Jul 27, 2026 · 49:17
Unqualified Fact Check
held up
Zero whiffs on the facts — the only thing we truly lost this week was the $250.
You already know from the title how the bet went. On the Fourth of July — America's 250th birthday — I sat in a living room in Dallas, fresh off watching Egypt beat Australia on penalty kicks with 71,000 of my closest friends, opened Kalshi for the first time in my life, and put $250 on the USA to win the whole World Cup. In honor of the birthday. The odds said 4%, my heart said 50/50 against Belgium, and my inner options trader said buy the pre-game price, sell the sentiment spike. Belgium had other plans. But the really unsettling part, as Dan and I get into, is the friction inversion: Dan couldn't figure out how to watch the World Cup — three different Fox apps and no answer — but funding an account and betting on it took five minutes. You can't pay for the game, but you can bet on it. Sit with that one.
Then Dan did something no guest of this show has ever done: he formally appealed a yellow card from our own fact-check. His Amazon "54 cents of every dollar" claim, it turns out, came from a US-sellers-only dataset — so the appeal is granted, the ruling is amended, and there's now officially a corrections desk at this podcast. Which led us to grade ourselves out loud: our fact-checks are holding up around 85%, our predictions a bit better than 50% — poor on timing, pretty good on direction — and you can now audit all of it yourself right here on this site, the thing I built when I kept telling you to go build something.
The back half is the AI economy from the ground up. I priced out the local rig I'd need to run my data-processing jobs at home and got a number north of $225,000 — a genuinely hard choice between a server rack and a Porsche 911 GT3 RS. Memory prices have gone parabolic, which raises a real question about who gets priced out of building. Dan brings the antidotes: a company called Etched rethinking AI hardware from first principles (founded by guys roughly 24 years old in an industry where founders have always been over 40), his architect-and-subagents doctrine for spending frontier-model tokens wisely, and a policy pitch I actually like — if the government takes stakes in frontier labs, year-old models should go open source by default.
And then it gets big-picture: a CS professor at IU who thinks freshmen should start their education in the CEO's office and work down, commencement crowds booing any mention of AI, and our long-running argument about whether the fear is justified. Dan says the creativity-and-jobs fears are completely unfounded and history backs him — the elevator operator is basically the only job technology ever fully deleted. I think the change rolls through one sector after another, a thousand little tweaks rather than one sharp shock. We land on 1929, radio, aeronautics, and why this looks to us like a boom with fragmented little bubbles on the financing side rather than one big systemic pop. So far. That can change.
We close on a question I keep chewing on: why shouldn't 100% of people own stocks? Dan's answer — that the best returns are hiding from the public on purpose — took us somewhere I didn't expect, and gave us the line of the episode: capitalism, like Harvard, is a filtering mechanism. Not a guarantee of results. But a pretty rigorous filter.
Stay frosty and go build something.
Cheers, Sean
Sean: The odds were right — Kalshi had USA ~5% in late June and ~4% entering the round of 16 fits, with Sean's 6.25% straight math correct — but $250 at 4 cents pays about $6,250, not $60,500; the payout gained an extra digit in the retelling.
Sean: Confirmed — Australia 1-1 Egypt through extra time, Egypt winning the shootout 4-2 at AT&T Stadium in Arlington; one asterisk, the match was Friday July 3, not Saturday, and the 71,000 attendance is plausible but unverified.
Sean: Confirmed — the round-of-16 match was Monday July 6 and FOX Sports headlined USA as slight favorites, so a coin flip was a fair read of the lines; the claim was the odds, not the 4-1 outcome that named the episode.
Sean: Confirmed and precisely hedged — DDR5 roughly doubled, CNBC called it an unprecedented AI-driven surge, and Tom's Hardware reported the surge cooling mid-2026 while prices kept climbing through Q3.
Dan: Confirmed despite Dan's loose-and-goosey caveat — the Sohu transformer ASIC, Harvard-dropout founders in their early-to-mid twenties in an over-40 industry, a recent Invest Like the Best episode (EP.480), and a $300M raise at $10.3B three weeks after recording.
Sean & Dan: The 1.4nm roadmaps are real and broke the week of recording (Samsung targeting 2029, Intel 14A), but TSMC was already in 2nm volume production and node names are marketing labels — real transistor features are roughly a hundred atoms across, not one.
Dan: Bessen's famous 1950-census study confirms exactly one occupation eliminated by automation, but spreadsheets cut ~400K bookkeeping-clerk jobs while adding ~600K accountants — net expansion, wrong species.
Sean: Textbook correct — RCA was the emblematic mania stock and aviation boomed after Lindbergh; real technology, wrong prices, unclear winners is exactly the parallel the literature draws.
Fears that AI destroys creative work and jobs are wrong — creativity and jobs ultimately expand as more individuals are enabled to create; only the transition timing (9 months to 10 years) is in question.
“I think both of those fears are completely 100% unfounded. It's just a matter of like, does it take us three years to get me to the right or nine months or 10 years? … I think the creativeness will actually explode because more individuals will be enabled to be creative and same with jobs.”
The AI bubble stays confined to the financing side — fragmented little pops, not one systemic crash; bankers and investors get hurt but everyday people don't feel it exceptionally hard.
“I think AI is more like a boom, less bubble. They'll be fragmented little bubbles, but not one big pop. … I think the bubble will exist in the financing side of it, but … it's all localized enough that it's not systemic.”
The open-source LLM layer will be won by Chinese labs.
“It seems like that layer is going to go to them.”
Entry-level CS grunt work (assigned-error debugging, junior troubleshooting) is essentially eliminated by AI by the time 2026's freshmen graduate (~2030).
“That grunt worker is basically gone or will be gone by the time they graduate.”
Fable-Five-level (frontier) capability will be back in consumers' hands within 6–12 months — "if it's not Fable from Anthropic, it'll be open source."
“Getting this back in our hands in six to twelve months seems like a guarantee. And if it's not Fable from Anthropic, it will be an open source.”
Open-source models at ~3/4 of frontier capability will run locally on ~$5K of consumer hardware (DGX Spark + Qwen-class weights), and rising post-IPO token prices will push more power users to home compute.
“All of a sudden I'm looking at a five-thousand-dollar piece of hardware and an open source model that is three quarters capable." / "If it is three-quarters as capable… Man.”
2026 will see meaningful pushback against AI-hype pricing — valuation compression, slowed capex, or narrative turn.
