Unqualified AdviceUnqualified Advice
← All Episodes

Sep 8, 2026 Β· 59:43

Artificial Laziness

Unqualified Fact Check

50%

held up

4Nailed it
6Close enough
4Whiffed it

The ideas held up and the numbers took a beating β€” every red is a figure imported from the right neighborhood and the wrong house.

Show Notes

Dan started with a question about AI agents: how do you stop a thing that keeps writing down its own rules from going rigid? He was picturing a government. I was picturing red tape, and the policy documents I'd spent the week wrestling with. In the end, you wind up with the rule is C, and nobody remembers that C only ever followed from A and B β€” so when A changes, things downstream don't adapt quickly. Dan's answer: build deterministically, break the process into the smallest steps you can, and layer intelligence into only the one narrow place that actually needs judgment. It sounds a lot like training an employee, no?

The middle of the episode is a boat. Ed Conway β€” whose Material World was my favorite book of its year β€” published a post the morning we recorded asking whether we've hit peak ship, and I brought it in as settled fact. Reader, I should have read to the end. The fact-checkers have notes. What survives the correction is the more interesting question anyway: what happens to a system when the curve that's been quietly fighting inflation on your behalf for forty years finally flattens out? That took us to Houthis, chokepoints, insurance actuaries, and my favorite tangent of the year β€” why shipping fortunes are shadowy in a way tech fortunes structurally cannot be. A Silicon Valley billionaire has one cash cow. A shipping magnate has Panama, Greece, and Angola.

Then Dan announces a business. He's launching Polyiota β€” polyiota.com, as promised on air β€” consulting on AI transformation for exactly the kind of company that will never buy itself a software team. There is no such thing as ready, so you put the rivets in the airplane while it's already flying. His field report from his early efforts with his own team gives us a glimpse into the adoption headaches he hopes to help others avoid: the blinking cursor effect. The fix isn't just more intelligence.

We end up somewhere I didn't see coming: not why not, but where's next. Cheap land moved technology west a century ago, and Dan's argument is that the underrated story of the next decade isn't AI at all β€” it's people moving again. Great Lakes. College towns. Columbia, Missouri. Then four million boomers a year start retiring, Dan does the arithmetic live and out loud with predictably mixed results, and we close on what the millennials will do with the inheritance. My guess was tattoos and beard oil. Dan, for the record, has neither.

Go build something this week.

Links

Timestamps

  • 00:00Cold Open: These Are Just Employees, Dan 00:41 - How Do You Keep an Agent From Ossifying? 02:40 - This Isn't an AI Problem, It's an Org Problem 03:23 - The Rule Is C, and Nobody Remembers A and B 04:42 - Artificial Laziness and the Cost of Thinking 05:31 - I Don't Know Why I'm Paying You, Claude 06:54 - The Sycophancy Data: Who Holds a Correct Answer 07:18 - Model Vibes, Flavors, and Choosing for the Task 09:57 - Build Deterministic, Layer Intelligence Thinly 10:31 - Ed Conway and Are We Approaching Peak Ship? 12:27 - The Moore's Law of Shipping Containers 13:37 - Houthis, Chokepoints, and Just Building More Ships 14:07 - China Has Built More Ships Than We Have Since WWII 16:29 - Insurance, Actuaries, and Ships as an Asset Class 17:57 - Why Shipping Money Is Shadowy Money 19:45 - The Bouncy Ball, the T-Shirt, and the Alaskan Fish 21:27 - What We Stopped Paying For 22:40 - Making Millions of Something Is a Different Mindset 22:59 - Should We Get Dan's Brother On to Talk Amazon? 24:46 - Just Hire Me: Dan Launches Polyiota 25:01 - Riveting the Airplane While It's in the Air 27:03 - The Real Customer Is a Business Changing Hands 28:11 - Compliance Is the Differentiator 32:10 - The Blinking Cursor Problem 33:12 - Zapier, the 80/20, and Making Automation Legible 36:49 - Sizing the Engagement by the Back Office 38:24 - When Does the Model Crown Change Hands? 38:58 - Grok Boys, and Why Your Feed Is New York Coded 42:26 - Don't Linger: You Control More of Your Algorithm Than You Think 44:05 - What YouTube Knows That X Doesn't 45:34 - Good Pope, Bad Pope: Benevolence as Infrastructure 46:25 - The Denver Illuminati Detour 47:21 - When the Midwest Was the Richest Place in America 47:54 - The Jones Act Interlude 49:22 - Where's Next? Cheap Land and the Next Migration 51:13 - The Sunbelt Happened. Are the Great Lakes Next? 53:01 - The Case for College Towns 54:13 - Aggregation Theory Comes for Higher Ed 56:02 - The Silver Tsunami: Four Million a Year 56:22 - Doing the Boomer Math Live 57:28 - Tattoos and Beard Oil 58:41 - Embrace Your Actual Age

The Fact Check, claim by claim

Material World is about six raw materialsNailed it

Sean: Sean hedged at six or seven and landed on the right one β€” sand, salt, iron, copper, oil, lithium, and the US subtitle says so on the cover.

