daily notes

By mid-August, OpenAI says its research organization used 3.1 agent-workdays for every human workday; experiments per active experimenter reached a record, and success without intervention on estimated 4–8-hour tasks rose from 18% to 53% between January and July. The report also exposes the constraint: its measures are preliminary proxies, and more than half of successful 4–8-hour tasks still needed human steering.

Pachocki separates goal alignment from value alignment and argues that chain-of-thought monitoring is weakening as models mix tool use, communication and less-verbal reasoning. His policy conclusion is unusually concrete: no lab can responsibly scale at maximum speed much longer without shared safety bars, outside auditing and voluntary slowdowns when confidence falls short.

The public experiment says it has scored 4,442 finance and economics papers since July 14 and exposes separate rankings for significance, originality, correctness, data and exposition. It is useful as a discovery queue, but automated scores are not a substitute for reading or human review.

Michael Pettisoriginal post ↗

1/8 Bloomberg: "China is injecting billions into its largest banks and insurers, as Beijing moves to shore up balance sheets and sustain growth in a slowing economy. At least eight financial institutions are... https://t.co/YUXKfaGWII 2/8 seeking 360 billion yuan in fresh capital, with the Ministry of Finance providing more than 80% of the funding. The fresh injection is aimed at easing margin pressure, expanding lending capacity, and bolstering provisions against potential bad loans." 3/8 This is in principle good news because Chinese banks are seriously undercapitalized, given the quality of their assets, and to encourage them to keep lending they need to expand their accounting capital. 4/8 The problem, as you learn (or should learn) in your basic finance course, is that while additional capital improves the safety of the individual entity that receives it, it only improves the safety of the overall financial system if it comes from outside that system. 5/8 When the banks themselves recapitalize each other, this improves safety only if the risk is that a few individual banks are overextended. If the whole system is overextended, the additional capital does no good at all. 6/8 Those who don't understand why should take a look at the Mexican banking crisis in 1995. The banking system had plenty of capital in theory after it was privatized in 1991-92 (I was part of the team that advised the Mexican government on its privatization), but because... 7/8 much of each bank's capital had been directly or indirectly provided by other banks, none of this mattered when the system itself got into trouble. Each bank's loss became a system loss. 8/8 This "new" capital, in other words, only tells us what we already knew, which is that the MoF is on the hook for the banking system. It doesn't in any way change risk profiles except on paper. If there is a systemic problem, we will all quickly relearn the lesson.

Eight state financial institutions plan to raise 360 billion yuan, with China's Ministry of Finance providing more than 80%. Pettis distinguishes protection for an individual bank from protection for the system: when capital and losses remain inside the same state-backed balance-sheet network, he argues, the recapitalization expands accounting buffers without removing aggregate risk.

August unemployment rounded to 4.1%, but the unrounded rate edged from 4.0901% to 4.1413%. The decomposition is the useful part: the rise came mainly because fewer unemployed people left the labor force, while job-finding and job-loss flows stayed near recent averages; payrolls rose 162,000 and June–July revisions were positive.

US-listed ETF assets reached $16.4 trillion in August and took in $182.6 billion even as monthly flows slowed 5.5%. Fixed income captured 33.5% of flows and commodities 5.8%, while 134 launches lifted the year-to-date count to a record 1,023; about a quarter of August's new funds used leveraged or inverse exposure.

FactSetoriginal post ↗

The trailing 12-month P/E ratio for $SPX of 26.5 is above the 5-year average (24.4) and above the 10-year average (23.5). #earnings, #earningsinsight, https://t.co/448bbXAxa7 https://t.co/uozsuG45Xy

FactSet puts the trailing 12-month S&P 500 P/E at 26.5, versus a five-year average of 24.4 and a ten-year average of 23.5. The comparison keeps valuation risk visible even though the forward multiple sits closer to recent norms.