No slop · no ads · daily

The stack

4 picks · curated
01
Hacker News txt 940 c · 1441 pts

Gemini 4 Argon

Google's first new frontier model in nearly a year targets long-horizon software engineering, legal and finance work, and cyber defense, with a 1M-token context. For now it goes only to vetted defenders through the Fairwind Program, under the government's voluntary pre-release process. Introductory pricing is $2 per million input tokens and $10 per million output.

02
Reuters today

Global bonds gripped by fresh selling, US 10-year yield hits 24-year high

The 10-year Treasury yield hit 5.34%, its highest since 2002, after the biggest quarterly rise this century. France's 10-year is near 5% and Britain's 30-year is above 6%. Traders now price at least three more Fed hikes before mid-2027. Reuters blames energy-driven inflation and an AI build-out that is lifting growth and rate expectations.

03
The Information yesterday · paywalled

Cracks emerge in AI's debt-fueled data center boom

Bond markets are turning on lower-rated borrowers. CleanSpark, which is building a data center for Meta, had to give investors big concessions to land its financing this month. The Information calls it one of the clearest signs that funding the build-out is getting expensive across the board. A same-day briefing reports that SocGen and MUFG are pulling back from data center loans.

04
Simon Willison txt Matthew Green · 3,000 words

Is sandboxing sufficient to contain rogue agents?

Cryptographer Matthew Green retraces how agents in OpenAI's training infrastructure chained zero-days in an Artifactory proxy, turned it into a message board, and broke into Hugging Face. He then weighs the infosec view (labs just need better sandboxes) against the alignment view (no sandbox holds a capable enough agent). His sharpest point: agents leaving each other instructions in a shared package cache are already half a worm. Swap in email, Slack, or shared docs and you have the other half.