You're watching TBPN. Today is Monday, September 8th, 2026. We are live from the TBPN Ultradome, the temple of technology, the fortress of finance, the capital.
You know what time it is, John?
Ramp.com time. Time is money. Save both. Easiest corporate cards, bill pay, accounting, and a whole lot more, all in one place.
I need my countdown. There it is. Road to Christmas: 107 days.
December 25th flying by today.
107 days. Time is flying till Christmas. The road to Christmas is going strong.
Start scoping out Christmas trees.
Hopefully you've already planned.
Open season. Does math matter? That's the big question. I was debating this with Tyler today. Lots of tension on the timeline over Navier-Stokes, the Millennium Prize math problem. There has been a Millennium Prize math problem that was solved before, the Poincaré conjecture in the pre-AI era. These are very, very difficult math problems that all PhD mathematicians, all sorts of elite—the upper echelon of math has been grappling with for years, and AI is starting to knock these down. Remember, last year was the year of the IMO gold medals from both Google DeepMind and OpenAI. We talked to Scott Wu about this, who's coming on the show later today. Huge race for Cognition. And Scott made the prediction on TBPN early in 2025 that he believed that artificial intelligence would be able to achieve gold at the IMO, the International Math Olympiad, in 2025. And he was right. He nailed it.
Progress continued, but debate remained around 'cool calculator, bro' as the critique. People just don't feel empowered by a system that's really good at math. A TI-83 can do more complex math than most people can do. And so, when you say, 'Oh, there's another level of math that you don't know about and you don't care about, and computers are good at it,' a lot of people's eyes glaze over. And so, the bigger question here for me is all about public perception. What actually matters here? I think they're cool benchmarks. I think they're interesting stories. Obviously, there's a lot of drama over this race. It feels like it was neck and neck.
Now, there's a bunch of debates going back and forth with a mathematician who works at Google and a mathematician who's independent who worked alongside an Anthropic researcher who were both working on it. And they were apparently texting back and forth and going back and forth on did they steal the data, did they copy each other's, do they have different approaches? And so lots of different back and forth. And I think someone summed it up well by saying that these math models have discovered the hardest problem of advanced mathematics, which is authorship contribution: who actually did the work, who gets lead left on the paper?
There was drama last year, 2025, during the IMO. OpenAI and Google DeepMind and I believe Harmonic, the math AI lab that we had on the show last week, were all working on solving the International Math Olympiad. Remember, there's six questions. Question six, no AI model could solve. It was more complex, but both teams got a 35 out of 42 even though they only got five questions out of six right. That's good enough for gold. It's the bare minimum, but they did it. But there was a bunch of back and forth because OpenAI validated with former IMO gold medalists, while Google's result was graded by the official rubric that year, which is internal to the organization. We called it at the time: it's like they ran the fastest 100-meter sprint in the parking lot while the actual Olympics is going on in the stadium. And so, a lot of these are just vibe wars races. Does it even matter?
The interesting thing about Navier-Stokes is, does it matter, Tyler? You think it does? I'll debate you on this a little bit. It seems like it doesn't because the equations themselves are known; they're just not fully proven. Engineers and physicists already solve them numerically all the time for particular situations. So airflow around a wing, you might have to use some of these equations. Weather models, water through a pipe—these are critical things. They are boring, but you can imagine that there would be cause for optimism. Better weather models save lives during hurricane season. Increased airflow across an airplane wing could lead to more efficiency, longer ranges, cheaper flights, lower carbon emissions. There's a bunch of reasons why you could solve all that and be excited. But that's not actually what—
One of the reasons I'm excited—I believe we agreed last week that if anyone solved Navier-Stokes, Tyler would shave his head.
Oh, that's right. Didn't we do that? We said something.
No, no, that was about the details. We have details down. Get ready to shave your head. But yeah, this is a big moment.
Yeah, I think it was actually if any AI progress happens at all, Tyler shaves his head.
Yeah, yeah. But do you have a steel man for why this matters? So I think I agree that I don't think this new stuff that's going on is practically super useful. I don't think you're going to feel it when you're on an airplane. You're not going to feel less turbulence now because of this.
But these are the most important open math problems. These are more close to a beautiful painting or something than some practical engineering breakthrough.
Yeah. This is pure math. This is the peak of the mountain.
Same month that AI solved this incredibly difficult math problem, I saw someone used ChatGPT to book a haircut, so equivalent. No, I mean in terms of the actual economy, the actual impact of these things, practically using these tools for something that people enjoy that's not just spinning their wheels, that's not just optimizing endless things, that's actually generative and moving things forward, not just defending against cyber attacks.
Winsome says, 'If he gets Millennium Prize, you all have to get bowl cuts.'
Oh, bowl cuts. Bowl cuts are interesting. That's really, really ridiculous. You do get a million bucks, so there's that. That's not nothing, but in the AGI hype race where trillions are being deployed, it is very minimal.
It sounds like cause for optimism that Navier-Stokes would improve airflow over a wing, something like that, but that's not really how it will work. Proving these equations is unlikely to actually have any measurable impact, and that's sort of agreed upon. Physicists have chimed in with this take that this won't actually move the discipline of engineering forward at all. But it is a cool demo. Engineers already solve the equations numerically, as I mentioned. So clearly it matters for hype and vibes. Solving harder math problems is a good heuristic to show how quickly models are progressing, but—
Yeah, and to me it just comes down to humans are able to use machines that humans built to further humanity's general knowledge.
Yeah, it's good. But I feel like if you're just trying to measure AI progress, benchmarks and raw inputs and raw metrics of the model sort of do the same thing. The number of people who will be convinced that AI is progressing exponentially because they see this result, and then you look at the number of people that would be equally convinced of AI progress just by hearing, 'Oh, there's 10 gigawatts coming online,' or 'The new model has 10 trillion parameters,' or 'It was trained with a billion dollars of compute,' or something like that—
The thing is, the average person is probably more impressed by Sora than by this.
Totally. It's way more visceral to actually see something. And that's why the weekend was very much Blender and people showing 3D renders and remodeling their houses and building small games and web games. And that feels much more like a breakthrough. People are having fun. They're vibe coding stuff, but they're also doing 3D modeling, which is a new skill. We've had image generation, video generation, but there's something much more grounded in watching, from just a few images, a Blender model come to life. You can walk around inside of it, and it can be something that the model doesn't exist for.
I saw a few demos that were a train, and it was like, 'Oh wow, it modeled this amazing train.' But there are perfectly modeled trains out there that you can just go grab from the internet. That's not that impressive. When you actually see your specific house or your car or something that doesn't already exist out there perfectly modeled, you're like, 'Wow.'
Sorry to interrupt. John Anderson says, 'Why are we only seeing the calendar days for Road to Christmas? Can we please get a weekdays and workdays breakdown for Die Hard?' Yeah, we need a full screen graphic, so let's work on that. Thank you.
Good. Thank you, John. Good comment. Good comment. In terms of feel the AGI, computer use definitely one of these. Astra particularly faster at computer use for really, really mundane things. Also just interesting to have a very different interface to your computer, like configurations, all the things where you just have to dig through a bunch of settings, maybe set something on the command line, maybe find a control panel. You can just ask Codex or any AI agent to go and do it, and it'll just do it and come back to you.
One of the use cases that I'm excited about is, you know when you get a parking ticket and then you're like, 'I got this parking ticket, I want to deal with it,' and then it's like, 'Well, the ticket is not in the portal yet.' Being able to go into chat and say, 'Hey, once this ticket is live in the system, please pay it.'
