daily notes

Artificial Analysis pairs a public, reproducible runner with a fixed 168-turn, 172-tool-call trace that carries the prior conversation forward. Its comparison makes an operational distinction that token benchmarks hide: among the four tested systems, completion time depends on memory capacity, bandwidth, prefill, decoding, and model fit; users can inspect recipes and reproduce runs instead of treating a leaderboard as a generic model ranking.

Using long-run U.S. GDP-per-capita decompositions, the Dallas Fed separates productivity from hours, employment, participation, and demographics. It finds productivity is the only component without a hard upper bound and the persistent positive contributor across eras, while explicitly stopping short of quantifying any AI effect; the data and accounting identity make it useful for assessing productivity narratives.

BEA raised Q2 real GDP from its 1.5% second estimate to a 2.2% annual rate, citing upward revisions to investment, consumer spending, and government spending. It is a third estimate of the April-to-June quarter, not a current-quarter forecast; corporate profits were revised down $16.9 billion to a $384.0 billion increase.

August headline PCE rose 0.3% month over month and 3.4% year over year, while core PCE rose 0.2% and 3.0%, respectively. Personal income rose 0.2%, spending 0.9%, and the saving rate was 4.1%; BEA also revised January-to-March compensation estimates, so the release is a current snapshot with historical revisions.