daily notes

previous notes →

Themes

Safety controls moved from compute cost to data-path design

Two independent systems moved automated controls into the live operating path: across model interactions and at skill installation. OpenAI previewed Private Safety Processing for pattern detection without personnel access to content, compatible with Zero Data Retention, and plans a September rollout, while Hermes added advisory install-time scans for PII, leaked secrets, Unicode smuggling, licensing and security, improving 11 bundled skills. Neither release reports measured detection quality or false-positive rates.

@OpenAIoriginal post ↗

We will continue to offer Zero Data Retention for frontier models. As AI takes on longer, more autonomous work and delivers greater value to businesses, safety systems also need to identify risks across related interactions. To help address those risks, we're previewing Private Safety Processing, which is designed to improve safety without giving OpenAI personnel access to the underlying content.

@NousResearchoriginal post ↗

Hermes now leverages NVIDIA's SkillEvaluator on skill installs, checking for PII, leaked secrets, Unicode smuggling, licensing and security issues before you confirm. We pointed it at our own bundled skills first, and used what it found to improve 11 of them. https://t.co/7l1cSz9UlM

OpenAI: Offering Zero Data Retention for frontier models · Hermes skill-system security documentation

Must read

Same agent, different search providerArtificial Analysis

A controlled 25-turn harness held GPT-5.6 Luna constant across seven search providers. Parallel advanced, Exa auto and Firecrawl scored 75, 74 and 73 versus a 33 model-only baseline, making retrieval quality, latency and total inference cost jointly measurable rather than interchangeable.

Signals

Claude produced wet-lab binders for 14 of 15 targetsAnthropic

Claude generated 1,320 designs and 354 binders, with Adaptyv Bio and Twist Bioscience independently testing the proteins. The paid validation found binders for 14 targets, but the model failed on MBP and had only modest affinity on BBF-14, so this is a measured design result rather than proof of autonomous drug discovery.

Sentiment

Controls moved into the operating path +0.21

3 posts · 88% confidence

Themes

$40 trillion marked a level, not a transferable risk threshold

Treasury's August 18 vintage put total U.S. public debt at $40.047T, up $60.8B in one reporting day and above $40T for the first time. A new BIS study found that fiscal expansions raised sovereign risk premia almost three times as much in high-debt regional emerging markets, but its 21-country 2015-25 sample excludes the United States, so the mechanism is a warning rather than a U.S. estimate.

@FToriginal post ↗

Breaking news: The US’s national debt has hit a record $40tn as borrowing rises at a historic pace, fuelling investor concerns about the state of America’s public finances despite Donald Trump’s vow to bring spending under control. https://t.co/JJIO7lHZc7 https://t.co/1N5OXPJtS3

@BIS_orgoriginal post ↗

With #PublicDebt and #InterestRateCosts up, countries in the Americas run the risk that #FiscalExpansions lead to higher risk premia and, in turn, put upward pressure on near-term inflation expectations. Read more here: https://t.co/J0o741qoNh #BISBulletin https://t.co/Y7t7lpmets

U.S. Treasury Debt to the Penny API, August 18 vintage · BIS Bulletin 133: High public debt in the Americas

Must read

No must-reads have been published yet.

Signals

Google's Marvell warrant turns chip spend into equityMarvell Technology

Google can buy up to 58,970,907 Marvell shares at $206.58. Only 1,360,867 shares vest with time; the balance vests in 240 tranches, one per $500M of custom-product revenue through fiscal 2033, tying most dilution to realized accelerator, networking, storage and memory-controller purchases.

Sentiment

Debt sensitivity overshadowed chip optionality -0.22

2 posts · 91% confidence