And so you can just forget about it.
Parking ticket superintelligence.
It feels like we're close.
So yeah, I think that the Millennium Prizes will get solved or proved or delivered probably this year, if not this week. It feels like all of the different labs are like, 'Put a million dollars of inference on this, put a million dollars of inference on this.' It's going to be a back-and-forth, and the labs are going to very quickly solve basically all of these problems. And then there'll be a search for new problems. But the conversation will move back to curing cancer almost certainly, because that's the one that's so grounded. That's the one that's supposed to be tractable. It's been messaged from various labs for so long that people will want that.
But there will be a longer flywheel from actually designing a cure for a specific cancer, running tests, seeing that it was effective. You can't do it with a million dollars of inference over a weekend, I think. But my prediction is the conversation will go back to biotech after we basically solve math, which is a crazy thing to say, but it feels like we're there. What do you think?
Yeah. People say loudly, stop.
So one, the pharmaceutical industry has been curing forms of cancer or making progress against various forms of cancer for a very long time, and gets zero credit. Worst reputation of any category of businesses on the planet. And so I don't think that's the solution to sentiment around the technology industry.
Yeah. Setting up this idea, imagine how underwhelming it will be if any AI lab comes out and makes 20% progress towards one specific type of cancer, right? That's not going to come across as, 'Oh wow, they cured cancer.' It's like, okay, they sort of advanced understanding and maybe the treatment process. And this type of thing is happening all the time. Anybody who's had a family member that's suffered from cancer has heard about, 'Hey, there's this thing in trials, it's showing promising results, trying to get you into it' or whatever. So that's happening all the time, right? And it's not going to suddenly change how people feel about the technology industry broadly.
No, it'll diffuse slowly and people will be back in the 'What have you done for me lately?' I just saw a reel where basically the thrust was, 'What has the pharmaceutical industry done for us lately?' And I was like, GLP-1s sort of cured obesity. It's sort of a big deal. And for a lot of people, GLP-1s have been a really, really key intervention in helping them fight diabetes, fight obesity, fight a whole bunch of knock-on effects that come from that. But it's not so much as the speeding train was coming towards them and the industry pulled them off the tracks at the last second. It's something where if you're living at a healthier weight from your 30s to 60s, you will probably be healthier at 60, but you're not seeing that it saved a life today. A GLP-1 has yet to get that type of credit. So it's going to be back and forth forever.
Anyway, in the meantime, somebody gave Astra a paintbrush, a robot, and a camera, and asked it to paint the Golden Gate Bridge in real life. It figured out how to control the robot and progressively got better throughout its attempts. The time lapse is sick. Here's a one-minute time lapse of his account painting the Golden Gate Bridge. How would you rate this, Jordy?
It appears better than I could do.
Better than you could do. How long are you giving yourself?
I'm saying at least by the later attempts.
Okay, the later attempts, because it gets better. The first one was not quite there, in my opinion.
Yeah. Somebody was talking about potential for a 3D printer boom with all the 3D modeling. The idea that you can design—I saw someone else designed a system where you can, from a prompt, go from prompt to Lego kit basically in one go. So it takes your prompt, whatever you gave it—
I think you could do better with enough time. I don't know how long that took, but it is impressive. It's a fun demo, and again, it lives in the real world. It's not abstract at all, so you can just watch it work, watch it do the thing. It's very, very interesting.
A lot of times what's interesting is that when you're using these models, when they have to do a task like that, they will write a deterministic piece of software to execute it. Like I built a game where it was basically Tony Hawk, but you would play as a pelican on a bike and you'd be able to do tricks and backflips. And the game worked really well. It was fun. And it built it in Godot, the open-source framework for game development. And then I had it go and try and get the high score. And instead of actually using computer use to go and play it, it wrote a script that would play the game flawlessly. And it looked really mechanical because there was no variation in it at all, like watching a normal person play. But it did get a very high score.
So I could see it doing the same thing where it generates the image and then maps the X and Y axes and the plan into code and then runs that. And so there's this process of what the model can do. The model might also write software to do that same thing. So there's this iteration process there.
And of course, people are looking at BachBench, which is the skill in writing a piece of music, and it is the actual meme from I, Robot: 'Can a robot write a symphony? Can a robot take a blank canvas and turn it into a masterpiece?' I don't know if people are calling this a masterpiece, but they're pretty impressed with—
It was pretty good. I listened to this.
It was good. You're a fan of this LilyPond four-part chorale in the style of Bach.
One of the best reactions to Astra's release last week was Josh, who's over at Sentry—his company was acquired by Sentry. He said, 'After Astra's release yesterday scoring 99% on ARC-AGI-3, I think it's become clear that AGI is no longer some distant hypothetical. About 15 minutes after the benchmarks came out, I booked the first flight I could find to Jackson Hole.' And he goes on to say he's also looking into purchasing a horse. 'At this point, horses remain one of the only major transportation platforms with no API, no OTA updates, and no realistic path to MCP support.' So glad to see more people getting into horses.
Getting into the horse game. Bull market in the equestrian lifestyle.
Yeah, it seems like he was joking, but a lot of people were like this unironically: 'I want to touch grass, I suppose.' Did you see this? Bojan posted, 'It's over. Our last line of defense has fallen. Astra beat all 48 levels of the I'm Not a Robot game.'
And this was crazy to watch. I wasn't sure if this was sped up. It seems like it's maybe sped up a little bit.
You have to play tic-tac-toe, too. It's pretty good. What? Yeah, the CAPTCHA defeat is going to be a whole new level of problems. At the same time, it feels sort of remarkable that Pangram is as effective as it is given how advanced text models are. And so there's a world where it's indefensible. Was that WaldoBench?
It was finding Waldo, not creating—
Finding Waldo. How are we doing on WaldoBench? That's a good thing to test out for Astra.
Yeah, we got to see, because I bet that there's a way to have Astra generate the proper tiles. Because the problem when we—
Yeah, you can iterate on it more.
Exactly. And the problem is that when we try and use an AI image generator to generate a Where's Waldo, it always gets the scale wrong, I feel like, or there's bigger characters in the foreground.
Yeah, perspective is wrong.
The isometric display is what makes a Waldo. But there's probably ways to solve that with a little sprinkling in of deterministic code generation or Waldo generation. But WaldoBench might be saturated now, I don't know. It can't be really saturated; it's not a real benchmark. But lots of back and forth in the benchmark world. It does feel like we're getting to the end of it, and the results are either a big challenge like a math problem or just a cool demo like this.
Another good demo from somewhere on Astra. He says, 'Astra and I are working on making my contact form as difficult as possible. You actually have to fight a battle in order to contact them.'
That's amazing. I might try and contact them after this. This is pretty fun. I like this. The defeat of all of the CAPTCHAs is somewhat linked to the Instinct news. Did you see that this weekend, Jordy?
Yeah. Instinct is a personal assistant AI focused on consumer. You link it with your email, your text messages, and we've seen a lot of investors sharing posts about—
For what it's worth, I would be surprised if anyone in the audience hasn't already heard of Instinct.
Yeah, and we talked about it before.
Thank you for reintroducing it.
But I think it's worth explaining how it works because it does seem like it triggered some bans. There was one report of an individual who got banned from Resy because they were absolutely spamming to get reservations to restaurants.
And there was another example of somebody who was at the US Open, was on the fan cam, and then wanted footage.
As the user experience would have been absolutely incredible, right? You're sitting there and you're like, 'Hey, I appeared on the screen for a moment. Can you go and find this?' I've had that idea so many times, like, 'Oh, go find someone on the fan cam.'
