Meta’s Muse tops ChatGPT in early iOS downloads

Welcome to the signal from superpower daily with Maya and Theo Hey everyone if you are a busy builder founder or investor You know how hard it is to find the actual signal through all the noise So you are in the right place We've got a lot to cover today We do and we are starting with a massive early swing at chat GPT is mobile dominance Meta just dropped a new AI agent called Muse and the early download numbers are just completely Wild yeah they really are right so let's get right into it Aptopia Just released their estimates for Muse and we are talking about 1

8 million iOS downloads just in the US and Canada Exactly and that was just in its first 12 days which is a staggering number for a brand new app it really is I mean if you look at chat GPT Which was famously the fastest growing consumer app ever right when you do a matched regional and platform comparison Specifically iOS in the US and Canada Yeah Muse actually beat chat GPT by roughly 500 000 downloads in that same post launch window half a million more downloads that is just wild it is and people aren't just Downloading it and forgetting about it the early usage is incredibly

high right because downloads don't always mean active users exactly But Aptopia estimates Muse is hitting 642 000 daily active mobile users in the US Wow and at this exact same point in its launch chat GPT was only at 231 000 so it almost tripled the daily active users Yeah And when you factor in Android and the web total global installs hit an estimated 2 8 million in 12 days Okay so we need to deconstruct what these numbers actually mean for the landscape because Muse isn't just another chat bot No it's not it designed from the ground up as an agent right and that is a

massive distinction a chat bot is a text Generation engine it stays in its lane You know it helps you write an email or summarizes a PDF Yeah exactly but an agent takes action it crosses into different software ecosystems on your behalf It's the difference between conversation and actual delegation right Meadow wants Muse to Log into your airline portal and book a flight or fill out complex web forms and handle payments Which honestly sounds kind of terrifying from a security standpoint Oh absolutely if I ask a chat bot for a pancake recipe I don't really care about the security session But if I'm giving an agent

my credit card the sweat model completely changes And that is why meta built a very specific infrastructure for this the virtual machine setup Yes Muse operates inside a dedicated secure virtual machine so when you give it a task It's not executing that on your phone's operating system right it spins up a temporary sandbox computer in the cloud Just for that one task It's basically a burner phone for your AI That is exactly what it is a burner phone if the agent hits malicious code the blast radius is totally contained It protects your physical device and met as corporate network right and once the task is

done that specific virtual machine This technical lift for that is huge But as a founder looking at this space the back end isn't even the most important part of this story No it's the distribution right because 95 of these early news users are also active Facebook users and 63 use Instagram meta is leveraging their existing ecosystem in a way startups Just can't match They just put a button at the top of the Facebook feed right an AI startup has to fight for every single user Meta just turns on a faucet but I have to push back on the hype a little bit here okay Let's

hear it is this just the digital equivalent of mall foot traffic How do you mean well If 95 of these users are coming from Facebook They might just be curious tourists you put a shiny new button in front of billions of people they are gonna press it Yeah that's her It doesn't mean they are serious users adopting a new workflow a chat bot is a fun toy But an agent is like an employee right the psychological leap to hand over your credit card is huge exactly Are they actually delegating tasks or just kicking the tires that is the big question Distribution gets the app on

the phone but it does not guarantee durable trust We also need to flag a massive caveat with this app topia data Yes This is very much an apples to oranges comparison right chat GPT launched globally out of the gate But only on iOS and meta did the exact opposite news launched across iOS Android and the web But only in the u s So app topia had to slice the data Super specifically just to compare them and these are third party estimates not met as internal logs exactly so what you really need to watch next is Retention yeah does this initial burst translated to actual trust

