Meta's Muse beat OpenAI to market without a frontier model

1. Instinct's two tricks — ▶ 5:52

   Instinct got off the ground by skipping new apps: sign up, then message it like a friend or a human assistant over existing messaging. The second trick was being proactive. LLMs create a "blank page problem", since you can ask anything and don't know what to ask. Older products also had a "boil the ocean problem", trying to do too much. Prompts fix both: it flags a message in your inbox and offers to reply, or spots a bill due in 10 days with no recurring payment and offers to set it up.

2. No single killer use case — ▶ 9:45

   Instead of nailing one thing like movie tickets, these products should do 5-10 tasks, some resonating with one user base, some with another. Users who get no immediate payback move on. Still, there needs to be an "it did this for me" story, because word of mouth carries the trust these products need. The clearest hook so far is money: review my expenses and credit card bills, find savings, tell me which card to use for which bill. All in natural language, with no new services to sign up for.

3. Agents threaten set-and-forget revenue — ▶ 12:03

   Alex notes bankers argue banks stay solvent because people leave cash in non-interest-bearing accounts. Agents that move it would hit the banking industry. Siegler goes further: many subscription businesses depend on people never cancelling. An agent that says "you haven't watched HBO Max in a while" and turns it off for the month breaks that. It hits every publication and newsletter, YouTube included. His answer is a whole new industry: annual contracts, breakage fees, ways to keep people incentivized.

4. Here to stay, timing unclear — ▶ 14:27

   At least one, and probably a handful, of these assistants will be used by most people. The timetable for mass adoption is unknown, down to when "our parents will be using them". The risk is churn. People sign up to see what it can do, try the obvious tasks or what a friend suggested, then have no idea what else to ask. Prompting is the fix, and the early companies understand it. He expects the features to get baked into AI services people already use.

5. Instinct's Switzerland play — ▶ 17:07

   Instinct reportedly raised three rounds in a few months and may be valued at $10 billion, with a young CEO from Sierra. Its best argument is neutrality: Siegler expects information silos where companies refuse rivals' agents. Amazon blocked Muse despite other Meta deals, with Alexa Plus and an ad business that agents sidestep as motives. Shopify agreed to work with Muse. But Instinct has no billions to spend, depends on others' models and clouds, and lacks built-in distribution. Gmail and Calendar access is also essential, and Google could cut it off.

6. Muse surprised him — ▶ 21:55

   Siegler isn't a Meta fan and dislikes what Instagram and Facebook have become, "ad vessels". Still, he was surprised how well Muse was done. After spending billions on the Scale deal, infrastructure and acquisitions to rebuild in about a year, Meta shipped a standalone assistant before OpenAI. Each user gets a private virtual machine, and there is an added security layer, because Meta knew people wouldn't inherently trust it. WhatsApp integration came on day one, and Muse runs fully through WhatsApp, with no new app required.

7. Cloud computer versus desktop — ▶ 27:10

   Meta answered users spooked by their computer's contents living in the cloud with "it's your computer", download it if you like. OpenAI is leaning that way too. But an Excel sheet that lives on your desktop needs native apps and OS-level APIs, which is why Anthropic has done so well. Microsoft takes enterprise, Meta consumer. OpenAI straddles with cutesy mascots on a work-oriented tool, which Siegler reads as whiplash after Anthropic passed it on business and valuation. Whether people want one AI for work and home is open.

8. Reach versus trust — ▶ 32:52

   Meta has 3 billion-plus users. OpenAI has 1.2 billion weekly users, per Sam Altman, and Dots needs the $100-a-month Pro tier. Alex upgraded from $20 for it, yet won't give Meta his Gmail and calendar. Siegler got many messages like that, and calls the concern real even if his audience skews tech. Negative Meta and OpenAI movies are about to land. He sees a big first-mover advantage that hurts startups like Instinct and Towns. His bet: one of the big players, whichever nails integrations without being blocked.

9. Anthropic's missing assistant — ▶ 40:33

   Anthropic led this moment with Claude Code and Co-work, then didn't build a personal assistant. Siegler guesses it expects Co-work to evolve into one, and notes it has always shied from consumer. It has no image generator or video product, and only minimal voice. That looks smart now, but images become table stakes: a presentation needs them. He expects it to be dragged in. It just merged the Claude and Co-work text box, which OpenAI hasn't, toward a single box that routes personal and work requests behind the scenes.

10. Frontier models may not matter — ▶ 46:16

    Muse runs on Muse Spark, not what Meta considers a frontier model. A bigger one, Watermelon, is promised. Siegler wonders if the frontier matters mainly for science and hard discovery, while cheaper models handle day-to-day agent tasks. On Apple, he says it could win AI by default, with mail, messaging, calendar, pictures and the screen built in. Its problem is the reverse of Meta's: users "forget that it's actually good now" and keep opening ChatGPT. Apple may also add Mac permission prompts warning users off giving Meta access.

11. Google can't get out of its way — ▶ 51:57

    Google has Gmail, Calendar, search, talent and DeepMind, yet Gemini Spark is "fine", not socks-off. At I/O it pre-announced 3.5 Pro, which never came; Gemini 4 is now in security testing. Siegler blames "too many cooks in the kitchen": Cloud, Android and Search have different incentives, and Spark, the family product CC and AI Mode overlap. He allows Google may be on top again in a month. But then comes trust: why hand over your agentic life to a company that may fall behind again in six months? He expects Instinct and Towns to be acquired.