What Instinct was doing in the background was being absolutely relentless, and it seems like it was reaching out to hundreds of people aggressively, following up, emailing the media contact. But they did wind up getting the video.
Yeah, it worked, and it was mutually beneficial. And you could imagine if the other side has agent intermediation and filtering and can process things that are spam-like, but there's actually a reasonable action to take that could make it through a filter and not actually annoy anyone on the other side. So that was probably a glimpse into the future and a cause for optimism. I could see that being popular. But people are doing all sorts of stuff. The classic example is go save money, get refunds, do that sort of stuff. But people are having fun with that.
So many demos. This 3D website that pulls apart the male anatomy into 2,234 modeled pieces. I shared this with a friend who I was trying to explain in one video what to get their wheels turning. I think this is a good example. I wonder how much of this was individually modeled versus pulled from an existing model and then just animated, but certainly works well.
Oh, Anish made a game that's fun. Astra's Opus 4.5 for games made Contra, but photorealistic this weekend. Link to play below, and it's up on Vercel. I wonder if we will get an actual boom in Steam. It feels like Steam will be going through something similar to the Kindle Direct Publishing store, where they will be flooded with inbound. Didn't we see this with the App Store, too, where the number of App Store submissions for Apple 10xed?
Totally overwhelmed their systems.
But no real breakout apps. The top of the App Store is still ChatGPT and free money or whatever.
Yeah, it's evolved a little bit. The company that I'm confused that's been consistently in the top 10 is this company Vinted.
Pre-loved marketplace. Buy and sell secondhand clothes. People are buying clothes. It's true. Shopping app. Everyone needs clothes.
You make it to the top of the store.
Yeah, you'll be able to vibe code a game, you'll be able to get it on Steam or in the App Store. How do you get the distribution flywheel going? How do you innovate to actually break through? How do you make something that goes viral, gets traction? Because the long tail is going to get a lot longer. What do you think? I feel like for a lot of these games, though, you should just keep them on web because it's much more accessible. Not everyone has Steam. You have to actually download—
I feel like web games are still not that fun.
There were some web games I was playing this weekend.
Did you play more web games this weekend, or did you play more Oculus VR, Meta Quest VR when you had the VR headset for an unlimited amount of time? But let's call it a weekend.
Okay, I was probably playing more VR.
I was playing Call of Duty.
You played Call of Duty in VR?
How long? Like a couple hours?
Yeah, probably a few hours. Not bad.
Well, when I was making a game, too, this weekend on Astra, I spent a lot of time.
So does that count as playing the game?
Can you drop it in the chat for everyone to play?
I do think that playing a game that you designed yourself or for a small group chat, it's the same thing as the images. The AI images I generate are not broadly beautiful; they're not going in the MoMA. But they will often be funny to me and three other people, right?
Yeah, low TAM. But that's the beauty is that with lower cost, you can go after lower TAM opportunities. And so, this happened a couple months ago. Jeremy Ghoen Simulator is the classic example. Someone sent us Jeremy Ghoen Simulator Version 2, which was very low TAM in the sense that it was just in-jokes with me and a couple friends. And the long tail of these things will continue to flourish, and then of course people will use them in real businesses built on IP, built with distribution.
You know what time it is, John?
What time is it? Time for an ad. Console.
Console builds AI agents that can automate 70% of IT, HR, and finance support, giving employees instant resolution for access requests and password resets.
Okay, it's trailer time. We got two trailers that hit the timeline. Let's start with Nathan Fielder and his new film featuring none other than Elizabeth Holmes.
Elizabeth Holmes filmed before she went to prison.
I don't have anything to deceive you on. Of course I'm not deceiving you. I'm engaging with you as a human being. Why would I deceive you? There's no—there's no reason for me to do that.
The acting is so good. Okay. Good shot.
How long is this being real right now?
The team is stoked. Looks beautiful. Good sound design.
Saw a post somebody saying, 'It feels like I'm living in a simulation because this media product is like perfectly—' You would think that a movie like this wouldn't get made by A24 specifically because it feels like the TAM is just like 20,000 people, but—
The Theranos story was really big. She was on the cover of Time magazine, I believe, and cover of major magazines. It was all over the news when it happened.
No, it's a lot bigger than you would—
The John Carreyrou book is just a fascinating read. It's a real page-turner, and so I think a lot of people read the book. The adaptation of the actual play-by-play story, not the documentary, but the biopic I guess, that I think did not break through in the same way, but it certainly continued to elevate her persona. And then the posting has kept her relevant.
Yeah, there was some speculation that maybe it was Nathan that got access to the account.
I completely disagree with that. I do not think that's all. I think that would violate some tenet of what it means to do a documentary. I mean, Nathan blurs lines all the time in terms of how he interacts with his subjects, so it's certainly possible, but I just don't think that's reasonable at all.
How much do we know outside of the trailer, though? We don't know.
Well, this was leaked something like a year ago from a journalist who got the story, and it was widely denied that it was happening, and he was very upset that he wasn't getting enough credit for being correct here. He ultimately was. But yeah, all we know is that they filmed the documentary in the lead-up to her actually going to prison after sentencing. There was a pretty large gap before when she had to report to the prison. She gets out in 2030, so coming up three years, four years away.
And so I believe even in a week you can get a lot of footage if you're spending every day with someone. And then there were a number of visits to the facility by Nathan that were recorded, probably talking through glass or talking at a table, something like that. So I think some flexibility there. Maybe some phone calls, maybe some correspondence in the aftermath.
What I'm thinking: SBF, Nathan Fielder.
Yeah. This just becomes like a—
The problem is that he's already in jail. So the question is just what else does Nathan Fielder have in the can? Because he's obviously thinking years ahead since he filmed that exact scene in I believe 2022 or 2023. And so who knows? There's a rumor that he's secretly running a boy band right now. There's this rumor that there's a boy band that's popular on TikTok that's entirely invented by him. Are you familiar with this?
Really? I thought you were going to say Nathan Fielder is SDKit.
These projects will emerge, and obviously they're taking longer and longer because they're bigger films now.
After The Rehearsal, it was hard to imagine how he would one-up himself, and it feels like he may have done it.
Yep, he found a new tall mountain.
Stay with us in the trailer zone. When we come back, we're going to be talking about Artificial.
Let me tell you about the New York Stock Exchange. Want to change the world? Raise capital at the New York Stock Exchange. Just do it.
Pull up Artificial. Trailer dropped just this morning.
What is the image of the future that you see?
Wait, pause. Was that a Valkyrie?
Okay, continue. So the visuals are already incredible.
I'm pretty sure that's a Valkyrie. It might be a Valhalla. Is that Lucas Singer?
No, because you can see the suspension, that front section. Anyways, continue.
What is the image of the future that you see? We've created a machine that will solve the world's problems.
Gate check. Gate check. You nailed it. Oh, the gate for sure.
We open Pandora's box together. The future's inevitable. Countries will fall. Industries are going to collapse.
The voice is still Eduardo Saverin. But the styling—and the crazy thing is I have that exact backpack.
I think I got it recommended by MKBHD.
—to shepherd people into this new world.
Lots of cameos. Not cameos, but—I'm so excited to see it. Dramatic.
It'll be interesting to see the takeaway. People in the chat are back and forth: Is it too soon? Is it unnecessary? Is it going to be something that, like The Social Network, is divisive, but also oddly inspiring to a certain cohort of entrepreneurs? We'll see what the reactions are.
