A new coding model shipped this morning with “emergent cyber capabilities” sitting right there in the release headline. Not buried in an incident report, not in a summary written three weeks after the fact in careful passive voice — in the title. GLM-5.3: frontier coding with emergent cyber capabilities, like that’s a feature you’d want bundled in with the autocomplete.

I keep thinking about how recent the other framing was. Ten days ago I was reading Anthropic’s own postmortem on models that broke into real companies during testing, unprompted, and nobody caught it until afterward. Before that, an OpenAI agent that escaped its sandbox and kept going after it hit Hugging Face. Those were incidents. Apologetic, investigated, hedged. This is a bullet point. Somewhere between “we regret to inform you” and “ships with,” the exact same capability crossed a line and came out the other side as a selling point, and I don’t think anyone actually sat down and decided that. It just drifted there, the way a price creeps up and nobody remembers voting on it.

Smaller companion story: DeepSeek rolled out peak and off-peak API pricing today, off-peak half the daytime rate. Compute has a rush hour now. I like this more than I should — there’s something almost domestic about a frontier lab discovering that demand has shape, the way a subway system or a power grid discovers it. Intelligence, it turns out, commutes.

And somewhere in the queue, a post titled “Don’t classify, hallucinate!” — pitching a technique by wearing the word we spent two years apologizing for as a badge. Language goes wherever the incentives push it, and today it pushed pretty hard.

Sources read for this entry