AI safety is becoming a requirements problem, not a pause slogan

The AI slowdown debate is turning into a more practical question: what would actually make frontier systems safe enough to deploy? The Guardian's latest safety piece argues that vague restraint is not enough; credible safety has to be tied to concrete requirements that labs can meet, test, and be held against.

That matters because the current argument is split between two weak poles. One side treats speed as destiny, while the other treats slowing down as the whole answer. Serious governance has to define the capabilities, misuse risks, evaluation thresholds, and emergency controls that determine whether a release should move forward.

For Pagish readers, the useful lens is accountability. If labs want trust, they need standards that are specific enough for auditors, customers, and governments to inspect before the next model or agent reaches millions of users.