daily notes

The eight-page report shows how confidential computing can keep model weights from evaluators and test prompts from the model developer. Its practical value is in the attestation, code-approval and encryption workflow, then the candid account of what the pilot did not solve: some proprietary code was not inspectable, builds were not independently reproducible, and Google services remained in the verification path. Read it before equating private evaluation with trustless evaluation.

Read the operating case studies, especially QuEra's laser controller and Genentech's failed retries. At QuEra, an agent developed a deterministic controller that recovered laser lock in 695 of 700 blind tests; the deployed controller did not need the model to reason at every step. Elsewhere, repeated attempts made a physical failure worse until an expert supplied missing context. The preview's shared device interface and enforced limits help with integration, but they do not remove the need for domain knowledge.

The useful detail is how the pieces fit: sparse attention selects context, gated residuals widen information flow, and a separate 51B-parameter N-gram table can sit in host memory with asynchronous prefetching. Read the architecture and deployment sections before treating the 6B active-parameter count as a memory budget. Reported speed gains depend on Qwen's test setup, including a 90% cache-hit assumption in its 1M-context comparison.

The August 26 tagged release adds disk-backed KV offloading and external secondary-tier managers, extending the cache architecture discussed on August 25 into a concrete upgrade. It also raises the default batched-token limit from 8,192 to 16,384 and moves bitsandbytes into an external plugin. Check capacity assumptions and plugin installation before upgrading; these are release changes, not a new universal throughput claim.

Reuters relays The Information's report, based on one person familiar with the deal; neither company had responded to Reuters. If confirmed, the transaction would bring a major model-and-dataset distribution platform under a chip supplier. Agreement, closing, and future platform rules remain separate questions: no completed acquisition or change in access policy is established.

This 21-page note explains why rate-path commitments made sense near the effective lower bound but became costly when supply shocks changed the inflation outlook. Its scenario-based framework separates the policy-rate stance from liquidity support during financial stress, then traces how compressed volatility can distort risk-taking. It is a useful guide to interpreting conditional central-bank language, not a forecast of the next Fed decision.

The useful distinction is between fast customer access to money and final settlement. FedNow had 1,725 participants in Q1 2026 but processed 2.73M payments, compared with RTP's 128M; the systems also serve very different transaction sizes. The authors trace how cards, ACH and instant-feel wallets reduce the incentive to adopt new rails, and how prefunding changes liquidity costs. Read the membership and usage comparisons before treating bank enrollment as evidence of consumer adoption.

The charts and method repay the short read: the authors strip out boilerplate, compare statements with minutes, and test word counts against inflation and unemployment. Four minutes sections lost at least 200 words relative to the final Powell meeting; the latest statements look unusually short relative to inflation, but not unemployment. The study measures a communication change without claiming that fewer words improve policy outcomes.

July's advance goods deficit widened to $118.8B from June's $101.4B: imports rose $11.4B and exports fell $6.0B. The import rebound changes the narrowing story in the June report covered on August 4. These August 27 estimates are seasonally adjusted and nominal, cover goods rather than goods and services, and remain subject to revision; they do not establish the real GDP contribution.

The August 27 vintage puts Q1 2026 global real house prices 1.2% below a year earlier, after a 0.5% decline in Q4. Yet the median jurisdiction's prices rose 2.6%, up from 2.3%. The weighted decline reflects weakness in several large economies, including China and Canada, not a synchronized fall across most markets. Keep the country weights and inflation adjustment visible when interpreting the global headline.

Marvell reported Q2 FY2027 revenue of $2.739B, up 37% year over year, with data-center revenue up 46%. Its Q3 revenue guide is $3.150B plus or minus 5%, roughly 15% above Q2 at the midpoint. That is another supplier's measured revenue base and forward delivery target; the guide is management's expectation, not booked revenue. Non-GAAP gross margin is guided to 57.5–58.5%, versus 58.9% reported in Q2.

The FDIC's August 25 report covers 4,238 insured institutions: Q2 net income rose 12.0% from Q1 to $90.1B, while loans grew 1.8% and domestic deposits 0.8%. Net interest margin increased just one basis point to 3.32%; past-due, nonaccrual and net charge-off rates declined. The aggregate is stronger, but the profit increase should not be read as a comparable expansion in lending spreads or evidence that every loan category improved.

EIA's August 26 weekly estimates put distillate stocks 14% below the five-year seasonal average for the week ended August 21, versus 12% in the July 31 data cited on August 5. Gasoline and distillate stocks fell 2.5M and 2.2M barrels despite 97.4% refinery utilization. Four-week products supplied fell 3.0% year over year: thin product buffers are not evidence of a broad demand boom.