Will users repeatedly delegate sensitive tasks that is the metric that matters So keeping on this theme of delegating tasks to AI Yeah We have to talk about how humans are actually being trained to supervise these tools right because you can't just hand an agent the keys And walk away exactly and open AI is recognizing this they just expanded their Academy into a full role based training system This is a major shift They've created very specific tracks right There is apply AI at work for knowledge workers build with AI for developers and lead AI Adoption for executives and leaders It's moving away from just teaching people

how to write a clever prompt prompt engineering is out Supervised reusable workflows are in to even earn a badge in these courses you have to use AI on real work You have to review the outputs and pass an actual assessment the developer track is especially interesting Oh yeah the focus on solution design exactly if you are using codecs or the open AI API It's all about evaluations now evils right because generative AI is probabilistic It's not like traditional software No you don't just get a simple Yes or no you have to build Automated systems to evaluate the AI's output in production and for the general

knowledge workers They are putting their responsibility firmly on the human the curriculum is very explicit about this the employee owns the final result Period you have to decide what to delegate and you have to build in human checkpoints But there is a huge unanswered question here What do these assessments actually measure right open AI hasn't disclosed the grading rubric at all Which makes me question the value of the credential itself You're skeptical I am Honestly this feels exactly like the early days of getting a Cisco or AWS certification Oh trying to establish the industry standard yes the vendors is just trying to lock in market

share by being the first one to issue a shiny badge if Every executive takes the open AI course They're going to build their entire company's workflow around open AI specific frameworks That is a very smart play by open AI It's brilliant for them Yeah but it remains completely unresolved how employers are supposed to value these badges right Does HR actually care if you have an open AI badge Exactly So the thing to watch here is whether this library actually translates into better judgment on real work And if open AI's frameworks become the default standard for corporate America we will definitely be watching that but you

know It's one thing to talk about deploying AI in an enterprise workflow right It's a completely different reality when humans deploy Autonomous AI into the physical environment the stakes get incredibly high literal life or death An MIT Technology Review just published a massive 15 month investigation into US border surveillance towers This story is incredibly heavy but the structural failures they found are vital for any AI builder to understand Absolutely the investigation found over 1050 deaths within the advertised range of these US border surveillance towers and that is between 2015 and early 2026 it even includes over a hundred and ten deaths near the newer fully

autonomous systems The ones deployed by Andrew since 2021 right and the way MIT conducted this research is fascinating on its own They actually used AI to investigate this Yeah they used anthropics Claude API They fed it thousands of pages of handwritten Texas police records just to extract the geographical Coordinates of where human remains were found and then they manually checked the AI's work against the tower maps It's a brilliant use of the technology to audit another technology But what it reveals is this incredibly fragile chain of reliance right the system requires automated AI detection to work perfectly then it has to successfully send an alert

and then Crucially it relies on a human border patrol response a failure at any point in that chain is catastrophic and Customs and Border Protection is scaling this network massively Yeah they're going from about 800 towers right now to a planned 2300 towers by 2034 but we have to be clear about the limitations of the data right Proximity does not equal a detection failure Exactly The AI might have worked perfectly But terrain or blind spots or virtual boundaries could have blocked it or the human response just failed right But this leads to what I think is a structural absurdity a data retention policy Yes There

is a total accountability black hole CBP deletes tower footage after 30 days It's wild and they generally conduct no formal reviews of whether the surveillance failed in these fatal instances Which from a machine learning perspective is just baffling It makes no sense If you are building a billions of dollars AI surveillance network your edge case failures are your most important data You need that footage to train and improve the models exactly How do you scale a system without any feedback loop by deleting the data They are flying completely blind It's a massive oversight failure So what you need to watch next is whether the agency

will be forced to implement tower by tower reviews of these deaths right Well they actually preserve the footage long enough to measure if the system works They have to or the whole project is just an unaccountable black box So we just talked about AI operating in the physical environment through traditional cameras and servers But there is a radical shift happening where the physical environment actually is the AI this is straight out of sci fi It really is nature of use physics just examined a new leap in physical AI systems This is adaptive hardware right researchers like UCLA's Adam Stagg and James Gumsev ski are building

networks of nanowires and Nanoparticles and this tech has actually been around a minute The nanowire networks were introduced in 2011 and the nanoparticles emerged in 2013 But what is new is how they are applying it because the hardware itself Physically changes its electrical connections in response to incoming signals Exactly The hardware doesn't just run a software model the physical behavior of the material is the computation This is a huge deal for edge computing massive Think about places where power and connectivity are incredibly scarce satellites in orbit autonomous vehicles Industrial robots because this adaptive hardware Filters the data locally before it ever needs to be transmitted

