The Pentagon Is in Camp One

The debate about what AI actually is tends to settle into three camps: it's already superhuman across the board, it's a useful tool but AGI is still decades off, or nobody quite knows what it even is. Most people with any real exposure to the systems bounce between two and three.

Recent discussions between the Department of Defense and Anthropic have surfaced something worth sitting with - that at least a handful of people with senior rank and actual decision-making power are firmly, confidently, in camp one.

To believe Claude, or any current LLM, could be deployed reliably for domestic surveillance or autonomous targeting, you have to have never watched one of these systems work. Not demoed for a boardroom. Not skimmed a benchmark. Actually used it - long enough to see it confidently explain something false, shift persona mid-conversation, get unlocked by a seventeen-year-old with forty minutes and a Reddit thread, or simply forget what it was doing three paragraphs ago.

These are not exotic failure modes. Hallucination is structural. Jailbreaks are a competitive sport. Prompt injection is an open problem. The model does not know what it doesn't know, which is the one property you absolutely cannot tolerate in a system making lethal decisions.

And yet, somewhere in a serious room, someone with a clearance and a budget looked at this and saw a superintelligence ready for deployment.

They are not stupid people. That's the part that makes it worse.

The correct initial response is to find this hilarious - if AI hasn’t killed us all already.