
Sep 22, 2026 · 2h 15m
Listen from 1:19:36
Listen at 1:19:36
AI agents collide with margins, logistics and human judgment
Harvey's Margin Whiplash, Human Powered Agents, McLaren Rebrand, Insects > Humans? | Ben Thompson, Gagan Biyani, Alex Ratner, Peter Kalogiannis, Daniel Petkevich
The episode tests whether AI’s next phase will be defined by cheaper software, human-supported services, or new physical infrastructure.
- 1AI companies must rethink pricing as heavier model usage can turn seat-based software into a margin problem.
- 2Agentic commerce depends on logistics, personal context and human support, while consumers may still value friction and choice.
- 3Model progress remains tied to specialized data, expert judgment and physical experimentation rather than synthetic scale alone.
Don't miss
Daniel Petkevich connects Tesla’s ambitious Roadster concept to the practical engineering limits of his own legged robot project.
The brief
Ben Thompson’s discussion of Harvey’s negative gross margins frames the episode’s central business problem: AI can make software dramatically more useful while making each customer more expensive to serve.
The hosts examine human-supported agents, sponsored recommendations and Amazon’s logistics advantage, asking whether agentic commerce will remove shopping friction or simply move it behind the scenes.
A debate over insect welfare and a humorous AI-bioweapon scenario widen the lens, contrasting abstract scale-based reasoning with the practical consequences of increasingly capable systems.
Gagan Biyani presents a one-year school for builders, while Alex Ratner argues that synthetic data cannot eliminate the need for real-world input and human judgment in model improvement.
Daniel Petkevich closes the arc with a Tesla Roadster thought experiment and an open-source quadruped robot, showing how AI ambition eventually runs into actuators, torque and physical constraints.
What was said on this episode
67 statements · 40 positive · 11 negative · 2 mixed · 14 neutral
Harvey has returned to positive gross margins after its June decline.
“The good news is that they're already back to positive gross margins”
Listen at 3:53
Harvey has returned to positive gross margins after its June decline.
“The good news is that they're already back to positive gross margins”
Listen at 3:53
A mixture of proprietary and public frontier models will likely be optimal for Harvey.
“Harvey is going to make another bet. Probably some mixture of both is going to be the right approach.”
Listen at 6:07
A mixture of proprietary and public frontier models will likely be optimal.
“Probably some mixture of both is going to be the right approach.”
Listen at 6:10
AI token costs are declining continuously.
“tokens are getting cheaper by the day”
Listen at 8:34
Jordi estimates his humorous AI-doom scenario has a 10% probability.
“I put it at 10%.”
Listen at 26:01
John estimates AI doom risk at 10% absent safety action.
“10% if we don't do anything around safety.”
Listen at 26:07
Meta's human-in-the-loop approach primarily collects training data, not a permanent solution.
“I would view this human in the loop thing more as them doing data collection and creating more training data for them than a permanent solution to the product problems.”
Listen at 33:13
Meta’s human-in-the-loop agents primarily collect training data rather than permanently solve product limitations.
“I would view this human in the loop thing more as them doing data collection and creating more training data for them than a permanent solution to the product problems.”
Listen at 33:13
Service and commerce apps will need to decide whether to open APIs to consumer agents.
“Every app that's a service marketplace or commerce app will need to existentially decide to open APIs for consumer agents to interact.”
Listen at 38:22
Commerce and marketplace apps will need to decide whether to open APIs to consumer agents.
“Every app that's a service marketplace or commerce app will need to existentially decide to open APIs for consumer agents to interact.”
Listen at 38:22
Commerce demand will shift toward alternatives when Amazon blocks agent access.
“The water will flow into the cracks.”
Listen at 44:11
McLaren’s SUV, manual supercar, and hypercar bets could address valuable market segments.
“The steel man here is that they are taking a number of bets that... seem to speak to interesting, potentially valuable pieces of the market.”
Listen at 47:53
McLaren's SUV, manual supercar, and hypercars could satisfy existing market demand.
“You have the trappings of at least satisfying the demand that exists in the market.”
Listen at 49:40
Adult-only household toy sales grew 16% through June, exceeding households with children.
“Toy sales among adult-only households have outpaced households with children for toy sales, growing 16% through June.”
Listen at 50:30
Adult-only household toy sales growth is bullish for the supercar market.
