daily notes

On-policy RL can silently stop being on-policy when training and rollout engines use different kernels, batch shapes, or parallel layouts. IsoExec turns the rounding-sensitive choices into an execution contract and aligns model kernels across vLLM and Megatron. On one 8×H100 Qwen3.5-35B-A3B run, it cut the average log-probability mismatch to below 10⁻⁶ with 25% end-to-end overhead. The useful part is the engineering trade: deterministic parity becomes testable, but it is not free.

The Speech Agent Arena separates human conversational preference from successful tool execution. Participants compare hidden systems on matched scenarios, while agentic runs are checked against the required final tool calls. That design exposes a product risk conventional voice leaderboards miss: a fluent interaction can persuade the user that a booking or order succeeded even when the decisive tool call did not.

@FToriginal post ↗

Anthropic’s best AI model struggles to attract users as cheaper tools thrive https://t.co/lBg4Bgs6Jj

Ramp spending data from 70,000 companies put Fable 5 at about 11% of Anthropic tool spending more than two months after release, while the cheaper Opus 5 had already passed it. Anthropic still reported July annualized revenue of $65B, up from $47B in May; this is a product-mix constraint, not a growth collapse. Frontier labs now have to defend the cost of their largest models against capable lower-priced alternatives.

@yuntiandengoriginal post ↗

PSA: @sytucr appears to be compromised. I received a DM from the verified account about a Bloomberg interview, followed by a Calendly-looking booking flow. It eventually asked me to authorize an X app called "Iphone IOS", which then posted crypto/token posts from my account. If you receive a similar DM, do not authorize the app.

Yuntian Deng reported that a verified account pitching a Bloomberg interview sent a Calendly-like booking flow that culminated in X app authorization; the app then posted crypto messages from his account. Verification status did not authenticate the outreach. Treat interview-booking flows that unexpectedly request social-app authorization as account-takeover attempts.

This 65-page BIS working paper separates three forces often bundled together when central banks estimate how far quantitative tightening can go. Liquidity rules need not raise reserve demand when banks can substitute other high-quality liquid assets; monetization frictions reshape the curve and move its satiation point; fragmented interbank markets can destroy the one-to-one link between aggregate reserves and the policy rate. The distinction changes whether the remedy is a larger balance sheet, better repo-market liquidity, or less fragmentation.

The New York Fed argues that wider stock-market participation spreads equity risk across more households, reducing the average exposure to a given change in market capitalization. Its time-series and state-level evidence links that diffusion to smaller consumption, valuation, and investment responses after rate shocks. The implication is easy to miss in policy models: household portfolio structure can change the strength of monetary transmission over time.

@michaelxpettisoriginal post ↗

But what a depressingly large number of economists don't understand is that exchange rate adjustments do alter the saving/investment balance. They effectively shift income between consumers of tradable goods and producers of tradable goods. An appreciation, for example, puts downward pressure on saving by shifting income from net producers (e.g. manufacturers) to net consumers (the household sector), thereby rasing the consumption share of GDP. In his 2008 paper, “The Real Exchange Rate and Economic Growth”, Dani Rodrik makes this point, arguing that a real depreciation is equivalent to a production subsidy plus a consumption tax on tradables. He then says that directly subsidizing tradable production would achieve the production-supporting effect without the consumption-tax effect. And it's not just Rodrik. Almost everybody used to know this. In his 1944 book, Ragnar Nurkse said that “the devaluation of a currency is expansionary in effect if it corrects a previous overvaluation, but deflationary if it makes the currency undervalued.” He was making the same argument -- if saving is too low, devaluation expands production, and if it is too high, it weakens production. Perhaps most famously, in his classic 1965 work on Argentina, Carlos Díaz Alejandro emphasized the redistribution caused by devaluation. A devaluation raises the domestic prices of tradable goods, reducing the real purchasing power of workers and consumers while benefiting producers and exporters. Because different groups have different propensities to consume and save, he explained, the redistribution can have major effects on aggregate demand and saving. Peter Montiel in 2000 and Montiel and Servén in 2008 said the same thing, arguing that an undervalued currency changes intertemporal relative prices, discouraging consumption in favor of saving and making investment in tradables relatively more attractive than investment in non-tradables. Consequently, when the currency eventually appreciates, consumption rises and investment shifts toward non-tradables. This is hardly a new idea. More generally, the distinction between policies that are intended to change an external imbalance and policies that are intended to change an internal imbalance is a spurious one. You cannot change one without the other, and either policy works by changing the saving/investment balance.

Michael Pettis connects Rodrik, Nurkse, Díaz Alejandro, and Montiel around one mechanism: currency moves shift income between consumers and producers of tradables, changing consumption, saving, and investment. An appreciation is therefore not merely an external-price adjustment, nor can an internal imbalance be fixed independently of the external balance. The post is useful as a compact test for claims that devaluation or appreciation operates without distributional consequences.

Alibaba priced 710 million newly issued shares at HK$112.70 for an HK$80B placement, with 100% of net proceeds earmarked for full-stack AI capabilities and infrastructure. Closing is expected August 26 and remains subject to customary conditions. The financing makes the next AI buildout an explicit equity raise rather than another inference from capex guidance.