
Sep 8, 2026 · 3h 12m
Listen from 1:30:43
Listen at 1:30:43
AI agents move from dazzling demos to everyday infrastructure
Math Wars, GG Navier Stokes, Greg Brockman Joins | Dan Wright, Eric Seufert, Scott Wu, Sahir Jaggi, Andrew Borovsky, Harry Mellsop
The episode tests whether advances in reasoning and computer use can become durable businesses across software, healthcare, finance, and robotics.
- 1AI progress now spans mathematical research, computer-use agents, coding systems, and practical consumer applications.
- 2Cognition’s Devin illustrates the shift from viral prototype to background software agent, while enterprise adoption still depends on security and workflow fit.
- 3OpenAI, Forrest, Split, and Antioch show how agent-based systems could reshape medicine, credit, and physical automation.
Don't miss
Greg Brockman explains why OpenAI’s reported Navier–Stokes result matters as the creation of new mathematical knowledge, even before a consumer application exists.
The brief
The show opens with AI’s reported progress on Navier–Stokes and other difficult mathematics, then asks whether such breakthroughs primarily expand knowledge or create economic value.
Computer-use agents such as Astra move the discussion from benchmarks to ordinary tasks, while demos involving CAPTCHAs, games, 3D design, and personal assistants expose the gap between spectacle and usefulness.
Scott Wu explains how Cognition turned Devin from a viral prototype into a practical background agent, while warning that enterprise adoption depends on security, reviews, onboarding, and concrete workflows.
Greg Brockman argues that generating new mathematical knowledge matters even without immediate applications, then connects Astra, product discovery, image generation, and medical memory to broader consumer utility.
The later interviews widen the frame: Forrest automates fragmented healthcare operations, Split rethinks consumer underwriting, and Antioch tackles the sim-to-real problem in physical autonomy.
What was said on this episode
78 statements · 60 positive · 10 negative · 1 mixed · 7 neutral
Proving Navier–Stokes is unlikely to measurably advance engineering.
“proving these equations is unlikely to actually have any measurable impact”
Listen at 7:22
Jordi Hays predicted the Millennium Prize problems would soon be solved.
“the Millennium Prizes will get solved or proved or, you know, delivered. Probably this year, if not like this week”
Listen at 11:05
Jordi Hays predicted AI discussion would shift back toward biotech after mathematical breakthroughs.
“the conversation will go back to Biden. biotech after we basically solve math”
Listen at 12:02
GLP-1 drugs are key interventions against diabetes and obesity.
“GLP-1s have been like a really, really key intervention in helping them fight diabetes, fight obesity”
Listen at 13:58
Agent-mediated communication may become popular and reduce spam friction.
“that was probably a glimpse into the future and like a cause for optimism”
Listen at 22:38
Agent-mediated communication may become common and reduce spam friction.
“that was probably a glimpse into the future”
Listen at 22:38
AI-generated games will substantially expand the market’s long tail.
“the long tail is going to get a lot longer”
Listen at 24:49
Lower creation costs make smaller-market products economically viable.
“with lower cost, you can go after lower TAM opportunities”
Listen at 26:06
John Coogan predicted The Social Reckoning would flop.
“I think this one's going to flop for sure.”
Listen at 38:01
AI-adjacent professional roles added roughly 730,000 jobs above trend.
“These roles have added roughly 730,000 jobs above the trend in recent years.”
Listen at 47:51
AI agents should be able to bypass some restaurant-booking platforms.
“an AI agent should be able to disintermediate those platforms”
Listen at 56:42
Uber is harder for AI agents to disintermediate than reservation platforms.
“Uber is much harder to disintermediate.”
Listen at 57:53
Apple’s rumored folding phone is expected to sell well.
“We're getting a folding phone probably this week. It's going to be exciting. I think it'll sell well.”
Listen at 58:32
Ms. Rachel’s second Netflix season performed substantially worse than the first.
“even with Ms. Rachel, like season two performed substantially worse than season one”
Listen at 1:20:54
Apple’s App Store monetization expansion will consist of advertising.
“No, it can't be anything but ads.”
Listen at 1:26:04
Apple has no alternative to monetizing App Store engagement with ads.
“I think they have no other option.”
Listen at 1:29:11
Apple will likely place ads in third-party apps and websites.
“I think they will. I think they will. I don't know what else they could do. I mean, I think they're going to do that.”
Listen at 1:30:43
Apple will place ads in third-party apps and websites.
“I think they will. I don't know what else they could do. I mean, I think they're going to do that.”
Listen at 1:30:43
Ads delivered directly to AI agents are unlikely to succeed.
“I just don't think ads to agents is ever going to take off.”
Listen at 1:32:57
OpenAI’s advertising growth will accelerate after conversion-optimized bidding launches.
“once you are bidding against a specific outcome then the growth in flex and i think you know they'll be off to the races”
Listen at 1:35:56
Meta AI now supports managing and optimizing advertising campaigns.