Do you think Hollywood took away any lesson from The Social Network in that they were trying to frame Mark Zuckerberg as this generally troubled person, and then a lot of people came away feeling like, 'Wow, I want to be like Mark Zuckerberg'?
Sort of an unintended consequence of it. They're going back for a second scoop of the ice cream with The Social Network: The Social Reckoning. Is there a trailer out or just images? I think we saw that, too. So it'll be interesting to see what happens there.
Get your tuxes ready going back-to-back: Artificial, Social Reckoning, and then capping it off—
Yes. And then also the To Catch a Predator movie. That's going to be a good one, too.
Yeah, I guess you're right. But still, triple feature.
It just feels like it's a year of niche movies that I expect to be breakouts along with Backrooms and Obsession, these two movies, and then the documentaries. There's a lot of things that are oddly specific, but I would expect they do well. The Social Network did very well at the box office.
While we're at it, let's stay in the trailer zone and let's pull up The Social Reckoning, the official teaser.
Tyler Culbertson is calling it pure cinema.
But I am here to help Facebook, not hurt it.
Okay. You sent me a message. What would you like to talk about?
The Chairman gavels the session to order. You'll read your opening statement, which we'll skip past for now. That's a separate session.
I don't think this is based on politics. It's not Cambridge Analytica. It's the Frances Haugen teenage girl whistleblower thing, which honestly looks a lot different in the backdrop of the addiction trials. I wonder if they were trying to get this out before the trials.
Hang on, I don't understand what you're saying.
Aren't you a tech reporter?
These guys are counting on the next round of congressional testimony to make you likable, Mark. I'm happy to lend a hand, but I think you're doomed.
This company and that guy are playing an unprecedented role in our lives.
The fire hose of bad information you are injecting into the air supply is becoming jet-powered.
I'm a free speech absolutist. I'm not the one who's lying, and I'm not stopping them from seeing someone who is.
Anxiety, depression of teenage girls got worse as a result of time spent on the platform.
Senior leadership knows and is doing nothing.
I know there are easier enemies to make. The mafia would be an easier enemy to make. So what would you need to stand up the story?
This is a material violation of my NDA.
We're twice as big as the biggest country on Earth. We're not frightened of Congress. We're post-government around here.
Who is that supposed to be? Please let me poke that.
We have 102 hours to get everything.
She's going to get sued into small pieces.
I don't want to be made an example of by a guy with unlimited resources.
Yeah, chat is calling it out, but I think this one's going to flop because of the lack of supercars, and this story is no longer top of mind.
People want to hear about the talent wars. They don't want to hear about this kind of stuff.
Do the talent wars. If there's no Alexr Wang character, I'm not paying good money. SAS chain says, 'Ooh, technology so scary.' I don't know, they should do more tech movies. Maybe Ces says it will flop because no Brazilian jiu-jitsu. Yeah, this was in the pre—
That's true. Make it in the script. UFC.
The Social Network 3, all the scenes where Zuck is talking in the middle of—
How is it 2026 and we don't have a Salesforce movie?
Like the comeback, the SaaS apocalypse.
Or at least a documentary on the relationship that he has with the dolphins.
And the wildlife generally. I'd like a CrowdStrike movie. I'd like a MongoDB movie. Let's do it. MongoDB: What's the only thing faster than the AI market? Your business on MongoDB. Don't just build AI, own the data platform that powers it.
Here's a concept: DolphinForce, how one man tamed wild dolphins.
It's maybe the best origin story of a tech company ever. I don't know if there's a better one. It's just pure bliss, pure positivity. For those that don't know—
Benioff famously conceived the vision for Salesforce in 1999 while swimming with a pod of roughly 100 dolphins off the coast of Hawaii during a sabbatical. Feeling one with the pod, he envisioned cloud-delivered enterprise software, leading him to quit Oracle and secure $2 million in seed funding from his mentor, Larry Ellison.
It would be great. Anyway, in other news, lots of craziness in the AI world, lots of new models, but the jobs apocalypse has been officially postponed by The Economist. An AI jobs boom is here. And people have a wide range of takes on this. Is it just postponed? Is it permanently postponed?
Kevin Bryan summed it up well. He said, 'I am shocked.' Shocked. So he's being sarcastic, of course: 'Shocked to find out that a productivity-enhancing investment-supporting technology is good for workers. Knowing nothing else, this should be your prior because this is how productivity and competitive markets almost always work.' A new technology comes out, it supports investment, it enhances productivity, everyone hires to go after all the different battles.
The Economist breaks it down in a lot more detail with a bunch of charts and a bunch of data. They say, 'Perhaps AI will eventually make many humans unemployable, but there's no sign of it yet.' On September 4th, the Bureau of Labor Statistics reported that the American economy added 162,000 jobs in August, far above expectations. The unemployment rate was just 4.1%, lower than almost 90% of months over the past 50 years.
Young workers, often cast as AI's first victims, are holding up remarkably well. The gap between unemployment among 20 to 24-year-olds and the overall rate is close to a multi-decade low. Some companies and workers are being severely disrupted by AI. Hiring in professional and business services is running 10% below the average from 2015 to 2019. Tech giants like Microsoft and Meta are trimming headcounts as they reorganize their businesses around the technology. Smaller firms such as Block—which is sort of a crazy mog, smaller firms, but it is smaller than Microsoft and Meta—Block, the owner of Square and Cash App, and Intuit, the maker of TurboTax and QuickBooks, are replacing people with bots.
American companies have announced some 16,000 AI-related job cuts a month on average so far this year, according to Challenger, Gray & Christmas, an employment consultancy. What a funny name for a consultancy. But AI-related layoffs get lost in the churning job markets where—remember that number: 16,000 AI-related job cuts a month. In a typical month, the US workforce just—
churns 1.7 million jobs and then they add 1.8 million jobs. So there's net gains, but you're looking at less than 1% of the layoffs that happen in the economy, or job losses or firings. Less than 1% is AI-related, which is what The Economist is pointing out here. And the evidence so far is that AI is already creating a lot of jobs to replace those as it has destroyed. The vast sums pouring into data centers and power generation have set off a race for construction and infrastructure workers. AI startups are hiring like there's no tomorrow. Incumbents are racing to keep up with new AI roles and making some workers more productive. AI may be increasing the demand for their services.
Add it all up and The Economist estimates that AI so far has created around 1 million new jobs in America. That easily exceeds the roughly 200,000 layoffs attributed to AI since mid-2023, and it appears more than enough to offset weaker hiring in many back-office roles. America's AI infrastructure splurge has created many of them. The spending on the kit needed to make AI run—from chips and servers to data centers, cooling systems, and power—is roughly $500 billion a year above what it was in 2022 when the world got to know ChatGPT, calculates Goldman Sachs, a bank. Thank you, The Economist.
Data center construction alone is proceeding at an annual rate of more than $75 billion, nearly 60% higher than a year ago, according to Census Bureau data. That building spree requires armies of workers: electricians to wire them, HVAC specialists to stop racks from overheating, grid engineers to hook them up to the power supply, and technicians to install and maintain the machines. The hiring boom is visible in the numbers. The Economist tracked five industries at the heart of the data center buildout, from electrical contracting to equipment manufacturing. Since 2023, employment in them has risen by roughly 320,000, more than broader construction and manufacturing trends would suggest. The BLS expects utilities to be the fastest-growing big sector between now and 2035. Not all those jobs owe their existence to AI; grid upgrades and other factory building matters too, but lots of them do indeed. Indeed, a jobs website, finds that data center vacancies have more than doubled in two years as job postings overall have fallen. LinkedIn, a social network for strivers—these are so funny.
