Welcome to the first edition of The Daily Archit. The theme found us: models are getting more capable and more open, and the labs are building tighter boxes to keep them in, so patch your Brotli before you build yours.
Hypothesis author David MacIver's case for tests that generate their own inputs is the answer to the pgrust affair. A fixed set of 43 benchmark queries can be overfit, and agents are very good at overfitting. Generated inputs are much harder to game, which makes them the kind of self-check Dan Luu says ambitious, agent-built software needs.
IncrementGlenn Fleishman's history of how Let's Encrypt made TLS cheap and automatic is good background for the eMudhra mess, in which revoked certificates linger on government sites. Automated renewal and short lifetimes are what make revocation survivable.
Increment