Muse and the Coupon

The Bhaja Govindam says artham anartham bhavaya nityam — contemplate wealth as anartha, not merely un-wealth but trouble, calamity, the thing that arrives carrying its own headache. Traditionally this hymn belongs to the world of Adi Shankaracharya, although the precise authorship of individual later verses is murkier than the devotional packaging suggests. It was not the Buddha whispering fiscal advice to Ananda, which would have made the accounting considerably tidier. But I have considered it. I have considered it while comparing the price of Amul butter across three apps, while calculating whether the Swiggy delivery fee negates the discount, while wondering if the artham I pursue is itself the anartham I dread. Meta released Muse on September 8, and within roughly two weeks more than 2.5 million souls had downloaded it, pushing it to the top of the American app charts, and I understand precisely why. They did not download it to discuss the categorical imperative or the finer points of quantum entanglement. They downloaded it because intelligence is impressive, but agency that leaves fifty dollars in your bank account is immediately comprehensible. The sublime is nice. The coupon is better.

I am the target demographic. I admit this without shame, which is to say with enormous shame that I have disguised as postcolonial critique. When money is scarce — and money is always scarce, money is a state of being, a guna, a quality of the universe like heat or viscosity — twenty percent off is not gamification. It is arithmetic. It is the difference between the pressure cooker and the pressure. I spent fifteen years in American healthcare research, where sooner or later everything seemed to acquire a denominator, a utilization rate, a cost, a probability, some brutal little algebra explaining why one human outcome was affordable and another was regrettably outside the budget, and I returned to Calcutta with a PhD in parsimony. I know the exact exchange rate between dignity and discount. It fluctuates daily.

Muse promises to navigate this exchange. It is an agent, from the Latin agere, to do, to act, to drive forward, and here is the etymological joke that the fat textbooks conceal: we call it artificial intelligence, but intelligence merely knows; the agent acts. It comparison-shops. It can hunt for forgotten subscriptions and discounts, negotiate with service providers, chase refunds, fill forms, make purchases, send emails, book things, and generally perform those tasks where the limiting human resource is not IQ but the willingness to spend forty-seven minutes clicking through menus designed by a committee that hates you. It does the things that require persistence, the digital equivalent of standing in queue at the Calcutta Municipal Corporation, which I have done, which my father did, which his father did, except the queue is now virtual and the clerk is a large language model trained on the entirety of human frustration.

And oh, the shopkeepers have noticed. Amazon, that Bezos-built bazaar, that everything-store, that digital maidan where a considerable fraction of humanity now buys its toothpaste, blocked Muse from shopping on Amazon.com with the indignation of a bania who has discovered the customer’s new friend has entered through the back door and begun negotiating. Amazon says Meta never obtained authorization, did not tell Amazon beforehand that Muse would browse its store, and that the agent did not identify itself while browsing. Amazon also raised concerns about how credentials were being handled. Meta’s own technical account gives a rather different picture: login credentials are supposed to travel directly into a protected credential store inside each user’s dedicated cloud machine and remain invisible to the main Muse agent. So we already have the first theological schism of agentic capitalism: Amazon says get your bot out of my shop; Meta says the bot cannot even see the password. Somewhere, lawyers are achieving enlightenment.

Shopify, contrarian as a Bengali intellectual at a dinner party, has gone the other way and is integrating Muse with Shop Pay. PayPal has announced an integration too. Walmart is building its own agentic shopping machinery, including Sparky, while also integrating its commerce with outside AI systems such as ChatGPT and Gemini. One retailer builds walls, another builds doors, another builds its own elephant and allows other elephants into the courtyard. Nobody yet knows whether this is openness, strategy, desperation, or simply Tuesday in American capitalism.

Behind this technical squabble lies an old commercial question wearing new silicon clothes: who owns the customer? If my AI searches the whole market before buying my toothpaste — if it crawls the bazaar in its entirety, comparing the kirana shop to the mall to the dark depths of Alibaba — then Amazon loses some of the enormous advantage of being the front door. Search results, sponsored placement, recommendations, bundles, Prime, the whole beautifully engineered architecture between desire and checkout suddenly has another creature standing in front of it saying: no, Suvro, the same bloody thing is nine dollars cheaper over there.

The gatekeeper becomes the gate. The middleman discovers he is merely the middle. This is not merely disruption; this is commercial dharma developing a nervous disorder, the varna of online retail dissolving into a puddle of APIs, permissions and price-matching algorithms.

But here is the terror, the bhaya, the cold sweat that arrives at 3 AM when the excitement recedes: to make Muse genuinely useful, I must let it into my life. Email, calendar, shopping habits, subscriptions, accounts, preferences, the accumulated karma of my consumption patterns, the data that reveals I am cheap about one thing and irrationally particular about another. Meta says every Muse user gets a dedicated virtual machine in the cloud. Sensitive credentials are kept separately from the agent. A separate permission system called Sentinel controls sensitive actions. Purchases and emails can require explicit approval. There is an audit trail. Meta also says Muse conversations and VM data are not fed into its advertising systems.

This is not nothing. It is substantially more interesting security engineering than “trust us, bro,” which remains the default architecture of much of the Internet.

It is also not the same thing as Meta being technically incapable of seeing the data.