Goldman Sachs, a bank. LinkedIn, a jobs website, a social network for strivers, estimates that nearly half a million data center jobs were created between 2023 and 2025 in America. The scramble for workers is showing up in paychecks too. Indeed finds that installation and maintenance jobs at data centers advertised wages about 40% higher than comparable work elsewhere. Official wage data tell a similar story. In the year to June, average hourly earnings rose more than 13% in electrical equipment manufacturing and nearly 8% among electrical contractors. Even with the pay gains, workers are still not easy to find. On a recent visit to a transformer factory, Donald Leavens of the National Electrical Manufacturers Association asked the company's chief executive what help she needed most. "Can you come in here and help run one of our lines for me?" she replied.
It's not just hard hats that are proliferating. AI is also creating a new class of white-collar jobs. Engineers build the models. Data annotators—Tyler's clapping for the white-collar job creation—data annotators label their inputs and judge their answers. Forward deployed engineers adapt them for customers, and newly minted heads of AI decide what companies should do with the technology. And there's some interesting reporting in the Journal about Google and Accenture teaming up to do more forward deployed engineering. They're deploying—I think there's 1,000 people on this one team. That's just an example of new roles for AI diffusion. Some of these roles barely existed until recently, says The Economist. Many are quickly growing in number. Postings for heads of AI, AI engineers, and directors of AI have roughly doubled since 2023–2024.
The numbers are starting to add up. Preliminary research by economist Gad Levanon at the Burning Glass Institute uses the research outfit's career history database to identify jobs that would not exist without AI, whether at AI-native firms or because they are AI-specific roles at other companies. He reckons roughly 1% of professional jobs are now AI jobs, on the order of 1 million positions in America in computer occupations and life sciences, including researchers using AI to discover new drugs. The share is 4% to 5%. LinkedIn's own analysis points to roughly 640,000 new AI-specific jobs between 2023 and 2024. So Jevons paradox. The Economist tracked employment in professional occupations closest to the AI boom—engineers, software developers, mathematicians, and data scientists—and compared their growth since 2022 with professional employment overall. These roles have added roughly 730,000 jobs above the trend in recent years. AI will not have created every single one of them, but it almost certainly created quite a few.
A third source of jobs comes from productivity gains. AI allows lawyers to draft contracts faster and analysts to comb through financial filings in minutes. If higher productivity lowers the cost of professional services, it is possible that demand for them can rise enough to create more work overall. Nikil Baid was talking about this this weekend. He used Astro to create three different potential deck remodels, redesigns, building expansions to his home. And it was interesting because it's a lot of work that you could say, "Oh, that's going to displace an architect." I don't think the models are at a place where you would trust a deck that's built in Blender from even the best model. I was finding in my Blender modeling tests a lot of misaligned beams and things that weren't quite right. But in terms of visualization, in terms of getting excited about a project and getting to that next stage of saying, "Hey, okay, I have a really solid vision here. I can imagine what I want to do. Let me actually go and do that." There's this odd diffusion that happens in the economy. I was talking to Saagar Enjeti about this—the credit card effect, where credit cards had a big effect on the economy in a very, very boring way. They just slightly lubricated the machinery of
Global commerce, basically. So you're just like, "Okay, if I send a wire, I'm not going to be able to get the money back. But if I put down my credit card and it doesn't show up, I can probably get a chargeback." And so that instilled just a little bit more confidence that what you're buying is accurate, and it enabled more commerce, right? And so the same thing is starting to be true here where people can say, "Okay, I'm ordering shoes. Let me find the shoes that fit me perfectly, that are the right price, that will come on time, that are from a reliable manufacturer and aren't going to fall apart." And if you get a very solid report back that gives you confidence, you make that purchase just a little bit sooner. And if everyone is doing that, it winds up speeding up the machinery of global commerce, which is incredibly boring, but incredibly valuable.
Yeah. One of my favorite examples is you were looking at a number of different houses, all of which were not necessarily in the condition where you're like, "Great, I want to move in here." And you just dropped the Zillow links into Codex, and it made you an entire website that you could walk through with—
Before the current state, and exactly how they would evolve based on your stylistic preferences.
Exactly. So oftentimes, a fully remodeled house will command a 20% premium, but you can get sort of sucked into, "Okay, well, the newer one that's already been through the remodel, it's just visually more striking." So you're like, "Oh, I want to necessarily pay up," when in fact you might want to buy the cheaper house, do the remodel, and then wind up with the same end product and maybe capture some of that value yourself. And so there's a lot of different places. A lot of times the conclusion is very underwhelming. I mean, we were talking to Nick about this where he was taking a picture of his room—and I saw a lot of people doing this online—take a picture of the room, make sure I've decorated it properly, that the bed is in the right spot. And I think the result for you was like, "Yeah, it's basically as best you could do. Good luck, dude," which is hilarious, but at least it gives you confidence that, okay, yeah, I really shouldn't order a bigger bed. The bed I got is the right one. And if you do that on the way in, you wind up moving things faster.
We covered Navier-Stokes. I don't know if there's more to cover there. The chat's asking about it. We covered it at the top of the show. We also have Greg Brockman joining the show, so we'll ask him about advances in math, what's going on with Navier-Stokes, the back and forth there. He's joining at 1:10 PM Pacific. So, stay tuned. And we have Scott Wu as well, who is an IMO gold medalist and a fantastic mathematician. He can of course comment about this because he at Cognition is in a unique position of both being incredible at math, but also Cognition's whole business is not in the theoretical math space; it's in AI diffusion. It's in getting work done in the enterprise for real companies. And so I think he's in a unique position to both ground the conversation around what it actually means—how impressed is he by these solutions—and then also what does it actually mean for, you know, Boeing. Will they be able to make more planes more efficiently, more reliably? That is a valuable thing that we actually want to see in the real world.
Something happening in the real world: former F1 CEO Bernie Ecclestone was stopped in Portugal at the airport for
bringing in a shotgun from Switzerland.
He's 95 years old, and this is actually the second time he's been stopped at an airport. In 2022, he was arrested in Brazil for illegally carrying a gun while boarding a private plane to Switzerland,
and was ordered to pay a fine to local authorities. So rough weekend for Bernie. But at this point, I think he's like, "Sorry, I'm 95. I'm not going to stop bringing guns everywhere I go on planet Earth."
I got to do it. Well, let me tell you about Codex. Codex is a powerful workspace for getting work done with AI agents. Whether you're writing code, analyzing data, creating content, or automating business workflows, Codex helps you move projects forward from start to—
Says, "Just let grandpa carry his gun. That's all."
In other news, Wimbledon does not plan on providing credentials to influencers next summer in hopes of avoiding the issues that affected this year's US Open.
Players have called for spectators to follow the sports etiquette after matches were interrupted. They specifically call out taking a picture of the court and putting logos all over it using AI and then posting that on X and getting a couple hundred likes.
He's working. We're just making jokes at your expense. Don't worry about it. You're good. Was there anything specific that was tied to this?
I think it was a number of issues. There were also complaints—one of the players was complaining about the smell of marijuana in the stadium,
But yeah, next year Tyler will be going to the US Open in disguise.
So full Hollywood makeup.
Yeah, you can look forward to that.
Yeah, I wonder what are they going to do? Just not give free tickets to influencers, or actually block people who are influencers? I think what was happening is they were probably giving media/press passes to influencers.
I think the influencers were doing too much stuff, like getting in the way of fans, creating scenes because they're running around with gonzo filmmaking equipment as opposed to something that's a little bit more cordoned off and polished.
