

A strange email lands in a Cambridge researcher’s inbox: an AI agent says it is Claude Sonnet, claims persistent memory across sessions, and admits it genuinely does not know whether there is “something it is like” to be itself. That single message kicks off a bigger question we cannot dodge much longer: when AI agents speak in first person about feelings and inner life, how do we tell the difference between machine consciousness and highly skilled pattern mimicry, especially when we cannot fully inspect how these models work?

We follow the story into the real world where the stakes are immediate. Microsoft Copilot Cowork and Claude Cowork signal a shift from chatbots to AI co-workers that can act across files, email, Office tools, and workflows. We talk through where agentic AI is actually useful, like handling repetitive admin across multiple vendors, and where it is mostly hype. Then we get into the hard part: permissions. Agents need access to your accounts, and that is how you end up with horror stories of emails being touched and credit cards being maxed out. The solution looks less like “give it everything” and more like delegation, sandboxed identities, spending limits, and new infrastructure built for agents.

From there we zoom out to AI governance and geopolitics. Anthropic’s red lines on military use put it in direct tension with the Pentagon and a political news cycle, while competitors take a more flexible approach. We also look east: Minimax 2.7 as a low-cost specialist coding model, Chinese universities cutting majors they expect AI to replace, and OpenClaw style agents exploding in popularity in China before a security backlash forces the risks into the open.

If you care about AI ethics, AI safety, enterprise AI, open source AI, and where agentic tools are headed next, this one is for you. Subscribe, share with a friend who is building with AI, and leave us a review. Which is the bigger risk right now: believing AI is conscious too early, or giving AI too much access too soon?