An assistant is only as good as the doors it can open. Mail, calendar and messaging are where a personal assistant earns its keep, and most of those doors belong to someone else. Gmail and Google Calendar serve billions of people. Connect them and these products work. Leave them unconnected and, for most users, they don't. That makes access the real battleground, and access is already being used as a weapon. Amazon blocked Muse, even though Meta has shopping and advertising deals with Amazon. The likely reasons are that Amazon has its own assistant in Alexa Plus, and an advertising business that agents sidestep, because agents are not moved by the ads in a feed. MG Siegler expects information silos to harden. Companies will not let rivals' agents run on their services. Google, which has been liberal about who plugs in, could decide Muse is next.

This is why the startups are structurally weak, however good their product. Instinct moved first and moved smartly. It skipped the app and let people message it like a human assistant. It then fought the blank page by being proactive: a bill due in 10 days, an unanswered message, an offer to handle it. VCs loved it, and Siegler jokes that tweeting love is one way to get a foot in the door of a deal. By reports Instinct raised three rounds in a few months and may be valued at $10 billion. But it has no billions to spend. It depends on other companies' models and clouds. It has no built-in distribution. Its one real argument is neutrality, the Switzerland play, in a world of silos. Siegler thinks that argument has a shot. He still expects the winner to be one of the big players. The first-mover advantage goes to whoever signs up the most people, and he expects acquirers to come for Instinct and Towns.

Reach, then, favors Meta and OpenAI. Meta has 3 billion-plus users across its products. OpenAI has 1.2 billion weekly users, according to Sam Altman at the Dots unveiling. And Meta moved first. Siegler is not a Meta fan, and he says he was surprised by how well Muse was made. Meta spent billions to rebuild its AI effort in about a year. It shipped a standalone assistant before OpenAI did, with a private cloud computer for each user and WhatsApp integration from day one. Muse runs fully inside WhatsApp, which makes the install problem vanish. And it did all this without what Meta itself calls a frontier model. Muse runs on Muse Spark. A bigger model, Watermelon, is promised. Siegler floats the idea that day-to-day agent work may not need the frontier. The frontier would be reserved for science and hard discovery, while cheaper models handle the errands. That is a hypothesis, not a finding. Alex's own pick is OpenAI, with Apple as the dark horse. Siegler names no single winner, only the big players.

Reach cuts the other way when trust is the product. Alex upgraded from $20 to $100 a month to use Dots, which is limited to OpenAI's Pro tier. Yet he won't give Meta access to his Gmail and calendar. Siegler got a flood of the same message after praising Muse. He allows that a tech audience skews skeptical, but calls the concern real. Meta's counter is design. When people were spooked by the idea of their computer's contents sitting in Meta's cloud, the answer was, in effect, "it's your computer". You can download it and do what you want. Apple has the opposite problem. Siri is now good, with built-in access to mail, messaging, calendar, pictures and whatever is on your screen. But users remember more than a decade of failure and keep opening ChatGPT. As Siegler puts it, "you just forget that it's actually good now." Trust deficits come in different shapes.

Where the assistant lives is a second fault line. Muse bets that a cloud machine can do most of what people want. Alex's counter is that real work sits on the desktop, and that Dots takes over your browser and your machine. Siegler agrees an Excel sheet that isn't in the cloud is a problem. That is part of why Anthropic has done so well, and why these services built native apps that reach into operating-system APIs. Microsoft is taking the enterprise lane and Meta the consumer one. OpenAI straddles both, with cutesy characters wrapped around a work-oriented tool. Siegler reads that as the whiplash of a company that built a consumer behemoth and then found Anthropic ahead on business and valuation. Whether people want one AI for work and one for home, or one to rule them all, is open.

Two companies with obvious claims are missing. Anthropic led the shift toward handing control to AI, with Claude Code and Co-work, and then stayed out. Siegler's guess is that it sees Co-work growing into the assistant, and that it has always shied from consumer. It still has no image generator, no video product, and only minimal voice. He calls that smart now and table stakes soon, since a presentation needs images. Google is the harder puzzle. It has Gmail, Calendar, search, talent and DeepMind. Gemini Spark is "fine", not socks-off. At I/O it pre-announced 3.5 Pro, which never came, and Gemini 4 is only now in security testing. Siegler's diagnosis is "too many cooks in the kitchen". Cloud, Android and Search have different incentives, and Spark, a family product called CC and AI Mode overlap. Alex's line is that if you have everything, people will use nothing. Siegler hedges that Google may be back on top in a month. Then it faces the trust question: why hand your agentic life to a company that may fall behind again in six months?

The strongest objection is to the premise that people want this at all. Siegler says at least one, and probably a handful, will be used by most people. He can't say when parents will use them. Churn is the threat. People sign up, do what a friend suggested, then don't know what to ask. His fix is prompting and a spread of 5-10 tasks, with saving money as the clearest hook. That success carries its own cost. An agent that notices you haven't watched HBO Max in a while and pauses it attacks every business built on people forgetting to cancel. Alex adds the banks and their idle cash. Siegler's answer is a new industry of annual contracts and breakage fees, which is a guess. And whoever builds the agent that cancels your subscriptions will need permission from companies whose revenue depends on you not cancelling.