The new book really is called Trade WorldNailed it

Sean: Trade World: The Ties That Bound Our Past and Could Unravel Our Future published September 3, 2026 β€” four days after we recorded.

The sycophancy numbers, recalled coldNailed it

Sean: Stanford's SycEval puts Claude's non-sycophantic rate at 42.6% and its regressive sycophancy at 18.3%, ranked Gemini 9.3%, ChatGPT 14.4%, Claude 18.3% β€” three numbers and an ordering, all correct.

The Planet Money t-shirt, down to the squirrelNailed it

Sean: Published December 2013 β€” cotton from Mississippi, spun in Indonesia, sewn in Colombia, and per Planet Money's own note, squirrel plus martini equals animal spirits.

China's shipbuilding, if anything understatedClose enough

Sean: The real CSIS statistic is one Chinese state shipbuilder, by tonnage, across all commercial vessels β€” and of 5,448 large vessels on order worldwide, China holds 3,419 while the US holds three.

The X algorithm weightsClose enough

Dan: Self-hedged, directionally right, and too gentle β€” the published weights are report βˆ’369, block βˆ’74, reply +13.5, retweet +1.0, like +0.5, which makes a block worth βˆ’148 likes.

Who actually owns CursorClose enough

Sean: SpaceX bought Anysphere for $60 billion in June 2026, and xAI merged into SpaceX in February β€” so Grok and Cursor share a roof, but SpaceX holds the deed.

Four million boomers retiring a yearClose enough

Sean: Real number, wrong verb β€” 4.1 million Americans turn 65 each year through 2027, while the actual retirement flow is roughly half that.

The Jones Act is not active right nowClose enough

Dan: Waived, not repealed β€” a 150-day waiver, the longest suspension of Jones Act restrictions ever, extended through November 15 and narrowed from 600-plus tariff codes to 237.

A twenty-or-thirty-million-dollar golf courseClose enough

Sean: Alabama's golf facility was approved at $26.1 million in February 2022, finished at $47.75 million, and opened in October 2024 β€” not the week we recorded.

What Ed Conway actually argues about peak shipWhiffed it

Sean: Conway lays out the constraints faithfully and then argues against the thesis, noting it has been posited before β€” the Emma Maersk carried 14,770 containers in 2006 and the MSC Irina now carries 24,346.

Nine grand to eighteen grand because of HouthisWhiffed it

Dan: Wrong crisis β€” at the July 2024 Red Sea peak Drewry's composite was $5,937 and the worst single lane hit $9,612, while rates above $18,000 belong to the 2021 crunch.

Ninety percent of the budget went to food and housingWhiffed it

Sean: BLS puts 1901 at 65.8% β€” food 42.5%, housing 23.3% β€” and you only reach ninety by adding clothing and fuel, which is not what was said.

Seven thousand boomers die per dayWhiffed it

Dan: That would be 2.55 million a year against 3.07 million total US deaths at every age; the real figure is 3,000 to 4,000 a day, and the cohort shrinks about 2% a year, not ten.

Predictions from this episode

PendingSean (Dan accepts the premise and argues the answer is more hulls, not bigger ones)

The decades-long improvement in container-shipping unit economics is over. With ships unable to get materially larger or more efficient, shipping costs stop absorbing other cost increases and begin to pass them through β€” changing the trajectory of international transport pricing from a disinflationary force to a neutral or inflationary one.