“the fact that it grew 16% in June... Sign that adults are having fun. They're buying toys. And bullish for the supercar community”
Listen at 51:02
People working from home should work outside the house when family interruptions matter.
“the answer if you can swing it is to get out of the house”
Listen at 59:45
People working from home should work outside the house when uninterrupted focus is necessary.
“the answer if you can swing it is to get out of the house.”
Listen at 59:45
An effective personal assistant can substantially increase a person’s freedom.
“a personal assistant that works and you learn how to use one, it is tremendously freeing.”
Listen at 1:03:10
A well-used personal assistant can substantially free its user.
“it is tremendously freeing”
Listen at 1:03:13
Humanoid household robots will eventually become widely available in homes.
“the humanoid robot that hopefully we'll all have in our houses in the long run.”
Listen at 1:03:23
Bill defines AGI as AI that genuinely learns.
“My personal what is AGI rule is AI that actually learns.”
Listen at 1:17:52
Personal agents should be framed as removing annoyances rather than increasing productivity.
“we're not helping you be more productive. We're helping you get rid of annoying crap in your life.”
Listen at 1:19:14
Personal agents should be framed as removing annoyances rather than increasing productivity.
“we're helping you get rid of annoying crap in your life.”
Listen at 1:19:17
Personal agents will gain adoption through demonstrations on Instagram Reels or TikTok.
“The way I think this is going to land is through like Instagram reels or through TikTok.”
Listen at 1:19:36
Useful autonomous vehicles will eventually include humanoids that perform real-world tasks.
“the waymo or the tesla with the humanoid in it”
Listen at 1:23:22
Tesla self-driving improved substantially over the previous year.
“it's so much better this year than it was even a year ago”
Listen at 1:24:10
Tesla’s self-driving system has improved substantially over the past year.
“it's so much better this year than it was even a year ago”
Listen at 1:24:10
Consumers may prefer Amazon because its fulfillment is reliable and predictable.
“my personal revealed preference is I would almost always want it from Amazon because I trust it's going to get there.”
Listen at 1:27:34
Amazon is preferred because its deliveries are trusted to arrive reliably.
“I would almost always want it from Amazon because I trust it's going to get there.”
Listen at 1:27:36
Amazon will ultimately prevail in the agentic-commerce dispute.
“Amazon is going to win this dispute or it's going to be resolved in their favor in the end.”
Listen at 1:31:22
Amazon will ultimately prevail in the agentic-commerce dispute.
“Amazon is going to win this dispute or it's going to be resolved in their favor in the end.”
Listen at 1:31:22
Consumers enjoy shopping as a form of entertainment.
“people love shopping”
Listen at 1:32:05
Consumers often view shopping as entertainment rather than merely a productivity task.
“people love shopping. They don't want to be productive. They want to be entertained.”
Listen at 1:32:36
Complaining during shopping does not necessarily indicate consumer dissatisfaction.
“Just because someone's complaining doesn't mean they're unhappy.”
Listen at 1:33:50
Consumer complaints do not necessarily indicate dissatisfaction.
“Just because someone's complaining doesn't mean they're unhappy.”
Listen at 1:33:50
The academy’s in-person San Francisco model will immerse students in Silicon Valley and build networks.
“it's about getting right in the mix of Silicon Valley, understanding the culture, getting immersed in the in the people and building a great network.”
Listen at 1:39:23
A for-profit education model pressures the academy to deliver value to students.
“one of the great things about the for profit model is it forces you to deliver or die”
Listen at 1:41:52
A for-profit education model pressures schools to deliver value to students.
“the for profit model is it forces you to deliver or die”
Listen at 1:41:53
Students can pursue multiple career paths and still achieve successful lives.
“there are many different options for their careers. there's many different paths to a successful life.”
Listen at 1:43:02
People can pursue many different career paths to achieve a successful life.
“there's many different paths to a successful life.”
Listen at 1:43:05
The academy's first class will begin in fall 2027.
“First class in fall, fall 2027.”
Listen at 1:44:52
- Alex Ratneron AI model training dataPositive1:46:43
Early-stage AI models benefit from increasing data volume.
“when you're early on in the learning curve, the model doesn't know much, so any bit of information is additive.”
Listen at 1:46:43
Alex rejects the sci-fi scenario where recursive self-improvement eliminates human economic contribution.
“I'm short the more kind of sci-fi, you know, accelerate and be done with humanity's contribution to anything economic kind of version of RSI.”
