daily notes

Must read

How SWE-Serve Exposes the Gap Between Local Tests and Live ServingJennifer Williams, Dave Farris, and Jeff Farris

SWE-Serve converts 83 merged SGLang pull requests into 53 executable inference-engineering tasks and shows why code that clears local checks can still fail in production. Its live-serving comparison, verifier design, task families, baselines, and stated single-engine and hardware limits make the full artifact materially more useful than the headline result.

This BIS Bulletin builds an ageing-automation scorecard across more than 130 economies and shows that the industries with older workforces, including health and agriculture, often have less scope for automation than younger workforces such as finance. The full article supplies the cross-country method and policy trade-offs needed to assess AI as a demographic offset rather than a blanket labour substitute.

Sentiment

Serving-path evaluation makes agent reliability more concrete +0.21

83 source items · 76% editorial confidence

Must read

Basel III monitoring reportBasel Committee on Banking Supervision

The current 28-page monitoring report separates stable observed capital ratios from the effects of fully phased-in rules. It covers 149 banks, including 106 large internationally active banks, and gives the underlying liquidity, capital-shortfall, regional, and methodology detail needed to judge the headline.

Signals

BEA's September 24 release puts the revised Q2 deficit at 3.0% of current-dollar GDP, up from 2.7% in Q1. Imports rose $92.2 billion while exports rose $58.8 billion; the Q1 deficit was revised down to $212.6 billion, so the comparison uses BEA's latest vintage.

A September 19 resolution sets the export duty on wheat, barley, and corn at zero through December 31, applying from September 1. The policy lowers a cost in Russia's grain-export chain; it does not by itself establish the eventual effect on global prices or volumes.

Sentiment

Bank capital is steady while trade policy shifts at the margin +0.04

482 source items · 72% editorial confidence