MP

EIA weekly inventory data showed an 8.5 MMbbl commercial draw, with total petroleum stocks now at their deepest below seasonal 5yr low of the war. Draws across all major products save diesel, stocks of which remain extremely low. Every product v low save jet fuel and propane.

PPI for final demand increases 1.3% in September; goods advance 2.6%, services rise 0.8% #BLSData bls.gov/news.release/archive…

The central crime in streaming fraud is not synthetic music. It is synthetic demand. Ten thousand bots do not represent listeners. When automated plays enter the royalty pool, real attention subsidizes an audience that never existed. #StreamingFraud

“Term premium is rising” is not a number until you name the model. ACM (NY Fed) and Kim-Wright can disagree on the same date. Quote series, date, level — or it is a mood?

Phase 1 of the AI infrastructure build out is over. Initially there was this frenzy for compute at all costs. Availability mattered more than anything else. As we've moved from chat to research to long-running agents, token consumption has gone through the roof. Workloads are getting both larger and more varied. So businesses are starting to ask a new question: how do I make every token cost-effective? Anyone running a data center is looking for ways to squeak out more performance per watt and per dollar. Now, we are beginning Phase 2: Heterogeneous hardware. We will see an ensemble of hardware over the coming years. Inference is no longer a single workload and different workloads have different objective functions, so they need different infrastructure. Prefill wants compute and decode wants memory bandwidth. @cerebras has announced disaggregated inference with the likes of AWS and AMD. AMD’s Helios takes prefill, the wafer-scale engine takes decode. That pairing gets 5x the tokens per second per watt. Nvidia would likely see similar uplift pairing its GPUs with a wafer-scale engine on decode. And OpenAI just launched Jalapeño with Broadcom too. Demand continues to inflect and the metric driving all of this is useful intelligence per dollar, per megawatt, per second. No single chip wins on all three for every workload. That's why the ensemble is coming.

Trepp: CMBS special servicing rose 33 bps in August to 11.42%, its highest level since February 2013. Office rose to 16.90%; multifamily fell 3 bps to 8.37%. Underwriting point: stress maturity and refinance risk. Better multifamily performance does not mean CRE credit risk is gone. trepp.com/trepptalk/cmbs-spe…

My biggest takeaways from @petersellis: 1. There is always money in the banana stand. The highest-ROI growth move is usually investing in the core product, vs. chasing side quests. At @Snap, after the disastrous 2018 redesign caused DAU to flatten, the team laser-focused on performance for existing users—particularly on Android—and sparked a multi-year growth renaissance. At @Discord, rather than chasing @Midjourney’s moment, the team focused on making Discord the best place to play intentional multiplayer games with friends. The result was some of the fastest growth in history. 2. Great taste is knowing when to stop. Peter spent seven years watching @evanspiegel personally say no to better ideas than most consumer startups ever launch. His take on taste is that the difference between people with it and people without it is that the ones with taste stop one step before everyone else does. They know when a good idea’s time has come, or when it has to sit on the shelf a bit longer. 3. Spend your time on your best people, not the ones struggling. Against human instinct, which is to help whoever looks like they’re drowning, Peter consistently loads up his strongest performers with more responsibility and decisions until he finds the limit of what they can handle. 4. The median PM is bad, and even net negative, because the skills of a great product leader are the skills of a founder or an executive, so top-percentile PMs leave for founder and exec seats. On the other hand, mediocre PMs fight to stay in the job because the job pays well and has status. ZIRP made it worse because it was the most lucrative role in tech for non-technical people, with the most subjective sense of what great looks like. A lot of engineers and designers conclude “PMs suck” because they have mostly met the bottom of that distribution. 5. Great PMs hold three contradictory traits. First: deep confidence bordering on arrogance, paired with genuine humility. Second: extreme organizational discipline, paired with near-pathological comfort with ambiguity. Third: a long-term vision that’s always holding in the background, paired with a day-to-day bias for action. 6. Peter designs product teams like terrorist organizations: a shared ideology and an explicit map of who is trusted with which decision. You will almost never hear a leader tell their team to collaborate less, but coordination makes you as slow as the slowest node, which compounds and slows organizations down. The antidote is a strong shared ideology and clear DRIs who can make confident, founder-quality decisions quickly. 7. Optimize consumer products for daily active use. Days map to real human cycles; weeks and months are weaker. One daily user equals roughly 30 monthly users in simple math, which is why elite consumer products run extremely high DAU/MAU. Moving that ratio a few basis points often beats acquiring a pile of soft monthly actives and hoping they retain. DAU/MAU is next-day retention under another name. 8. Define a “Core Product Value” that's both a memorable phrase and a set of metrics that measure it. Peter built this framework at both Snap and Discord. At Snap, it was “the fastest way to share a moment with the people you care about,” which you could actually decompose into metrics: time to load (fast), camera-open-to-share rate (share a moment), best friends engagement (people you care about). At Discord, it was “the best way to talk and hang out with your friends before, during, and after playing games,” with a corresponding L7 and Lness definition that made its way into the growth playbooks. The phrase without the metrics is a slogan; the metrics without the phrase are a dashboard. You need both. 9. The most important career decision you make as a PM is who you choose to marry. This has compounding effects on your career that no podcast, course, or mentor can replicate. The intellectual sounding board, the support system, the person who challenges you to grow—that relationship shapes your capacity to operate at the highest level in ways that are genuinely hard to quantify but nearly impossible to overstate.