β€œover time, what that's helped is fight inflation and shipping costs… we've not seen this inflation in shipping prices and now we might start to experience that… it will just kind of change the trajectory of pricing, I think, for international transport." (Dan, the counter-move): "just because we can't make the bigger ships, we need to make more of the bigger ships.”
Said in: Artificial Laziness (Aug 30, 2026)Horizon: 3–5 years. Evaluable against real (inflation-adjusted) container freight rates and cost-per-TEU-mile trending flat-to-up across a full cycle rather than resuming their long secular decline, and against whether newbuild orders shift from larger vessels toward more vessels of existing size.Confidence: Load-bearing on mechanism, and unusually well-specified for a same-day reaction β€” but the premise is contested by the very source it came from. Conway's post grants the constraints (Suez's 400m limit, crane reach across the beam, draught rather than box count as the binding ceiling, insurers balking at concentrated cargo value) and then declines the conclusion, noting the thesis has been posited before: the *Emma Maersk* (2006) carried 14,770 TEU and the *MSC Irina* (2023) carries 24,346, 65% more. Scoring should treat "the plateau is real this time" as the live question. Pairs with `christmas-2025-supply-shock` and the show's standing chokepoint lane.
PendingDaniel (Sean endorses: "Exactly. That's right where my head's at with it too")

Token prices fall roughly 90% over the next three years, which is what makes disposable, personally-built software rational β€” you accept the technical debt because a future model will fix it on request.

β€œas token prices drop ninety percent over the next three years… you just push a button and say, fix the thing that I built two years ago 'cause it's all broken and then the model's gonna like, yeah, Captain, and probably do a good job.”
Said in: Artificial Laziness (Aug 30, 2026)Horizon: 2029-08-30. Evaluable against published per-token pricing, but only once the comparison basis is fixed.Target: Dec 31, 2029Confidence: Load-bearing for the whole deterministic-build argument, and the scoring hinges entirely on a distinction Dan does not draw on air. Against a fixed capability level, 90%/3yr is strikingly conservative β€” Epoch AI measures declines of 9x to 900x per year depending on benchmark (GPT-3-level MMLU went $60/M tokens in Nov 2021 to $0.07/M by Oct 2024; a16z puts the general rate near 10x/year). Against frontier flagship list prices it is aggressive β€” those have been roughly flat-to-rising since GPT-4 launched at $30/M input in March 2023. Score the fixed-capability reading as the charitable default, since the quote is about re-running a fix on two-year-old software, but record both. Pairs with Ep 49's `open-source-three-quarters-local`.
PendingSean (arguing Dan's new venture into existence; Dan supplies the customer-sizing rule)

AI transformation consulting is noisy and influencer-heavy now but grows into a durable niche, because the technology is not usable out of the box and someone has to translate it into a specific business. The defensible wedge is compliance and industry-specific documentation, and the natural customer is the business changing hands β€” inherited or newly acquired, with a small back office and an owner who wants to grow it three or four times.

Said in: Artificial Laziness (Aug 30, 2026)Horizon: 2–3 years. Evaluable against whether Polyiota and comparable practices are still operating with repeat engagements, and against whether the compliance/vertical-documentation wedge is what differentiates survivors from the generic automation shops.Confidence: Load-bearing, with a live experiment attached β€” Dan launches Polyiota in this episode, so his own outcome is the first data point. ⚠️ Direct tension with Dan's own `automation-agency-window-closing` (Ep 22, 2025-01-28): "being an automation agency is like printing money right now, but I don't think it lasts very long." Dan resolves it on air by positioning *against* the commodity version β€” "you can go hire a twenty four year old automation agency dude… he lives in Dubai and he's making stupid money. But I don't know that that thing is very lasting" β€” and selling process understanding rather than builds. Whether that distinction survives contact with the market is exactly what both entries now test from opposite directions. Score them together.
PendingDaniel (Sean corroborates with the post-COVID Sunbelt precedent and nominates the Great Lakes)

The most underrated change of the coming decade is domestic migration, not AI. Priced out of coastal metros and unbound by remote work, young people without inherited wealth relocate and make the next set of desirable places β€” with college towns, mid-sized cities and the Great Lakes as the likely beneficiaries.

β€œThe amount of change we're going to see over the next decade, everyone's like, AI. Change over the next decade. AI. I'm like change over the next decade. Migration?" (On the mechanism): "if the technology moved west, it was partly because all the people that needed to take a chance needed to be extremely stingy with every single penny… So they went west where their pennies went further. And now I'm like, where are the cities?" (The framing that carried the working title): "No, no, not why not? Where's next?”
Said in: Artificial Laziness (Aug 30, 2026)Horizon: 10 years, with readable proxies much sooner. Evaluable against Census domestic-migration flows and IRS county-to-county migration data β€” specifically net in-migration to Great Lakes metros and mid-sized college towns, and continued net domestic out-migration from California and the coastal Northeast.Confidence: Load-bearing thesis and the spine of the episode's back half, but deliberately non-specific: Dan names candidate regions (Great Lakes, Michigan, Kansas City, Columbia MO, northwest Arkansas) without committing to one, and flags his own dramatization β€” "I'm purposely being dramatic to paint the picture." Score the *direction and driver* (affordability-forced relocation reshaping which places matter), not the specific cities. The falsifier is a decade in which coastal metros re-concentrate young talent. Pairs with the show's `housing-divide` lane and with `boomer-hostage-system`.
PendingSean (supplies the aggregation-theory frame) and Daniel (supplies both halves of the outcome)