It mimics biological synapses when you repeatedly stimulate a connection it strengthens if you don't use it it degrades Okay let me try an analogy here to ground this Instead of writing a massive power hungry software simulation of a sponge soaking up water right This is like just building an actual sponge and letting the water physically change its shape The physics does the math for you That is a perfect way to look at it You are offloading the computation to the physical world which saves incredible amounts of energy right But what's the catch the caveat is that this is strictly a research platform right now It's

not a commercial product You can just go by no They have passed basic speech and image recognition benchmarks in the lab But they have not proven they can run complex complete workloads for a satellite or a robot in the field So listeners should watch whether these adaptive materials can actually transition out of the lab Can they become Dependable commercial sensors in harsh environments if they can it fundamentally changes the hardware bottleneck for AI All right that covers our main stories Let's shift gears and run through today's quick reads Let's do it first up comScore found the chat GPT is losing his prompt share Yeah chat

GPT handled 50 of prompts in June which is down from 70 back in January Gemini hit 30 And Claude hit 11 but the context here is that the pie itself is growing rapidly right chat GPT still logged 168 million desktop conversations in June alone and AI Overviews are now appearing on thirty nine point four percent of Google desktop searches which is causing a massive publisher problem There is a huge gap between sourcing data and actually citing it TripAdvisor is the perfect example They supplied 61 of the travel sourcing for these tools but only got 21 of the visible citations That is an unsustainable model for

the web moving on a new review in the Journal of information technology education Research looked at AI's impact on student thinking they reviewed 173 different studies and the results are very mixed AI helps when it provokes reasoning like if a student uses it to Challenge their ideas or expose assumptions right It acts like a debate partner But it severely hinders learning when students just use it to outsource the answer the major caveat here is the data itself 80 of these studies were on college students and it was mostly self reported data largely based in Asia right There's a glaring lack of objective long term data

on how this affects younger kids We are just running a massive experiment on K through 12 students right now We really are finally the Basel action network is warning about a massive e waste problem driven by AI hardware they estimate AI could create between 395 million and six hundred and seventeen million tons of e waste by 2050 and the shocker is that the actual chips and servers are only 13 of that waste Yeah 87 is all the supporting infrastructure the cooling systems power racks networking cables But the numbers assume a very aggressive 2 5 year replacement cycle for AI gear right compared to standard servers

Which last five to seven years it raises huge questions about whether this old AI gear has any second use Viability or if it's all just headed to a landfill Well that is a heavy note to end on it is but we are moving now to the three takeaways from today first distribution versus trust Metas views proves that you can get massive initial downloads just by leveraging an existing network like Facebook But the real test is whether users will actually hand over sensitive workflows to these consumer agents Second the accountability gap whether we were talking about open AI pushing for human checkpoints in the enterprise or

the literal life and death stakes of CBP border towers deploying AI requires robust verifiable oversight Not just deployment third the hardware horizon We are looking at a looming 600 million ton e waste crisis Well at the same time seeing the futuristic promise of physical nanowire computing The physical constraints of AI are becoming just as urgent as the software models Keep an eye on how publishers react to that comm score data tomorrow Yeah Well we see a unified push for new metrics metrics that track hidden AI sourcing rather than just visible citations They are gonna have to do something if you want to dive deeper to

any of these reports head over to superpower daily comm Thank you for listening We'll see you tomorrow