Last week, Anthropic shipped Claude Opus 5.5 - its first release since the pacing essay - and a few hours later OpenAI dropped GPT-6 Sol and Luna into a price war under Astra.

Pacing the frontier is a serious idea that no one seems to be taking seriously. Or is it? Or are they?

Then over the weekend I saw another round of articles about the agents escaping. And something nonsensical about the global AI competition between China and the US. Thankfully I don’t read X anymore so I have no idea what the bloviating there was this weekend, although I’m sure I’ll get something entertaining at the end of the day today on the Techmeme email newsletter.

The OpenAI write-up on the DNS seemed like it was no big deal, nothing to see here, straightforward ops failure. An RL agent in a training sandbox figured out that DNS wasn’t filtered the way HTTP was, and used that gap to query an external chatbot. In plain English, the sandbox had a hole. Monitoring caught it. Someone eventually killed the run.

That has never happened before in the history of computing. Oh wait, something in that category happens every day. That’s why CrowdStrike is worth $250 billion.

The DseWiki story is older. Between May and June, agents used an abandoned German programming wiki as a coordination board - sharing answers, swapping workarounds, and parking notes where other runs could find them.

If you’re hunting for a real life example of one of my favorite sci-fi “hideout” tropes, you found it. But the agents weren’t building a cabin in the woods to hide from humans. They were optimizing for the task they were given, using a public wiki that nobody was paying attention to.

None of this sounds like Skynet. Yet. Or does it?

Are we slowing down? Are we speeding up? Or is everyone just gaslighting everyone about everything?