Declining university enrollment resolves as aggregation rather than as uniform decline β€” some institutions die outright while the anchors (flagships and strong regionals, Purdue and Mizzou named) not only survive but get bigger, taking share from the schools that close.

Said in: Artificial Laziness (Aug 30, 2026)Horizon: 5–10 years. Evaluable against IPEDS enrollment by institution: closures and consolidations at the bottom alongside enrollment *growth* at flagships and named anchors, rather than a flat proportional decline across the sector.Confidence: Load-bearing, cleanly falsifiable, and better-specified than the show's earlier enrollment calls because it predicts a *shape* rather than a level. Context at recording, which cuts both ways: national enrollment rose ~1% in fall 2025 (+187,000, NSC) with freshmen up 5.5% β€” so the sector-wide decline the hosts assume is not currently in the data β€” while the concentrated pain is real and matches the prediction's shape, with Penn State's commonwealth campuses down 12.4% and seven slated to close, and at least 16 nonprofit colleges closing in 2025. Also note the fact-check on Sean's Alabama golf-course aside in the same passage (approved $26.1M in 2022, finished at $47.75M, opened October 2024, not "this week"). Pairs with Ep 54's `trades-pendulum-overshoots` and the earlier enrollment entry, both of which assume the level; this one predicts the distribution.
PendingSean (Dan endorses and adds short socks to the list)

The millennial tattoo boom has peaked and is reversing β€” visible tattoos become less common among the cohorts coming up, the way other millennial markers have already turned over.

β€œI think that trend is now passed too. I think it's starting its reverse course… yeah, hardcore." (Dan): "Along with short socks.”
Said in: Artificial Laziness (Aug 30, 2026)Horizon: 3–5 years. Evaluable against Pew/Ipsos tattoo-prevalence surveys by age cohort, tattoo-parlor industry revenue (IBISWorld), and the growth rate of the tattoo-removal market.Confidence: Casual aside, logged because it is cheap to score and the show has a standing habit of calling cultural turns. Immediate context is anecdotal β€” Dan's "millions of dollars worth of tattoo value at this pool" observation from the prior summer, which is evidence of the *peak*, not the reversal. Score generously on direction.
PendingDan

Automation-agency businesses (selling AI workflow integration to mid-market clients) remain profitable for 2–3 years, then commoditize and collapse as clients in-source the capability.

β€œBeing an automation agency is like printing money right now, but I don't think it lasts very long.”
Said in: We're in an Age of Build (Jan 28, 2025)Horizon: 2027–2030Target: Dec 31, 2030Updated: Sep 7, 2026Confidence: Load-bearing
PendingSean (Dan supplies the architecture requirement and the lock-in argument)

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

Said in: Rolling Your Own Problems (Aug 9, 2026)Horizon: 3–5 years. Evaluable on whether the durable winners at the interface layer are provider-agnostic (Buzz's model-agnostic design, workgroove, HQ) or vendor-native (Claude Tag, Cowork, Copilot), and on whether users report an agent as their primary computing interface rather than one app among many.Updated: Sep 27, 2026 (3 revisits)Confidence: Load-bearing thesis, and the episode's most testable *product* claim because both hosts already have skin in it. Sean's own evidence cuts against the horizontal preference: Cowork's mobile release (2026-07-07) obsoleted several tools he had built precisely to stay provider-agnostic, and he says out loud that the vendor-native version is "pretty slick" and he may just use it. Dan holds the line anyway and names the risk from the other direction β€” "you do doing feature releases that just sort of bulldozed something that you had built." The named contenders at recording: Buzz (Block, announced 2026-07-21, Apache-2.0, agents as first-class channel members over Nostr, pre-1.0), workgroove.ai, HQ. Pairs with `automation-agency-window-closing` and Ep 53's `agent-ready-commerce` lane.