“Meta AI now does that. They've added that as a whole feature set.”
Listen at 1:39:05
Meta AI ad optimization could produce major secondary revenue effects.
“this could have dramatic second order effects too”
Listen at 1:41:11
Eric says direct-to-consumer commerce would not exist without Meta’s advertising platform.
“D2C would not exist absent Meta”
Listen at 1:45:08
Affiliate links are viable at small scale but not optimal for monetizing AI attention.
“affiliate makes sense. But this is the argument I made when ChatGPT introduced Instant Checkout. That's not the optimal way to monetize that attention.”
Listen at 1:47:23
Eric recommends conversion-optimized advertising over affiliate monetization at scale.
“The optimal way is to do conversion-optimized. uh advertising”
Listen at 1:47:32
Scott says current agents are capable enough to handle delegated tasks end-to-end.
“the capabilities are just so good that obviously you should go and delegate to something that just does the task entirely”
Listen at 1:52:33
- Dan Wrighton AI model routingPositive1:55:06
Scott says routing tasks across specialized models outperforms relying on one model.
“by combining the different models, as long as you know which use case to route to each model, then you can do a lot better than anyone alone”
Listen at 1:55:06
- Dan Wrighton AI model routingPositive1:56:11
Scott predicts model selection and routing will become increasingly important.
“I think it's sticking around, yeah. I mean, I think if anything, it's going to be a bigger thing as time goes on”
Listen at 1:56:11
- Dan Wrighton Coding agentsPositive1:58:30
Scott says coding agents are improving software quality.
“short answer is definitely yes, it is better”
Listen at 1:58:30
- Dan Wrighton Navier–Stokes problemPositive2:03:09
Scott claims AI has solved the Navier–Stokes problem.
“we just solved Navier Stokes with AI”
Listen at 2:03:09
Scott predicts the Riemann hypothesis will not be solved by end-2026.
“the rebound hypothesis will not be solved by the end of 2026”
Listen at 2:03:26
Scott predicts the Riemann hypothesis will be solved by 2027.
“if not 26, it will be solved in 27”
Listen at 2:03:33
Cognition prioritizes real-world applications over fundamental mathematical research.
“this fundamental basic science research is much less our focus as opposed to going and doing real-world use cases”
Listen at 2:07:21
AI models can generate new knowledge useful for solving difficult problems.
“we can actually generate new knowledge that we can learn from these models”
Listen at 2:12:15
Advanced AI makes disease cures and new medicines more attainable.
“talking about curing diseases and new medicines we're going to be able to develop, all of that starts to become much more real”
Listen at 2:13:24
AI assistants should proactively explain how to unlock additional capabilities.
“it should also be able to proactively say to you, hey, actually, if you ask me this other way, or if you added this connector, or if you authorize me to do this, or if you hook up your credentials in this way, I can go and do this other thing for you.”
Listen at 2:20:57
A unified healthcare platform could improve information sharing across providers.
“good sharing of that information across different providers, that becomes possible if everyone's on one platform.”
Listen at 2:28:57
Clinical-trial enrollment is a major bottleneck in drug development.
“clinical trial enrollment. That's a huge bottleneck to drug development”
Listen at 2:29:03
OpenAI aims to build the world’s best healthcare platform.
“what I view us as building is really trying to build the world's best healthcare platform”
Listen at 2:29:21
AI healthcare assistance is among AI’s most positive applications.
“it's one of the most positive applications of AI that you can think of.”
Listen at 2:30:06
John says AI memory may identify medical patterns humans take years to detect.
“it may be able to like pick up patterns that sometimes would take a human years to figure out”
Listen at 2:30:32
AI will generate potential medicines faster and for more diseases than before
“that is going to work. Yeah. We're going to generate more potential medicines for more diseases faster than we ever have before.”
Listen at 2:37:09
Turning a molecule into an approved medicine takes over a decade and billions of dollars
“it takes more than a decade and billions of dollars to turn a new molecule into an approved medicine.”
Listen at 2:37:17
Medicine will remain a permanent and increasingly important investment industry
“medicine is going to become this, like it's a permanent industry to invest in is increasingly become the most important thing that's happening in advanced society.”
Listen at 2:40:29
Healthcare lacks a standard case because workflows vary widely
“there's no standard case in the system.”
Listen at 2:42:43
SplitPay has one million users
“we now have a million people using us”
Listen at 2:45:39
- Dan Wrighton Split cash-flow underwriting modelPositive2:46:52
Split has the country’s best cash-flow underwriting model
“I think we're the best cash flow underwriting. model in the country today.”
Listen at 2:46:52
FICO is outdated, especially for younger borrowers
“what we find and we believe is that it's outdated.”
Listen at 2:47:58
ACH is ten times larger than a credit-card network and the largest US payment network
“ACH, right, is 10 times bigger than a credit card network. It's the biggest payment network in the country.”
Listen at 2:49:53
ACH is ten times larger than a credit-card network.
