B
Benny Doda
Guest
Muse looks like another AI agent on the surface. When you look closer, you realise it’s a much bigger bet they are making. They are assuming the real prize in AI isn't the smartest model, but the system that knows who you are, understands what you want, and has permission to act on your behalf.
The easiest way to misunderstand Meta's new Muse agent is to treat it as another entrant in the increasingly ridiculous AI model horse race.
OpenAI releases their something-or-other, Anthropic responds, and then Google moves. A benchmark changes and Reddit declares somebody dead, and then X spends 48 hours announcing that one company has achieved AGI and another has become completely irrelevant and we are all cooked. Short race finished, and then everybody moves onto the next race.
Against that background, Muse seemed to arrive rather suddenly. Meta launched it on September 8, and within days it had climbed to the top of the US App Store. Reuters reported 2.8 million downloads in its first 12 days. Muse can manage emails, research purchases, book travel, use financial information, complete transactions, and continue working after you have closed the app. Automation invasion for the normies has finally arrived.
It looked like Zuckerberg had suddenly turned up at the agent party, but if you have been watching closely, he hadn't.
Meta founded its Fundamental AI Research operation in 2013. Meta AI was already embedded inside Facebook, Instagram, WhatsApp and Messenger at scale by the time 2024 finally rolled around. In April 2025, the company launched a standalone Meta AI app specifically designed to remember context and personalise itself using information people had already chosen to share across Meta products. By May that year, Zuckerberg said Meta AI had reached one billion monthly users; how they count that is very interesting. Later in 2025, Meta announced that interactions with its AI products would become another signal used to personalise content and advertising across connected Meta accounts. This is 4D chess right here.
Then came Muse Spark, Meta's new model family, followed by agentic capabilities that connected Meta AI to email and calendars and allowed it to make plans and complete work. Muse is not the beginning of that story; it’s just the thing that makes you take notice when they deemed it was the right time. I think a lot of the AI industry is still looking at the wrong scoreboard.
The model race matters. It just might not matter as much as we think.
For the one percent of users who spend their lives switching between Claude, Gemini, ChatGPT, Grok, Kimi, and whatever was released last Thursday, model performance matters enormously. They notice a five-point coding improvement; they change subscriptions because an agent suddenly became better at spreadsheets. Most of the planet doesn’t behave like this; they are the exception.’
Meta's Family of Apps averaged 3.6 billion daily users in June. In the second quarter of 2026, the company generated $60.8 billion in revenue, of which $59.36 billion came from advertising. That means roughly 97.6% of Meta's quarterly revenue still came from ads. I recently wrote about Amazon banning access to agents on its platform. It fundamentally ruins the platform economics for Amazon, as they do $68 billion a year in advertising revenue, approximately 9.57% of their total revenue.
That matters because Meta has spent twenty years becoming extraordinarily good at one particular problem: understanding what gets us human beings to look at something, click something, share something, buy something, or come back tomorrow. AI is now making that machine better.
Meta told investors that Instagram time spent grew by double digits year-on-year in the second quarter, largely because of improvements in ‘Feed and Reels’ recommendations. Facebook video time increased 9% globally. Meta says its newer recommendation architecture draws on deeper user history, while its LLMs increasingly analyse content, interests and past interactions to improve predictions about what individual users will care about next.
Zuckerberg was explicit about the advantage during Meta's earnings call. He described Meta as a company that builds consumer products reaching billions of people and, once something works, knows how to scale it to an enormous audience.
That may be the part of the AI race Silicon Valley keeps underestimating, and why I think Meta could still be hugely underrated as a stock, because the best model can win a benchmark, but distribution wins Tuesday morning.
Muse doesn't just want prompts. It needs context.
For a personal agent to actually be useful, it can’t know only what you typed into a chatbot yesterday. It needs to know your calendar, your email, what you bought, where you need to be, what your partner told you, which flight you booked, what you can afford, which people matter to you, and which things you always forget.
Meta's own Muse page encourages users to connect email, calendars, Instagram and other everyday apps. Its Mac application can, with permission, work with files, Messages, Calendar, and Notes. Plaid is providing connections to financial accounts so Muse can offer personalised financial assistance. Meta says 1Password integration is coming. There are more than 2000 connected integrations already. That sounds terrifying when you write it as a list, but it also sounds exactly like what a useful personal agent would need.