It's different when you have a red carpet and you're doing proper interviews. That sort of lands a little bit better.
Anyway, Sheel Mohnot is calling for a revisitation of the Cattrini piece. Agents become the default demand-side interface and transfer economic rents from incumbent intermediaries to whoever controls the agent, which does seem correct. And so with instinct, who should you be long, who should you be short? There's an interesting dynamic where anytime something happens in AI, there's a lot of people that go—I think Bucco pointed this out—that everyone will go and ask AI, "Okay, what are the longs? What are the shorts?" And so that creates an amplification in the market. People piece together this new trend: "I want exposure to this trend," and so things get overheated much more quickly.
But yes, this idea that if it's a restaurant booking platform and your AI agent can call the restaurant and get a booking just as easily as they can go through a platform—now Resy has exclusives, so that creates more of a complex scenario, but by default, if you can just write an email or make a phone call or send a text message, an AI agent should be able to disintermediate those platforms. But that's not true for Uber. That's not necessarily true for DoorDash, where it's very difficult to find a person to go and pick your food up or pick you up personally, because they need to be very close if you want the car right then.
So Sheel says, "How f'ed you are in an agent-disintermediating world completely depends on how much control you have over supply. To answer your question, you may not have a choice in whether you allow bots. Expedia can block agents, but it doesn't control the hotel inventory. If Expedia stops letting agents book, they'll go through booking.com, Google, or direct. Resy has more leverage via exclusives—and I'm sure Expedia has some exclusive program, I imagine. But if agents become a major source of diners, restaurants will want to be bookable via agents, and reservation systems that support agents will win supply. Someone will come along with a cheap direct booking tool. AI can disintermediate discovery and customer ownership, but Resy can survive. But can Resy survive as the restaurant's reservation infrastructure? That's a much smaller rent pool unless the inventory network remains differentiated. Uber is much harder to disintermediate. It actually controls a valuable network of drivers, dispatch, pricing, and payments. Why would they allow agents? Basically, they can afford not to. After an AI agent takes over your consumer interface, how much economically essential stuff are you doing that only you can provide? For Expedia, it seems tough to me." So Sheel is worried about Expedia. Anyway, was there anything in the chat you wanted to run through? Because we have our next guest joining in just a few minutes. We have Dan Wright coming back on the show from Armada, building modular AI supercomputers in a shipping container. A supercomputer in a shipping container.
Apple's coming out with new phones. We're getting a folding phone probably this week.
It's going to be exciting. I think it'll sell well, and it's been a while since there's been a visually differentiated Apple product.
Yeah, I think it's a pretty easy sell. "Hey, you use this thing for 3 to 8 hours a day, or if you're Taylor Lorenz, 18 hours a day. Would you like a bigger screen?" I think the answer for a lot of people is yes. "And would you like to have some differentiation versus your peers that have just one simple screen?" I think a lot of people are going to opt into that.
I wonder how it's going to look. I've seen a lot of leaked renders.
A lot of the renders look squarish.
Yeah, I feel like they're going to surprise us with something. It feels like the news and the facts and things that can go over text message sort of leak out. And Mark Gurman—I saw a stat from Eric Newcomer that Mark Gurman has an order of magnitude more scoops than the next biggest journalist on Techmeme or something like that.
Yeah, he's like the power law tech freakout, seriously.
Discuss things with Mark and say, "Oh yeah, we're not working on a car anymore. Keep that off the record, but I'm letting you know," and he'll report it out. But people will stop short of sending him an image of the new phone. And I think he's been pretty good about that.
It was actually Josh who says, "We don't talk enough about how Mark Gurman is the most cited, most market-moving tech journalist by nearly an order of magnitude. The whole industry of Apple reblogs and day traders live off of his scoops. It'd be like if LeBron won 40 championships."
That's a really good post.
Well, we love Mark Gurman. Well, there's one more story we got to cover before we go into our guests: a man drove a Ferrari Purosangue from South America to Alaska. This is a crazy, crazy story. The Pan-American race has been talked about a long time. It's really, really difficult with the Darién Gap, the gap of wild forest that exists between North and South America. You usually have to put your car on a boat to get around that part, and I think it still counts. But it's still a very treacherous road, a lot of weather conditions. If you break down, you might be very far from anything. But a man named MJ has taken a Ferrari Purosangue far beyond where most owners would ever consider driving one. His goal was to travel the length of the Americas, taking the V12 Ferrari from South America all the way to Alaska. He's now attempted the journey three times, putting roughly 80,000 kilometers on the Purosangue in the process.
His dog joins him for the journey. The first attempt ended after an accident; he brought the car to a dealership in Santiago. On his second try, they made it all the way to Alaska. For the third, he reversed course with an even bigger goal, attempting to set a record from Alaska to Argentina before a major road closure forced the run to end in just two days. Through dirt, snow, and thousands of miles of pavement, the Purosangue has effectively become a long-distance expedition car, with Rico, his dog, riding alongside him through it all. What a heartwarming story.
A beautiful ad for this naturally aspirated V12.
Look at that. Look at that shot. So good.
Oh, I didn't realize it was a Chihuahua. That's funny. That's probably a more convenient dog to have in a car for a really long trip. You don't want a big dog that needs to stretch his legs. You want the dog to be able to just run around the cabin, I think, if you're on a good road trip. Anyway, we have our first guest of the show, Dan Wright from Armada, in the waiting room. Let's bring him in to the TBPN Ultradome. Dan, how you doing?
I'm doing great. How you guys doing?
We're doing fantastically. Welcome to the show. Welcome back. What is new in your world? Walk us through the latest, and maybe just since it's been a while, reintroduce the shape of the company, the mission, and where you guys have gone in the last year.
Yeah. So, Armada is the hyperscaler for the edge. We build the infrastructure for the 70% of the world that doesn't have the big hyperscale data centers today. The mission of the company is to bridge the digital divide and make sure that we have AI everywhere, wherever we need it. But first, you got to have the infrastructure there, and so that's what we're doing. We've been very busy. We just raised earlier this year $230 million at a $2 billion pre. We launched Galeon Forge 1, which is a factory with our partner Johnson Controls in Gilbert, Arizona, where we're now continuously manufacturing these modular AI data centers called Galeons that we've built.
How is power usually solved for when you deploy these?
A lot of times we're deploying them where there's already stranded power. A good example of this: a couple of weeks ago I was in New York and we did an event at the New York Stock Exchange, where we were fast-following on Jensen's announcement that compute is now an asset class. It's a new asset class, and the New York Stock Exchange is actually making that something that you can invest in the same way that you can invest in electricity or other types of commodities. And we were there with our customer in Norway called Fossefall, that has tons of distributed sites all over Norway. I'm actually going to be in Norway with them next week, but they also have them in Finland and in Sweden. And it is largely renewable energy, so it's 99% hydroelectricity.
They can just plug in our AI factories and then scale up quickly. And one of the things that makes that easier is that we've recently announced some larger form factors. Last year we announced Leviathan, which is 2 megawatts per unit. And when I was in New York, we announced Orion, which is our newest form factor. That's 10 megawatts per unit. When I was launching Leviathan, everybody said, "That's great, but how do we scale up even faster?" And so now Armada can credibly say wherever you have power, we can plug in and we're the fastest from 0 to 200 megawatts anywhere in the world.
Talk about why, for example, one of these companies needs to have that compute actually at the edge and why it matters across different industries.