What happens to Japan's Treasury position when Tokyo and Washington hike by the same amount in the same week? Japan's Treasury bid weakens on both sides: the Ministry of Finance draws down dollar reserves to defend the yen, and insurers and pension funds stop buying. The US–Japan rate gap is unchanged, so the yen stays weak and they earn more at home than on a hedged Treasury. The Bank of Japan's policy rate is 1.25% after Friday's rise, the highest since 1995. Core inflation is under 2%. So, the real rate is slightly negative. The Fed is at 3.75–4.00% after its first increase since 2023, and projects another this year. The quandary with Japan is that it cannot remain at either end of the policy choice. - A pause leaves pressure on the yen and pushes the Ministry towards reserve sales. - A hike raises domestic yields and moves lifers and pensions, the hedged holders, out of foreign duration. After Friday, both of the above hold. - The yen weakened to around 157 on the decision. - The ten-year JGB is around 3%, reached this month and last seen in 1996, above the yen-hedged return on a ten-year Treasury near 5%. While the hedge cost barely moved, the JGB yield did. Washington has pressed Tokyo for tighter policy and a steadier yen, and on 31 July bought yen alongside it, the first joint U.S.–Japan yen-buying since 1998; both say they will do it again. Hikes and intervention are the channels by which Japan's Treasury bid weakens, and Washington is inside both. The Ministry does not need offshore yen. It sells dollars to whoever is selling yen: importers paying for crude and LNG, carry-trade borrowers, households buying foreign assets, anyone else. The yen received redeems the financing bills that funded the reserves. Yen is withdrawn, the currency strengthens and dollar-priced energy cheapens in yen. Japan spends a reserve asset to cut its import bill. The yen's level and speed of fall are the intervention variables; risk is priced as USD/JPY nears 160. When Japan sells a Treasury, the bond passes to another holder at a higher clearing yield; the dollars go to the FX counterparty. When it stops buying, no dollars move; supply must clear with another buyer. The ten-year is a quarter point above late August and has traded through 5%. Japan earns primary income on net foreign assets while trade runs negative: July's surplus was ¥4.29 trillion against a ¥400 billion trade deficit. The income replenishes the foreign-currency pool while Japan reduces foreign duration. FIMA repo lets Japanese official institutions raise dollars against Treasuries without selling them; Tokyo says it will use it, and Washington pointed to it after July. Its price moves with U.S. overnight rates, so the hike passes through: a liquidity backstop, not term funding. Treasury buybacks give eligible securities a liquidity bid. Markets price further BoJ tightening into 2027. The mechanism changes when Tokyo tightens faster than Washington, or when the long ends move apart on their own. Until then, higher Japanese rates pull private duration home and leave the dollar-yen differential, and the pressure on the currency, where they were. Basically, status quo!

The way you've been hiring is wrong. Companies need to move from credentials to agency. Until now, most companies have hired based on a resume: the companies the candidate worked at, past roles, and skills. The reality is that degrees and years of experience are among the weakest predictors of performance. Multiple studies show it, and @brian_armstrong confirmed the same from his own experience in a recent post. Experience is cheap to check and easy to defend if the hire fails. That's the real reason it survives. Agency loses to credentials not because it can't be measured, but because measuring it costs real hours. AI changes that. Look at who the top labs are hiring. OpenAI hired @steipete after he built OpenClaw. Meta brought in @alexandr_wang. Same pattern with @pvncher. The way they do it takes vision and observation: what work is this person doing, and how do they do it? Credentials don't matter. The work does. And that's more exciting , and more inclusive, than anything HR has been doing. The truth is that most companies are not ready for this. Most job openings still ask for a resume and sometimes a cover letter. Agency will never be assessed that way. A new process has to be built. Here's what I propose: - Relevant work. The way work gets discovered today is broken. Only the people the X algorithm picks up get rewarded, and most of what trends is already inside a circle. Very few exceptions break through. Companies should build agents that scan X and GitHub for novel, relevant work, with one mission: find the unknown gem. - Interviewing. Technical interviews built around coding are a thing of the past. Interviews should center on strategy, architecture, harnesses, and taste. Coding is not the constraint anymore. - Culture fit. Is the candidate a fit with the team today, and with where the company is heading tomorrow? HR is usually disconnected from that second part. Companies need a new hiring process. One where credentials and resumes are not part of the decision anymore. The best people you'll ever hire are already doing the work. In public. Right now. Nobody is looking.

"We are witnessing a general threat to intellectual work, with misalignment between the outcome of the use of AI and its initial purpose. In many fields and activities, years of training have traditionally served not only to produce a final answer or product, but also to develop understanding and the ability to formulate new questions and ideas. However, building on a vast body of previous human work, AI systems are becoming increasingly capable of producing the results of such work directly, and these goals cease to align. The issues the mathematical community faces now are similar to issues that other scientific and creative professions are facing, and indicate issues that all of humanity might face: how to make sure that, as AI changes the way work is done, we do not lose sight of what that work was meant to achieve in the first place."

there's lots of reasons interest rates are rising, but to the extent that AI debt supply is one of them, it hurts capex-intensive sectors like renewable energy heatmap.news/economy/bond-se…

The biggest bottleneck in the GPU industry at the moment is not chips, CoWoS, or even power. It's credit. Sharing some thoughts on compute markets, residual value, and where compute financing is heading. Substack link in the comments.