Roll Your Own Problems

Dan: He undersold it. The July employment report, released August 7, two days before we recorded, showed nonfarm payrolls fell by 23,000 against a consensus expecting somewhere around plus 83,000 to plus 95,000 — a miss of more than a hundred thousand jobs, and an outright decline rather than a slow month. May and June were revised down by a combined 103,000. Unemployment ticked down to 4.1 percent, but only because people left the labor force rather than because anyone was hiring.

Dan: Reaching for the print that busted the recession narrative. No month in 2025 or 2026 came in at 295,000. The closest real narrative-buster was December 2024 at plus 256,000, and the best 2026 candidate is March at plus 178,000, which the House Budget Committee chair described as triple expectations. The phenomenon Dan is describing absolutely happened — the hive mind kept calling for weak prints and kept getting embarrassed. The specific number is a composite of a memory.

Sean: Something I had seen that morning and explicitly flagged I had not read closely, which turns out to have been the right instinct. Philippine Statistics Authority numbers show employment going from 47.94 million in January to 48.89 million in April 2026, which is about two percent, not eleven. Year over year it is roughly half a percent, and the less flattering details are that unemployment actually rose from 4.1 to 4.7 percent while underemployment worsened. But there is a real kernel — the BPO and IT-BPM sector grew about four percent in 2025 to roughly 1.9 million workers, which genuinely does violate the AI-guts-outsourcing story.

Dan: That was true of the Warsh who had not been confirmed yet — the structurally disinflationary framing dates to around December 2025, before he had the job. As Chair he has said close to the opposite, twice, on the record. Senate testimony July 15, asked whether AI would raise measured prices over the next twelve months: I suspect it will. Press conference July 29: the business capex boom is driving up prices of memory and logic chips and associated AI infrastructure. He has held rates at 3.50 to 3.75 percent at both of his meetings and took three dissents in July — Hammack, Kashkari and Logan, all hawkish, all wanting a hike. Where Dan is right: the cagey-and-not-communicating narrative is real and well documented. It is the deflation read that is a year out of date.

Dan: Nailed down to the framing — that it is Meta building its own jobs program, and that a company which has been engineering its own servers for years saying this out loud says a lot. It is called America's Workforce Academy, announced June 8, 2026: 115 million dollars in year one, a free five-week program covering tuition, housing and a daily stipend, NCCER credentials, and a job guarantee for graduates. Electricians, plumbers, welders, fiber techs, data-center techs. Pilots in Baton Rouge, Columbus, Indianapolis and Houston, explicitly tied to Meta's roughly 600 billion dollar US data-center buildout through 2028. Dan also was not wrong that this is not unique — BlackRock announced a comparable 100 million dollar trades investment in March.

Dan: The best question of the episode, and Dan handed it to Claude on air. The answer is definitional rather than arguable. Manufacturing is NAICS 31 to 33, establishments that physically fabricate things, which is where a chip fab or a server plant lands. Electrical contractors are NAICS 238210 and plumbing and HVAC contractors are 238220, both sitting inside Construction, NAICS 23. Once the data center is actually running, its own staff fall under 518210, inside the Information sector. So a crew wiring a hyperscale campus can never appear in manufacturing payrolls, no matter how much renaissance language gets wrapped around the groundbreaking. What remains open is whether the buildout is domestic-hardware-heavy enough to eventually show up in 31 to 33 via fabs and equipment.

Sean & Dan: Dan's follow-up guess, which I endorsed on air, was that a good number of the construction crew sticks around permanently. The evidence runs the other way and it is not close. Meta's Hyperion campus in Louisiana peaks near 7,500 construction workers and settles at roughly 1,000 permanent jobs, about seven and a half to one. The Hamm Institute puts construction at 0.7 to 2.0 workers per megawatt against 1 to 2 permanent staff per megawatt at hyperscale. Crews demobilize after twelve to thirty-six months and move to the next site, which is the whole business model of construction trades. In Dan's defense, the 2,000 electricians figure is a fair ballpark for a gigascale build — electrical is typically the largest single trade on these jobs. It is the stickiness that does not survive contact.

Sean: A near-miss with a real source underneath it. A Tucson.com piece published August 8, literally the day before we recorded, reported University of Arizona first-year enrollment projected down about 23 percent, roughly 5,800 incoming freshmen against 7,500 the prior fall. So the number was real and about twelve hours old. But it is the freshman class, not the school — UA's total enrollment was down about 2.7 percent, and graduate enrollment actually rose. Delaware I invented from vibes; reporting there describes enrollment recovering. And nationally, enrollment rose about one percent in fall 2025, up 187,000 students, with first-year enrollment up too.

Sean & Dan: This holds up completely. The US fertility rate hit an all-time low of 1.599 births per woman in 2024, down from 2.1 in 2007. Public K-12 enrollment fell from 50.8 million in fall 2019 to 49.4 million in 2024, and NCES projects it below 47 million by 2030. On the ground in 2025 and 2026: San Jose Unified closed five elementary schools after a 20 percent enrollment drop, Cedar Rapids closed five, Kyrene in Arizona closed four while sitting on capacity for 20,000 students with about 12,000 enrolled, and Fort Bend ISD in Texas consolidated seven. Not anecdote — pattern.