“ACH, right, is 10 times bigger than a credit card network.”
Listen at 2:49:53
- Dan Wrighton SplitPositive2:51:19
Split grew fivefold within five months
“within five months, we were 5x bigger.”
Listen at 2:51:19
Physical AI technology is heading toward world models
“We think that's where the puck is headed”
Listen at 2:56:08
Tesla and Waymo already have technology close to full self-driving
“we see close to full self-driving already yeah with you know, Tesla and Waymo and all those things. So the technology actually exists”
Listen at 2:56:56
Full autonomous driving will arrive before 2050
“I don't think it's 2050, right?”
Listen at 2:57:28
Autonomous driving will arrive before 2050
“I don't think it's 2050”
Listen at 2:57:28
- Dan Wrighton Tesla, Waymo, and Wave autonomous driving systemsPositive2:57:30
Tesla, Waymo, and Wave have systems that work extremely well
“we're seeing Tesla and Waymo and also companies like Wave actually now doing these requirements that work extremely well”
Listen at 2:57:30
- Dan Wrighton autonomous-driving systemsPositive2:57:30
Tesla, Waymo, and Wave have autonomous-driving systems working extremely well
“we're seeing Tesla and Waymo and also companies like Wave actually now doing these requirements that work extremely well.”
Listen at 2:57:30
Tesla, Waymo, and similar companies have built powerful data flywheels
“these companies have built this incredible data accumulation flywheel”
Listen at 2:57:40
- Dan Wrighton Tesla data-collection modelPositive2:57:44
Tesla car buyers help fund and supply data for Tesla system training
“Tesla's is particularly brilliant because if you buy one of those cars, you're essentially paying to help them train the system and get kind of more and more data into that engine”
Listen at 2:57:44
- Dan Wrighton autonomous driving timelinesNeutral2:58:14
Autonomous-driving timelines depend on how quickly companies build data flywheels
“that timeline is going to be purely a question of how quickly can you build that flywheel?”
Listen at 2:58:14
Robotics must become more sample-efficient to bootstrap deployment
“you need to figure out a way to be much more sample efficient than we are today.”
Listen at 2:58:35
Robotics systems need substantially greater sample efficiency to scale
“you need to figure out a way to be much more sample efficient than we are today”
Listen at 2:58:35
Antioch is betting on a hybrid simulation approach
“that's our hybrid simulation approach”
Listen at 2:58:41
- Dan Wrighton Hybrid robotics simulationPositive2:58:52
Hybrid simulation can accelerate deployments and create a virtuous data flywheel.
“You can scale your deployments faster and it becomes this kind of virtuous flywheel effectively.”
Listen at 2:58:52
Antioch's present market includes tens of thousands of companies
“even present day, it's massive. Like we're talking tens of thousands of companies”
Listen at 3:00:11
The physical-AI market already includes tens of thousands of companies
“even present day, it's massive. Like we're talking tens of thousands of companies.”
Listen at 3:00:11
- Dan Wrighton Antioch target systemsNeutral3:00:31
Antioch targets systems whose real-world testing is difficult and expensive
“it's any system that has this. hardware software machine learning component where testing in the real world is really difficult and really expensive”
Listen at 3:00:31
Many product categories will benefit from robotic components
“everything from a robotic vacuum cleaner to a robotic drone, camera drone, sports. There's just going to be robotic pieces. Even if it's not a robot, a humanoid in every category, there will be something that benefits from this.”
Listen at 3:00:50
- Dan Wrighton AI back-pain wearablePositive3:04:18
The wearable could help many more people with back pain
“really kind of like bring this out to the world and hopefully help Many, many more people.”
Listen at 3:04:18
Current AI models enable building almost any desired health software
“we can build pretty much whatever we want on the software side”
Listen at 3:05:19
Data collection is the key constraint for AI health decisions
“it's the data collection that really rules.”
Listen at 3:05:23
Body patches can collect data from the rest of the body
“we can cover the rest of the body”
Listen at 3:05:49
The back-pain wearable is targeted to ship in January
“I think we're aiming to ship January, the start of January.”
Listen at 3:07:31
- Dan Wrighton AI back-pain wearablePositive3:07:48
The wearable aims to be affordable while remaining viable
“I'm trying everything I can to charge less, basically, and still make it work for people”
Listen at 3:07:48
ChatGPT Images 2.5 produces high-fidelity zoomed-in text
“some of the zoomed in text is really, really good. It's pretty high fidelity, I think. We're getting close.”
Listen at 3:08:51
Modern AI may be able to solve the Waldo benchmark
“I think it's possible with modern AI.”
Listen at 3:09:09
The new image model produces more consistent animations.
“So it's much more consistent. That's like one of the big consistency, big improvements.”
Listen at 3:10:11
The new image model substantially improves animation consistency
“So it's much more consistent. That's like one of the big consistency, big improvements.”
Listen at 3:10:11
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.
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.