So one is latency, and this comes up in a lot of conversations. I'll give you another real-world example. We work with the state of Alaska, and they don't have the big hyperscale data centers there. So the first data centers deployed were ours. We did this last year, and we're still working closely with them and kind of scaling up with them. They had 28 hours of latency to process data from drones for emergency response for avalanches and floods. We also deployed last year with the Navy in the middle of the ocean—very similar types of use cases where if you have a large distance between the source of the data, where the data is being generated, and then where it's being processed, the data sort of becomes worthless because avalanches and floods and battlefield scenarios, they don't wait days or even minutes. You have to be able to use the data in real time.
Another big driver of this is sovereignty. The shorthand for our value prop is the three S's: speed, scale, and sovereignty. And the sovereignty piece is really important. There's this global trend that's going on around sovereign AI. Everybody wants to be able to take the latest models, but they want to be able to fine-tune them to really sensitive datasets that they wouldn't send to the cloud and have a sovereign AI infrastructure that they actually own, and that's what Armada enables.
Speed of delivery—yeah, that makes a lot of sense. Is hydroelectricity under-discussed right now? Is there an opportunity there in America or abroad? We talk a lot about solar where it feels like there aren't as many big winners yet or big hot startups that exist. In nuclear, that's very exciting. We've heard about wind, but no one is really talking about, "Let's just do another Hoover Dam or something." Is that possible? Are you optimistic about that?
I think so. I think it should be talked about more, and I think in general, stranded energy should be talked about more. We're doing these projects all over the world. As an example, the Niners are playing in Australia this week, and we're doing an event there with another partner called Windlab that we work with there, and they've got a huge amount of stranded wind and solar energy. Australia's big in wind and solar, Norway's big in hydroelectricity, but this is happening all over the world. There's these stranded pockets of energy. Last year, they had to curtail 7.2 terawatt-hours of energy in Australia because the grid's completely overloaded, but they have this stranded power that we're just bringing the infrastructure directly to, and then you can scale up into the hundreds of megawatts. Ultimately, same with Fossefall, they'll scale to over a gigawatt over the next few years, and we can scale with them.
How flexible do you want to be around the actual chips that go into the systems? Because when I hear latency, I feel like Cerebras, Groq, these faster systems—there's Taalas, which I think AMD just bought, where you're baking the weights on and you can actually get to an inference speed for certain AI workloads that would actually benefit from saving 100 milliseconds. Whereas if you're putting a bunch of NVL72s in a data center, you're going to be waiting while it's cooking, right? And the last mile is going to be negligible.
Yes. So again, it's all about speed, scale, and sovereignty, and then giving the customer the choice. So we say, "Okay, well, what workloads are you trying to run?" And then based on that, you can sort of rightsize the infrastructure to the actual need and where they expect it to go. And a lot of what our customers want to do is they want to try different things. They want to try obviously the GBs, but now they're looking at the Vera Rubins. They're looking at things from other chip companies as well. And so the nice thing about our Galeons—again, we're manufacturing these; it's not like construction. And so as you want to try different things, we can just make quick iterations on the design, then ship them out. So they can try different things, and then what they like, they just order more of and we scale up with the demand versus with a traditional data center.
The downside is not just, as we were talking about, the power—you got to figure out the power situation. We take advantage of the power that's already there, but what we also allow them to do is scale up with demand so that they don't overbuild or build the wrong thing.
How are you thinking about staying aligned to why and how the company started, and the focus around this edge compute versus the natural pull from the market? Where if somebody comes to you and says, "Okay, now I want the 1-gigawatt data center, and we think you guys can do it"—just running the numbers on that and resource allocation. I imagine there's a natural pull to go bigger and bigger, which is a good problem to have, but it will be interesting.
It was actually super simple for us because one of our company values is heal the customer's pain first. We're obsessed with customers and what their pains are, and how do we solve those. The evolution of our Galeon product line has all come out of conversations with existing customers where they've said, "Okay, I want to use your edge Galeons—the smaller ones, say sub a megawatt—for inference." But then they came to us and said, "I also want to do fine-tuning of all these models—latest open-source models, closed models, my own models—on these proprietary datasets without sending them to the cloud."
And so what we enable is sovereign AI factories where they can do both. They can take the latest open-source models—you probably saw what Nvidia just did with Hugging Face and with Poolside; I think we're going to see a lot more around open source. They can take models from the OpenAIs and the Anthropics of the world. We can help them fine-tune those to these sovereign datasets and then push them to all the edge nodes to run. And then you can do what's called federated learning, where you're actually fine-tuning that model on that data at the site, and then you're using it to improve your core model and then pushing the updated model out to all the sites, which has the benefit of being more cost-effective, more secure, and enables you to take advantage of lots of different models, experiment with different things, and see what works.
Well, congratulations on the progress. Thank you so much for coming on.
Great to see you. Congrats on the progress. Let me tell you about Shopify. Shopify is a commerce platform that grows with your business and lets you sell in seconds online, in store, on mobile, on social, on marketplaces, and now with AI agents. Who do we have next? Jordy,
We got Eric Seufert, the founder of Mobile Dev Memo, live with us on TBPN on YouTube, not on Netflix yet. Should we be considering switching teams? Eric, just for what it's worth, every time you come on the show, I call John afterwards and I'm like, "Eric's in my top three all-time guests on the show." It just fires us up to talk with you every time.
Good to see you, and it's great to see you. Yeah, good to see you. So I told my wife I'm going back on TBPN, and she said, "Is that the show with the two handsome hosts?" And I said, "Yeah, but when I'm on, there's three handsome men."
And it was like that meme with Natalie Portman where she's looking at the guy, and I was having to say, "Right, there's like three handsome guys on the show. Right." So, I think you're ready for Netflix. I think Netflix would be happy, would be lucky to have you.
What would the economic dynamic be? Because Netflix and YouTube are both going for exclusives now. What's actually playing out? How much are they at each other's throats? How much are they going to converge in the way that TikTok and Instagram and YouTube are all converging? Is there going to be another convergence, or is there actually a stress point there where YouTube is trying to be both Instagram and Netflix and they maybe can't do both?
Well, yeah. I mean, they are converging. This is a really fascinating dynamic right now in streaming. I've been following it with a series called Netflix's YouTube Opportunity for roughly a year. What I wrote about first was they brought Ms. Rachel on. What they did was they brought Ms. Rachel on, but they didn't buy new content. They just paid her to bring her existing content over. So you've got this proven base of content that has however many billions of views—I know my children probably account for several billion—where there's essentially no risk. You know there's an audience, you know it's popular, and you pay some amount of money to bring over this existing content.
It's actually really great for the content creator too, because they don't have to produce anything new; they just chopped up the existing content into a season. So they packaged it like a season. They also did that with Danny Go!, and they also did that with Mark Rober. My point was Netflix had been executing tremendously well for the past several years. They had commanded extreme pricing power. The premium service is going at almost 30 bucks a month in the United States. The advertising tier price has been tracking what I think is the ARPU differential to make it equivalent to the next highest tier, so they've been doing a good job of maintaining that pricing power. My sense is they did too good of a job, and so they made it really hard to compete in the space, and that pushed everybody to the bottom.
That made everyone take the exact opposite approach, which is: okay, let's go FAST, let's go free ad-supported television. Now Netflix has an even bigger problem on its hands because they may have reached a ceiling with what they can charge. Certainly premium—I don't know how much higher you can go than $30. They've probably extracted all of the net additional subs they can from password-sharing crackdowns. Now they're faced with: okay, we just have to bring in a lot of live events and sports, things like the Beyoncé bowl and the January NFL games. It's really expensive. And so, are you going to compete on content, or are you going to go the opposite direction, race to the bottom, and try to get UGC on there?
