A detailed systems guide to the bottlenecks behind large-scale RL post-training. It makes the trade-offs concrete across colocated versus disaggregated rollouts, gang scheduling, sandbox fleets, asynchronous-policy staleness, weight transport, and failure recovery; the cited examples make it more useful than a generic account of agent training infrastructure.
daily notes
Must read
Introducing Ember-1Fireworks Research
A detailed first-party efficiency report rather than independent validation. It publishes the training rationale, five benchmark tables, and two production A/B tests: Fireworks reports roughly 35% fewer tokens per task at comparable quality, while its model-specific claims and preview status remain reasons to test the result against a reader’s own workload.
Sentiment
Distributed-systems constraints dominate the useful update +0.04
10 source items · 70% editorial confidence
Must read
Banking Analytics: U.S. GSIBs Steadily Increase Exposure to Less-Traditional LendingAmalia Estenssoro / Federal Reserve Bank of St. Louis
The full data note makes a narrowly useful balance-sheet distinction: U.S. GSIBs’ less-traditional loans rose to 12.6% of assets at June 30, 2026 from 10.1% a year earlier, with NDFI lending and securities-based lending central to the shift. It also records the 2025 asset reclassifications, so the reader can separate a real composition change from a clean credit-cycle signal.
Signals
$SPX is expected to report Y/Y earnings growth of 29.1% for Q3 2026, which is above the estimate of 26.7% on June 30. #earnings, #earningsinsight, https://t.co/nWKwAnOzaw https://t.co/nrnUHEd1oO
FactSet now puts expected third-quarter S&P 500 earnings growth at 29.1%, above its 26.7% estimate on June 30. It is an estimate revision rather than a reported result, so the useful read is the direction and magnitude of expectations—not confirmation of delivered earnings.
Sentiment
Primary research adds monitoring context, not a market call +0.08
46 source items · 78% editorial confidence