Muse doesn’t really need to be the smartest model in the world. It needs to be smart enough, reliable enough and sufficiently embedded in people’s lives that they start giving it jobs instead of prompts. That’s a completely different product than our current AI apps.
Meta is describing something that can open a browser, fill forms, negotiate, send emails, book travel, continue working after you leave, come back for approvals and eventually transact through checkout rails. Once you frame it that way, this stops looking like “Meta has a chatbot” and starts looking like Meta is trying to put an agent between the user and the internet.
There is a huge difference between AI that produces information and AI that produces economic outcomes. You don’t search for the cheapest electricity contract, you tell Muse to lower your bill. You don’t spend Saturday comparing hotels, you tell Muse roughly what kind of holiday you want and let it come back with something sensible. You don’t open six tabs to sell a car, negotiate with three buyers and coordinate logistics, you tell Muse to handle it and interrupt you only when a decision actually needs you. That is a much more powerful interface. Search was “find me information.” Social was “show me something I might want.” Agents become “go do this for me.” And I think that shift is potentially much more important for Meta than whether Muse wins any model leaderboard.
Meta has spent twenty years monetizing attention. Muse potentially lets it sit much closer to agency. Facebook and Instagram are great at discovering intent before the transaction happens. You like running, then suddenly Nike wants to advertise to you. You start watching videos about Italy, then travel ads appear. Fine. But an agent sees something much stronger than inferred intent because you explicitly tell it what you want. It potentially knows the budget, the constraints, the schedule, the people involved, what you bought before, what you hated last time, whether you are willing to pay more for convenience and whether you want it to ask before spending money. That is a very strange amount of economic context sitting inside one interface. Meta currently says Muse conversations and VM activity are separated from the ad system, so I would not jump immediately to “this just becomes better Facebook targeting.” I actually think the more interesting possibility is bigger than that. If Muse becomes the interface through which people delegate economic activity, Meta potentially owns a new layer between human intention and transaction.
And the monetization is almost embarrassingly obvious if the product works. Consumer AI has always had this weird problem where the infrastructure might cost tens of billions and yet the user sits there wondering whether $20 a month is worth it. Part of the reason is that the ROI is fuzzy. Maybe ChatGPT wrote something better. Maybe Claude saved you two hours. Maybe the answer helped you think. Those can be huge benefits, especially at work, but the consumer often doesn’t see a literal number. Muse can potentially show the receipt. Bills renegotiated: $87 saved. Subscriptions cancelled: $46 a month. Refund recovered: $190. Tasks completed while you were away: 14. Time saved: 9.2 hours. Now suddenly the subscription price stops being the reference point. A $20 AI that answers questions feels optional. A $20 AI that reliably creates $200 of value every month feels cheap. And once consumers get used to seeing AI as something that produces measurable ROI rather than occasional amusement, the willingness to pay ceiling may move a lot higher than people assume.
This is also where Metas distribution becomes scary. A startup has to convince people to discover an agent, download it, understand what an agent is, trust it, connect accounts and then form a new habit. Meta can potentially drop Muse into WhatsApp, Meta AI, eventually glasses, and let people encounter agency inside products they already use every day. Billions of people do not need to understand “agentic architectures.” Someone sends them a message saying “ask Muse to sort it out” and that is basically the onboarding. This is why I think the conversation around “who has the smartest model” is increasingly missing the economic point. Once baseline intelligence is good enough, the winner may be whoever owns the best harness, distribution and execution loop around the model. Meta has developed a product that leverages a lot of its already existing moats.
And that brings us to the part I find really interesting: the harness itself. There has been an obsession with frontier IQ because intelligence was historically the bottleneck. Bigger model, better benchmark, smarter reasoning. But once the models are capable enough for ordinary work, the bottleneck moves outward. Can the AI reliably navigate a website? Can it remember what mattered three hours ago? Can it recover when checkout breaks? Can it understand that the cheapest flight isn’t actually the one I want? Can it notice conflicting instructions? Can it ask me when it truly needs me and otherwise stay out of the way? Can it coordinate several subtasks without forgetting one? Can it survive a malicious webpage trying to hijack its instructions? That is not simply a model IQ problem. It is a training environment, harness and reinforcement learning problem.