And this is where I think Meta deserves more credit than some of the current commentary gives it. Muse runs inside a dedicated secure virtual machine, and they say credentials are stored separately so the model itself can’t see them. Sensitive actions such as sending an email or making a purchase require approval so at least there is an audit trail.
Users apparently can control individual connections. Meta says Muse conversations and the data stored in the VM are not passed to its advertising systems, and a more strongly encrypted Confidential VM is planned in which even Meta would not be able to access the contents.
Those are not trivial protections, but they do not make the strategic implication any smaller. Because Meta does not need to secretly shove your Gmail inbox into an advertising model for Muse to become enormously valuable to Meta. It merely needs Muse to become the place where you go when you want something done. That is why they are running a different race at the moment.
This is where the burner-account problem gets weird.
We have spent the last decade learning to fragment ourselves online.
There is the work email, personal email, spam email, and the email created because some website demanded one. There is the public Instagram, and then the private Instagram and the account you use to look at things you would rather not train your public algorithm on.
Meta itself recognised this behaviour years ago; yep, they were all over it. Facebook now allows users to create up to four additional personal profiles under their main account specifically so different interests and communities can remain separate and, of course, not to tie your alter egos to yourself. WhatsApp can remain separate from Meta's broader account system, although users can voluntarily connect it. In 2026, Meta began rolling Accounts Center into a broader Meta Account designed to manage apps and devices under one identity system. That all makes sense in the social-media era, but… the agent era creates the opposite incentive.
A privacy-conscious user may decide to give an AI agent a secondary email address, a limited calendar, a burner profile, and only a carefully sanitised slice of their life. I know I do it that way. But I’m continually told that every layer of separation makes the agent slightly more stupid.
It can’t remind you about the restaurant reservation buried in the other inbox. It can’t realise that tomorrow's meeting conflicts with the dentist because the dentist is on another calendar. It can’t understand that the person emailing you is the same person appearing in your WhatsApp conversations. It cannot optimise your spending if it only sees one account, and it certainly can’t see that you’ve already dated and blocked that one person on another dating app.
So we are heading towards a fascinating identity problem. At what point does the account you trust with your inbox, calendar, relationships, purchases, finances, and personal history become your real account? Are these becoming a series of best on which combination and stack works best? I certainly don’t want to get arrested in 20 years for trading Lima beans on a black-market lookmaxxing site, because unlike burner email accounts, burner agent-connected accounts decide they want to ‘live on’.
Maybe the future "real account" isn't Facebook, Google, an email address, etc.; maybe it is an agent. The account with the richest authenticated context about you and permission to operate on your behalf may eventually matter more than the hundred accounts sitting underneath it, but hey, that’s a very different internet.
And then there is the slightly absurd AI contradiction.
For several years, the public has simultaneously been told that artificial intelligence may be the most dangerous technology humanity has ever created. This whole AI doomer thing is already getting old, but it’s certainly occupying its fair share of space on mediums like the dying New York Times.
Researchers discuss alignment and Governments discuss catastrophic risk. AI companies publish preparedness frameworks and meltdown philosophical crisis papers, thinly veiled and geared toward regulatory capture. Meta's own Advanced AI Scaling Framework explicitly evaluates risks including cyber capabilities, biological threats, and potential loss of control.
Then the consumer product arrives and essentially says, hold my beer, I’m going to need access to all the things you use, own, frequent, and operate, and I want to be able to fill things out, buy things, work while you are asleep, and do all the things that could actually expedite the very thing that feeds the fears.
That contrast is one of the most bonkers things I have seen in my ample years on the blue marble.
I don't necessarily think it is a bad thing. In fact, I suspect ordinary consumers may eventually reach a more pragmatic conclusion about AI than much of the industry has. People have always traded some degree of control for enough utility. We gave smartphones our locations because maps became useful. We gave cloud services our photographs because having every photograph on every device was useful.
We stored our credit cards in browsers because typing sixteen digits repeatedly became annoying. Agents will probably cross the same line. Let’s face it, we are VERY good at creating lies to ourselves about those trade-offs, and we are excellent at saying, “ah, it will be OK”.