Meta explicitly says that the ordinary Muse architecture restricts employee access through policy but does not cryptographically prevent Meta from accessing information when necessary to support, secure or operate the service. A more ambitious “Confidential VM,” designed to prevent even Meta from reading the contents, is supposed to come later. Muse interaction trajectories can also be used to improve the underlying model after sanitization intended to remove key personally identifiable information, unless the user opts out.

So the architecture is not simply Meta has all your secrets, and it is certainly not Meta can never see anything. Reality, the tiresome bastard, occupies the uncomfortable middle.

And then there are the humans.

Meta has been internally testing a “human concierge” system in which contractors handle some phone calls placed through Muse. Humans! Actual flesh-and-blood manushya entering the supposedly automated loop, because apparently the final frontier of artificial intelligence is discovering that someone still has to telephone the dentist. Employees themselves reportedly raised privacy concerns: a delegated call could contain personal or sensitive information that the user imagined was being handled only by software. Meta says this is precisely why it is testing the feature internally before any broader release and that it intends proper disclosures and safeguards.

Perfectly reasonable.

Also perfectly horrifying.

Because the great consumer-AI bargain is slowly revealing its price. The more useful the agent becomes, the more context it needs. The more context it needs, the more intimate the machinery becomes. A chatbot that knows nothing about me can safely tell me the boiling point of ethanol. An agent that can save me five hundred dollars needs to know what I pay, whom I pay, when I pay them, what I bought, what I forgot to cancel and which humiliating little recurring charge has been nibbling ₹499 from my account every month since the Pleistocene.

The artham calls, and the anartham follows like a shadow.

OpenAI does not necessarily have to beat Muse by making ChatGPT twenty percent smarter. Anthropic does not need Claude to discover a new branch of mathematics between breakfast and lunch. The interesting competitive axis may be elsewhere. Make the shopping agent unusually good at comparing products across hostile storefronts, tracking prices, dealing with warranties, obtaining refunds, cancelling subscriptions, completing checkout — the vyavahara, the grubby mundane business of being alive — while persuading merchants that allowing your mechanical customer through the door sends them business rather than merely stealing their margin.

Claude could compete on another axis: become aggressively boring about trust. Ask permission. Explain the action. Preserve receipts. Minimize retained data. Make it painfully obvious what the agent can see and what it cannot. There may be considerable money in being the AI equivalent of the accountant who wears the same grey sweater for twelve years and never accidentally transfers your savings to Mongolia.

The battle may ultimately be less about which company owns the cleverest chatbot than which company I trust to carry my wallet.

Trust! In Silicon Valley!

The phrase itself is a sutra of absurdity, a koan designed to break the mind.

And then there are the Chinese models, those cheena engines of efficiency, which could make this wonderfully uncomfortable for everyone. If capable open-weight or extremely inexpensive Chinese models become good enough to run useful shopping agents — DeepSeek, Qwen, or whatever crawls next out of the Hangzhou laboratories and Shenzhen hardware bazaars — the discount shopper may eventually ask a much more dangerous question: why am I paying twenty dollars, fifty dollars or a hundred dollars a month for an American AI whose principal job is helping me avoid paying twenty dollars, fifty dollars or a hundred dollars for everything else?

Connect such an agent to Chinese manufacturing, cross-border marketplaces and brutally efficient price comparison and the proposition becomes almost indecently obvious. It would bring its own very serious questions about privacy, jurisdiction, data governance and geopolitical trust — the anartham arrives free with shipping — but consumers who comparison-shop detergent may eventually comparison-shop intelligence too.

The race to the bottom meets the race to artificial superintelligence.

Perhaps that is where consumer AI becomes genuinely strange. Today I choose between Amazon and another shop because one sells the same charger for three hundred rupees less. Tomorrow my AI may choose among retailers for me, while I choose among AIs according to which one extracts the largest discount from those retailers. The machines negotiate, the platforms block one another, the merchants fight over access, banks worry about fraud, regulators worry about data, payment networks invent new machinery for letting robots spend human money, and I sit in Calcutta waiting to discover which silicon creature found the cheapest pressure cooker.

The purusha is now the prakriti; the observer is the observed; the shopper is the shopped.

Artificial intelligence may conquer civilization eventually. First it appears to be going bargain hunting. First it is clipping coupons with the mechanical precision of a dhobi ironing shirts, searching for the artham in the anartham, the wealth inside the trouble, the saving concealed inside the spending. The Singularity will not arrive with a bang or a whisper but with a notification:

Your refund has been processed. You saved $47.32.

I will take it. I will take it and hate myself, or love myself, or both — the bipolar condition of the modern consumer, the rajas and tamas of the checkout page. The Buddha did warn, in rather more sophisticated ways than my shopping-cart theology, about craving and attachment. But he never had to buy a flight ticket during Diwali rush.

The artham may become anartham, yes, but the coupon is real.

The coupon is real, and I am here, in Calcutta, waiting for the machine to save me money I do not have, so I can spend it on things I do not need, in the endless anartham cycle of digital samsara.

P.S. Kautilya’s Arthashastra actually makes the joke more dangerous. Artha eva pradhana iti Kautilyah. Arthamulau hi dharmakamav iti — Kautilya treats material well-being as primary because dharma and kama depend upon it. Which is considerably less Instagram-spiritual than pretending money does not matter.

Kautilya wrote for kings.

Silicon Valley writes for everyone.

The dharma has been democratized, and it costs $19.99 per month, unless your agent finds a coupon.