Which is why the possible Innodata connection is interesting. Hunterbrook published a piece arguing that Meta may be Innodata’s largest customer and that Innodata’s new long horizon agent work could be related to Muse. Important caveat: neither Meta nor Innodata has publicly said “Innodata trains Muse,” and Hunterbrook disclosed a long position in $INOD, so this should be treated as a circumstantial thesis rather than fact. But Innodata itself has said something unusually specific: it won a significant program with its largest customer involving personalization of long horizon agents, and separately a program creating reinforcement learning environments for computer use agentic tasks. Put that next to Meta’s description of Muse: personalization, long horizon work, computer use, autonomous execution, agentic harnesses. You don’t have the missing confirmation, but you can see why people are connecting the dots.
The bigger $INOD angle is that the nature of the data problem may itself be changing. The first phase of AI ate the internet: books, webpages, code, papers, Reddit, everything humans had written down. Then post training became important and the industry needed better answers, preference data, domain expertise and human evaluation. Agents introduce another category entirely. You cannot teach an agent how to successfully run pieces of somebody’s life by scraping another trillion tokens. It needs situations. Fake inboxes, browsers, travel sites, broken checkouts, conflicting calendar events, malicious prompts hidden in pages, ambiguous instructions, changing websites, preference conflicts, error recovery. The model has to enter those worlds, act, fail, get feedback and try again. The old data company provides answers. The new data company potentially builds little universes where agents learn how to act.
And this is where Muse succeeding could actually make the Innodata opportunity bigger. The intuitive model is that Innodata helps train Muse, Muse launches, project ends. But if Muse works, Meta expands it into travel, commerce, personal finance, small business administration, insurance, healthcare bureaucracy, procurement, customer service and whatever people decide to delegate next. Every category adds thousands of workflows. Then personalization explodes the state space again. Booking an airline ticket is easy compared with booking my airline ticket. Window seat unless it’s short. Don’t land after midnight. Lisbon is fine but Porto can work if the savings are large enough. Never a 50 minute Frankfurt connection. Remember that my partner needs checked baggage. Pay slightly more if the cancellation policy is flexible. Multiply that by hundreds of millions of people and “personalization of long-horizon agents” starts sounding less like a one off training job and more like permanent infrastructure.
That is what makes Innodata potentially interesting here. The company has already been showing improving economics, higher value work and more reusable datasets/IP, while talking more openly about reinforcement learning environments and agentic workloads. The key question is whether this becomes a repeatable, IP heavy business or remains mostly expensive outsourced labor. If it becomes the former, then the market may eventually stop thinking about $INOD as a data services company and start thinking about it as something closer to agent training infrastructure. The layer that continuously constructs environments, reward systems and evaluation loops that teach these agents how to behave in the messy real world. That is a very different narrative.
And this loops all the way back to Meta’s capex. For years the question has been where the return on all these GPUs actually comes from. Better ads were the obvious answer: better recommendation systems, better targeting, AI generated creatives, more engagement. Muse potentially opens a completely different answer. What if the giant compute buildout isn’t just about making Instagram marginally more addictive? What if Meta is trying to build the default execution layer for ordinary human life? WhatsApp becomes communication. Instagram becomes taste and discovery. Facebook provides identity, communities and marketplace infrastructure. Meta’s models provide intelligence. Muse provides persistent agency. Secure VMs give each user a computer. Payments provide transaction rails. Glasses eventually provide an always present interface. And underneath it all, somebody continuously teaches the system how to operate in an increasingly complicated digital world.
That is why I think Muse matters. Maybe it is the smartest model, maybe it isn’t, doesn’t really matter. The important question is whether normal people begin to say “Muse, deal with this” and then stop thinking about it. If that behavior sticks, we have moved from software that waits to be operated to software that operates the economy on our behalf. Meta has spent decades owning where people communicate, discover and express themselves. Muse is potentially an attempt to own the next step: intent becoming action. And somewhere underneath that shift, companies like Innodata may discover that training AI to answer questions was only the warm-up. The much larger job might be teaching billions of agents how to actually live in the world.
Leave a comment