But it is entirely reasonable for someone to look at the wider AI safety conversation and ask why, if the machines are supposedly one capability jump away from becoming HAL 9000, the onboarding screen would ask permission to access their bank account. The industry has never completely reconciled those two messages, and Meta is simply making the contradiction visible.
Zuckerberg may have learned which race actually matters.
There is another reason I think Muse matters. Zuckerberg has spoken openly about Meta's historic strategic weakness: it did not control the dominant computing platforms underneath its products.
In a 2025 conversation with Stripe co-founder John Collison, Zuckerberg said that if Meta arrived at the next computing platform at roughly the same time as Apple, Google or Microsoft, the incumbents would have the advantage because they already possessed the underlying ecosystem. His conclusion was that Meta needed to be earlier and better.
That helps explain the enormous investment in glasses, and it also somewhat helps to explain Muse.
Yesterday Meta announced that Muse is coming to its AI glasses. The progression starts to make sense: social graph, messaging, recommendation engine, AI assistant, personal agent, persistent identity, then an ambient hardware device that can see and hear the world alongside you.
Meta isn’t gunning for the best chatbot; they are trying very hard not to spend the next computing era living inside somebody else's operating system. So all eyes are on attention.
The moat can’t be only about data. It is the feedback loop.
People often say Meta possesses enormous amounts of data, which is obviously true but not particularly insightful because we all know the more important advantage is what happens to the data after it arrives.
Meta sees enormous volumes of human behavior. Its models use that behavior to improve recommendations, because better recommendations increase engagement and more engagement creates more behavioral signals. Those signals improve advertising and personalization, and then those improvements attract businesses and creators, who create more things for people to interact with, and now they are putting a personal agent inside that loop.
Meta has already said interactions with Meta AI can be used as another recommendation signal across connected Meta accounts. Ask Meta AI about hiking and, under the policy introduced in 2025, that interest can help shape the content and ads you see elsewhere within the Meta ecosystem. Muse itself, ‘they say,’ is treated differently; Meta says its VM data and conversations are not shared with the ad system, but the broader strategic direction toward AI-mediated personalisation could hardly be clearer.
The doomers will be quick to state that Meta built an AI to steal more information about you. There really isn't evidence for that claim.
The flip-side version would be that Meta has built an AI that becomes more useful the more accurately it understands you, while Meta already owns one of the largest identity, attention, and distribution networks ever constructed.
I, for one, don’t need to be tricked into giving it context, but I have to make a decision on whether the utility of the offering is great enough for me to volunteer my information. That is a much better business for them to run, and a much smarter race to try and win.
This is the 4D chess move.
Muse may become the world's dominant personal agent, or, yeah, it may not.
Google has some very tip-top advantages of its own, and OpenAI has massive consumer mindshare. Anthropic is strong in professional and enterprise workflows, and Apple controls some of the world's most valuable consumer hardware. Personal agents remain unreliable, security is difficult, and consumer trust is far from guaranteed.
Meta's not-so-squeaky-clean history means some people will quite rationally be more cautious about handing it access to sensitive information. Reuters reported this week that even an experimental Muse feature using human contractors to complete some phone calls generated internal privacy concerns.
Meta doesn’t need to win over all the Frontier-AI obsessed people and declare Muse the winner. The people swapping models every fortnight are interesting, vocal and disproportionately visible online. They are not three and a half billion people.
Meta's advantage is that it already knows how to reach everybody else and in a language they all understand.
The company can survive losing a benchmark or a model arriving six weeks late. It can survive somebody on X declaring its AI strategy dead before breakfast. It has done something similar through mobile, Stories, Reels, messaging, short-form video and multiple rounds of technological panic. It is a king of dying fast and not skipping a beat.
What it can’t afford is to lose control of the interface through which billions of ordinary people encounter the next generation of computing. Muse suggests Zuckerberg understands that perfectly.
The only question that will ever matter is who owns the relationship between the intelligence and the human being using it.
And, eventually, which account the human considers to be the one that actually represents them. If that becomes the real AI race, Meta didn't arrive late at all because it’s been standing on the starting grid for twenty years.