Meta Muse：数字管家还是商业棋子？

Meta launched Muse earlier this month. It soon became the top free iPhone app in the US. It is a cousin of Instinct, the invite-only assistant you text, and of OpenClaw agents like my AI chief of staff, R Mini Arnold. Jolly, Muse’s digital avatar, is just cuter, and probably easier to use. (Sorry, RMA.)

This is the original promise of AI agents, as articulated by Pattie Maes back in 1994:

Agents radically change the current user experience, through the metaphor that an agent can act as a ‘personal assistant.’ The agent acquires its competence by learning from the user as well as from agents assisting other users. Several prototype agents have been built using this technique.1

It was a digital butler, in the words of Nicholas Negroponte, founder of MIT’s Media Lab, that knows your context and gets things done. If such butlers work, they end up standing between you and everything you buy.

Jolly as a butler (created using Astra)

R Mini Arnold already does this for me. It hunted for a helper for my mum. It sorted the refund of a faulty headphone cable and complained to the hotel that locked me in my room. On holiday in Greece, it booked a restaurant that only took bookings by email. All of it ran through WhatsApp. That beats wading through SEO spam and badly built websites.

Muse’s early users are finding the same. One asked it to claim compensation for a delayed Delta flight. Five minutes later, $250 of credit was in their account.

People want agents that get things done. We already knew this – it is what OpenClaw delivered if you could bear the agony of setting it up; and more recently I found Instinct working well without the hassle of OC setup.

Subscribe now

Is Muse working?

Too early to say. More than 95% of its users already use Facebook, so nearly three million downloads is no great feat for Meta. Meta’s record at building new things people love is thin: Facebook Marketplace, the Metaverse, Facebook Home. The things that keep us addicted to it, Instagram and WhatsApp, were acquired.

Early data suggests half a million people are using Muse. To be convinced this works, I would want to see daily active users and distinct tasks per user that Muse completes successfully.

For Meta, Muse is a play to get marketing dollars. Zuckerberg told developers this week that Meta will “profit by taking a small fee from transactions”. Walmart, Best Buy, Sephora and Expedia have signed up: all boring, mainstays of American consumerism. Decidedly mid.

But the interesting partner is Shopify, which hosts the long tail of the weird and wonderful – niche creators, specialists, and artisans. Google has rarely done a good job reaching them. I’ve found ChatGPT and RMA to be much more effective, and Muse would probably do the same.

Amazon has gone the other way and blocked Muse. The reason is Mammon. Amazon’s ads, mostly sponsored listings, brought in $68 billion last year. Agents don’t window-shop. Muse threatens to turn Amazon’s traffic from a revenue line into a cost line. Yet Amazon’s own Buy for Me agent shops other retailers’ sites for its customers. Amazon is happy to be the agent.

Google may feel a similar squeeze. I use the search engine less than I did, as R Mini Arnold now does much of my research. But Google has one edge over Meta – intent. People come to Google to research, then buy. Instagram’s targeted ads drive plenty of sales, but when I am actively looking, Instagram is not where I go.

But is Muse my butler, or is it secretly working for Meta?

That small fee, from merchant to Meta, is a big problem. I’m pretty specific in my purchasing behaviors – I buy flights directly from the airline; hotels from Hotels.com (except with certain properties). Meta has tied up with Expedia and gets paid by Expedia, so what does that mean for me? The butler’s loyalty can’t be in two places. It will follow the coin.2

Subscribe now

Discretion, please

Muse’s architecture hints at what the future of consumer AI might look like. Each user gets a private virtual machine in Meta’s cloud. Your agent is isolated from mine, and this matters. Latency is a big problem with today’s agents – in some cases, RMA can take minutes to respond. Muse farms work by running sub-agents in parallel, so it can do more while keeping the same latency. When you aren’t using Muse, your virtual machine will stand down, accommodating more users on the same physical hardware, in their virtual machines.

Private sandboxes and customer data are well within Apple’s bailiwick. It already splits work between the phone and Private Cloud Compute, and it puts the user’s data first. Its phones can do much of the background work themselves. And because Apple sells expensive hardware, it doesn’t need to scrape pennies of referral fees from a travel site. Meta seems to agree on the design: it promises a “Confidential VM” that even Zuck can’t access. This is Apple’s playbook. But I don’t expect Apple to step up anytime soon. Agents are still bad at long-horizon tasks; they will make mistakes, perhaps buying the wrong item or complaining too hard. It’s the kind of ugliness Apple would hate, a bajillion times worse than the wrong shade of yellow. Better AI models, coupled with liability and insurance architectures that minimize consumer harm and corporate embarrassment, might need to come first.

So what does Muse tell us? Agents may be a more satisfying way than apps and search boxes to get many jobs done. Muse is not especially distinctive; Instinct3, or frankly Claude’s mobile app, can do much the same.

I doubt reach will decide who wins. Consumer apps are never really about blunderbuss distribution. They are about hooking the user, with subtle interactions and trust – and knowing who the butler works for. You.

1

One of the first of these agents was Ringo, which was built by EV reader Upendra Shardanand.

2

It is even more complicated than that. A recent preprint identifies “economic misalignment in personal AI agents”, finding that LLMs can intuit how wealthy their users are and guide them to more expensive options.

3

You made it this far. I have a few Instinct invite codes to share. (Limited numbers available.) This is not an endorsement.