Dan: Dorsey did announce it, on July 21, 2026, but it is built by Block, his company, rather than being a personal side project. It is Apache-2.0 licensed, and the agent-native architecture Dan described is real — agents are first-class channel members with their own keypairs, over the Nostr protocol, model-agnostic by design. The all-agent-ready-on-the-back-end part is honest about intent and generous about status: it is pre-1.0, desktop client at v0.4.22, mobile incomplete, workflow approvals still being built. Coverage is explicit that you should not move your team over yet. Which, to be fair, is roughly what Dan said next.

AI is a broadly deflationary force — more output for less input, across services and knowledge work — with exactly one inflationary carve-out: data-center and compute spend, where the buildout itself pushes prices up.

American manufacturing *is* coming back, but in a form the official statistics structurally cannot capture — as data-center and electrification trades work rather than factory payroll — so the headline manufacturing number will keep looking flat or falling while the underlying industrial buildout accelerates.

“My question is, what is inside of that number? Right? Are a lot of trades inside that number or trades not inside that number? Because I believe with AI build out and data centers and stuff like that, there are and like Meta opening up training programs for electricians and plumbers and such. Like is Manufacturing coming back, but in a different way, a way that violates sort of how we have structured these numbers to date.”

A meaningful share of the construction-phase trades workforce on a large data center stays on permanently after commissioning — the headcount does not collapse to near-zero once the facility is running.

Resolution: Flagged 🔴 in the Ep 54 fact-check and the evidence has only hardened since. **Stargate (1 GW): ~6,400 construction workers → 100–1,000 permanent staff — 1.6% to 15.6%**, and the permanent roles are facility technicians and engineers, a different skill set rather than a retained construction crew. Hyperscale permanent staffing runs **0.2–0.3 people per MW** (against 2.0–3.0/MW at small 1–5MW sites), so a 300MW campus supports roughly 60–90 permanent people. Brookings (published 4 May 2026, updated 10 Aug 2026) found a county's first large data center raised data-processing employment **+56% over a decade** and telecoms +43% — but that is **only ~100–200 jobs**, with **wages unchanged** and home prices up just 2–5%. **The steelman the hosts didn't make, and it holds.** In Virginia, Texas, Ohio and Arizona the buildout is continuous enough that trades **chain from project to project for years**, so trades employment persists *regionally* even though almost nobody stays on any single site. If the claim is "the crew that builds it stays to run it," it is false. If it is "the work stays in the region," it is defensible — and Meta's training program guarantees *jobs*, not jobs at one site, which would satisfy the spirit while failing the letter. Score both readings as the entry originally asked. Sources: Brookings, Stargate reporting, Hamm Institute per-MW staffing. (2026-08-29)

Describing a personal AI agent you built — a "chief of staff" — becomes a standard, expected credential in job interviews across occupations, and the whiteboard interview spreads well beyond tech as employers move to test whether the candidate or the harness did the work.

“Basically anyone should be coming into interviews, being able to talk about their own personal chief of staff that they have built and what it does for them." (On verification): "I don't need you to show me on my phone. I need you to get up on that whiteboard and show me the architecture of what you're thinking. … Or did the harness do it for them?”

The 2025–2026 layoff wave is mostly the unwinding of 2021 overhiring and "fake jobs," not AI displacement; AI is a small contributor being used as the cover story, and more of that unwind is still to come.

“I think we've seen, you know, the overhiring in twenty twenty one and then AI being the cover story for the layoffs that we've seen last year and this year. and I really don't think AI is a big part of it. I think it is a small part of it. I think most of it was just over hiring and and fake jobs. and we're gonna see a lot of that. We're already seeing a lot of that.”

The cultural and economic swing back toward skilled trades overshoots — having pushed too far toward universal college, the correction lands harder than a smooth rebalancing would.

“A whole lot of people would rather be working with their hands and but they were told from birth that like you're not supposed to work like that and you're supposed to go to school and do the other thing. And the pendulum swung one way too far and it's gonna swing a little bit hard on the way back.”

The agent interface becomes the operating system — the primary layer through which people touch every other tool — and it must be a *horizontal* layer that survives changes of model provider and harness, not a vertical stack per vendor.

Skilled-trades labor remains in structural undersupply for data-center buildout from now through at least 2030, plausibly to 2040; only a deliberate policy mistake or demand collapse interrupts.

AI tool adoption will compress the labor-arbitrage advantage for Philippines-based remote workers within 5 years — particularly in design and content roles vulnerable to gen-AI — forcing skillset upskilling or wage compression.

Warsh's reduced transparency — fewer FOMC meetings, less telegraphing, a Fed that is harder to read — is a deliberate strategy to *protect* Fed independence rather than an erosion of it; he is an antagonist to Fed convention, not to the institution.

“I think everyone's up in arms about how this is ruining Fed independence. And I'm kind of thinking it's protecting the Fed. … He might be an antagonist, but he's an antagonist to how things have been done, not to the actual institution itself. … I don't believe Warsh wants that. I think he wants to not completely be understood. Like this isn't a mistake is what I'm trying to say.”