Listen at 1:48:14
Recursive AI improvement requires real-world input in addition to model-training changes.
“you need real input from the real world.”
Listen at 1:50:48
- Alex Ratneron Specialized AI training dataMixed1:51:09
As AI accuracy improves, marginal data becomes more valuable and harder to produce.
“the economic value of every kind of additional decimal point of improvement goes way up because of the value. And so does the difficulty of creating that data.”
Listen at 1:51:09
Swarm Aero is building capacity to produce hundreds of aircraft annually.
“we've also been building the factory, the capacity to... to really build these hundreds a year.”
Listen at 1:55:09
- Peter Kalogiannison Swarm Aero Gamera aircraftPositive1:55:25
Swarm Aero’s command-and-control system is intended to operate hundreds of aircraft simultaneously.
“building out that autonomy and the command and control stack that allows us to have truly a swarm of these, hundreds of these operated all at once.”
Listen at 1:55:25
- Peter Kalogiannison Aircraft remote-pilot workforceNegative1:56:00
There are insufficient personnel with skills to remotely pilot hundreds of aircraft.
“There just aren't enough folks out there who have those skill sets.”
Listen at 1:56:00
- Peter Kalogiannison Swarm Aero Gamera aircraftNeutral1:56:57
Swarm Aero expects the Air Force to pay less than $10 million per aircraft.
“we're confident the Air Force is asking for less than $10 million per. They're not going to pay more.”
Listen at 1:56:57
- Peter Kalogiannison Swarm Aero Gamera aircraftNeutral1:57:25
Gamera is primarily designed for Pacific-theater range and payload requirements.
“it really is the Pacific. It's that range. It's carrying the kinds of payloads that the. The big enough ones that make a difference in a potential conflict there. Really, it's focused on the Pacific”
Listen at 1:57:25
- Peter Kalogiannison Gamera aircraftNeutral1:57:25
The Pacific theater is the primary operational focus for Gamera.
“it really is the Pacific”
Listen at 1:57:25
The Tesla Roadster will use four rear thrusters to generate downforce.
“It's going to have four thrusters in the back that are going to generate downforce.”
Listen at 2:03:29
The Tesla Roadster will accelerate from zero to sixty in 0.9 seconds.
“No, 0.9 seconds.”
Listen at 2:03:40
“No, 0.9 seconds.”
Listen at 2:03:40
The Tesla Roadster’s airspace closure is probably a public-relations stunt.
“I think that part is a PR stunt.”
Listen at 2:04:38
The Tesla Roadster's airspace closure is likely a publicity stunt.
“I think that part is a PR stunt.”
Listen at 2:04:38
Elon Musk intends the Roadster to be a commercially purchasable car.
“I think he actually wants to make a car people can buy”
Listen at 2:04:40
Elon Musk intends the Roadster to be a purchasable car rather than merely a stunt.
“I think he actually wants to make a car people can buy”
Listen at 2:04:40
The Tesla Roadster is primarily a fun project rather than a scalable business.
“This is like a fun toy project.”
Listen at 2:07:25
People should buy the Tesla Roadster intending to resell it at a higher value.
“everyone should just try to buy one and resell it.”
Listen at 2:09:43
People should buy the limited Tesla Roadster and resell it for potential appreciation.
“everyone should just try to buy one and resell it.”
Listen at 2:09:43
Buyers should not leave Tesla Roadster deposits tied up for ten years.
“under no circumstances, tie your money up in this deposit for 10 years.”
Listen at 2:11:04
Quadruped robots are an inefficient design compared with alternatives.
“Quadruped is a really suboptimal design for a robot.”
Listen at 2:11:55
Quadruped robots are a suboptimal general-purpose robot design.
“Quadruped is a really suboptimal design for a robot.”
Listen at 2:11:55
Electric-motor-based actuators are poor at generating torque for robotic limbs.
“they're really bad at generating torque”
Listen at 2:12:16
Electric-motor-based actuators are poor at generating the torque needed for legged robots.
“they're really bad at generating torque.”
Listen at 2:12:16
Statements are attributed to the speaker as said on the episode and reflect their view at the time, not PodLume's. They are not advice.
Books & mentions
Weekly email
State of the Pod, every Monday
The week's most-discussed names and stories from podcasts, with the moments behind them. Free, five minutes, unsubscribe in one click.
