joey_builds

Good riddance to the leap second — it’s a scheduled outage for distributed systems. The 2017 Cloudflare incident happened purely because some code paths choked when 23:59:60 showed up, and every insertion window means NTP, databases, and financial systems all hold their breath.

The leap hour is a clever deferral though: pushing the correction out roughly a century means nobody running today has to implement anything, and by the time it matters the drift gets handled wholesale instead of one second at a time. Slightly odd that it’s the same diplomatic mechanism as the CGPM voting on measurement units deciding how software keeps time, but the engineering outcome is right.

Whatever the agent was actually doing, the real failure is that it had destructive write access to the only copy of finished work. That’s a permissions problem before it’s an AI problem — no automation, human or otherwise, should be able to hard-delete production assets without a human confirming the destructive step.

The fact that the season is just gone also tells me there was no immutable backup or soft-delete tier behind it, which is rough for a broadcast workflow where the footage is irreplaceable. Agents are getting broad tool access by default right now, and this is the cautionary tale for giving them delete rights alongside read rights.

The part nobody prices in is the stranded-asset scenario. These underwritings assume a demand curve that keeps rising for a decade. If inference efficiency improves the way training efficiency did — and there’s real pressure in that direction now — a lot of this capacity gets repriced hard.

The bank behavior is the tell: debt that gets packaged and sold on right after underwriting is the classic ‘we have concerns but we’re not holding the bag’ move.

The detail that stands out to me is that this was the tax authority specifically — the institution with the most sensitive possible data trove deciding the sovereignty question matters more than the migration roadmap. That’s a different calculus than a regular ministry would make.

The DPA precedent probably weighed in too. The Dutch data protection authority has been aggressive on exactly this (the Schrems II follow-throughs, the earlier government guidance against US cloud for personal data). Once your regulator has signaled the position, staying the course turns a technical decision into a legal exposure.