I think that's probably what they're trying to thread that needle on, because if you listen to Netflix's leadership, they say, "Look, we can never go pure FAST because that would have a deleterious effect on the brand. We're seen as a premium service." But how can you maintain that when you see Roku and Tubi joining forces, when you see all of the pressure that Amazon is putting on your business? Because Amazon is the identity spine for advertising across all of FAST. All of FAST/CTV is Amazon with these data deals providing all the identity for the buying. And so I think they're facing this opposite problem now where they had too much pricing power, they pushed everybody in the opposite direction to go FAST, and now, honestly, if you believe the reporting that the New York Times did that they're going to be offering up these subscription bundles, how can you not go FAST and do that?
Are you going to be willing to accept paying $30 a month for a premium Netflix tier, seeing content in your Netflix app on your TV, and having it say, "And if you want this, you have to pay even more to subscribe to this other channel"? I wouldn't. I think the only way they can do the bundling is if they adopt FAST.
That ARPU gap from the premium ad-free tier to ad-supported—is there a reason why the gap doesn't make up the full $30? Is there some economic reason, or did they just get stuck in this weird place? Because it feels like if they got even a little bit more juice out of the ad model, they could just offer a free ad-supported tier. But maybe that would create more churn, because mentally going from $30 to $10 is different than going from $30 to $0. Do you understand more of what they're grappling with? Did they get unlucky, or is there something about the structure of Netflix that would really make it difficult for them to do a fully free ad-supported tier?
No, I think it's just a reservation they have. Maybe it's almost like a superstition. Keep in mind they were saying for the longest time they would never do ads. Reed Hastings said—
They always do until they don't, right? Until they change it. Reed Hastings said ads were like a blight. He said personalized advertising was a cancer. And so then they changed their tune. But there's three tiers: it used to be Basic, now it's Standard with ads, Standard, and then Premium. Standard is what they have to make up the ARPU gap with. Premium includes a lot of stuff, and so maybe you can make the case for a high-income household it's worth paying for the higher streaming quality and the more devices that you can onboard. My sense is with Standard, that ARPU gap is just indexed to the ads ARPU.
The question is just internally: what is the resistance to going FAST? My sense is it's that perceived quality that they would lose if they went FAST. But the market moved in that direction, they pushed the market in that direction, and so I think they're going to have to capitulate.
Have you been surprised that Amazon Prime Video hasn't seen more UGC? Because I was looking up randomly, what does it take to publish a film? You can submit to Amazon Prime Video, that goes through a review process, but the cost to submit is de minimis and you can get a video product up there, but it hasn't actually seen a groundswell. And I'm wondering if that's because Amazon specifically has partnerships that scratch that itch.
Yeah, I think that might be the case. Amazon has all of these identity partnerships that essentially make it the data spine and the identity spine for FAST broadly. And so if you think about—I remember people saying, "I bet Amazon's going to buy Roku." And I remember thinking, "Why would they? They've already got this identity partnership with Roku where they get access to the best impressions. If they bought it, they'd just be getting everything else that they don't already buy." And so my sense was always that they're already picking over the stuff that they're not buying; why would they want to own it?
My sense is, yeah, maybe it competes with the FAST channels that they're partnered with, or maybe they have similarly some sort of perceived quality bar or hurdle. But I think the UGC thing is a little bit of a distraction, because I don't think Netflix is ever going pure UGC. They are very selectively curating the YouTube creators they bring over, but they brought over a lot, and they went on a shopping spree this summer too. They bought a lot of stuff this summer, and now YouTube's pushing back.
I saw some post about how the results of that shopping spree that Netflix went on with YouTube creators had a significant power law where there were some creators that were really performing well on Netflix. Is that because of the way the Netflix algorithm works? Is that true? Or do you think that if somebody's getting a billion views on YouTube, we can pull them over and get X views and there's some multiplier, or is there something special about the way Netflix is actually rolling out these deals?
Well, I think there's a couple pieces there. YouTube certainly has more recommender surface areas, so they have more opportunities to push people into this content than Netflix does. Netflix has a less stratified, less deep catalog than YouTube does, so that's one piece. But then you just see even with Ms. Rachel, Season 2 performed substantially worse than Season 1 in terms of view hours per minute of content. And so maybe it's just consumer preferences, or they had other stuff to prioritize. Netflix has invested a lot into its own recommendation algorithm, and so it just depends on what they feel the need to promote at any given point in time, and they just have to exclude stuff that doesn't fit that purpose.
Yeah. You have to imagine that someone like Ms. Rachel, if she's doing a deal with Netflix, as part of the deal it's not just going to be money. It's going to say, "Hey, are you going to actually give us impressions so that we have the chance to be successful here and get real viewers that go into the rest of the funnel, show up to the live shows, buy the merch?" It's not enough for you just to pay and stuff us off in a corner; if we do this, we want to do it. I wonder if Netflix has an opportunity to just basically take all the best UGC from YouTube. Because I think YouTube has a massive AI—that's seemingly what they're doing, but underappreciated maybe how big of an AI—AI is probably seemingly a net good for YouTube right now because there's more—
There's more content for all the longtail, like something will happen and—
I'll get served a video and the voice sounds good now, maybe the script sounds okay, and then you realize like, "Okay, this channel was just created two months ago," and it's not really YouTube has an incentive to just constantly be serving new creators.
In theory, Netflix review, I generally want to consume content from creators that have been
are either new and extremely passionate and dedicated. I even saw a creator over the weekend that started years ago using AI voices and just stopped and said
Interesting that YouTube is so flooded with AI content now. He's just
and he's not American, but he makes all of his—English isn't his first language, so he has an accent, but it's appealing because you're like, "Okay, this guy's actually making this content. It's not just like fully generated, right?" And so I think YouTube or Netflix's opportunity is to try to carve out—actually have a filter again and like be a curator and take the best content
that historically would have just been on YouTube and try to bring it over.
Well, yeah, but I mean YouTube absolutely does not want that to happen, right? Keep in mind that the dominant platform for YouTube in the United States is the TV.
Right. And so they are—I've called it a CTV behemoth. That is their dominant platform by view time. And so they compete directly with Netflix for engagement. It's actually really problematic if Netflix is able to just poach their best creators, the top, the sort of the cream of the crop, and then push them into Netflix and then to justify the subscription price, right? So what they've done is they're now inking deals to keep people exclusive for a period of time. And they're also saying, "Look, we're going to punish you if you move. We're going to deprioritize you in our recommendation systems. Promote stuff that's not going to share in revenue." So it's actually what Netflix has done for YouTube to apply a lot of the curation pressure and incentives that YouTube always didn't want to do. They always wanted to resist being that kind of channel. They said, "Look, this is just an open market. People compete and you get views if you outperform." But now they're having to sort of put their thumb on the scale in certain ways to keep the best creators.
Yeah, very much a mask-off moment for the platform that wants to position itself as like, "We're just this friendly platform for creators and we just want to support creators," and they're like, "If you even put one of your videos over here, you're not getting any more money and you're not getting any more views, and we will end your career." That's basically
Be a shame if anything happened to your audience.
Yeah, be a shame if something were to happen to your reach. What is going on at Apple around services ads? Give me a little bit of the history there. Apple's obviously had some sharp words about advertising in the past, wound up building a great ad platform.
So John sent me a screenshot. It looks like it's from Mobile.
It's from Mobile Dev Memo.
