Thinking out loud
Notes I write to work out what I think. They change as I learn.
What we’re figuring out Promises Build it with me Why I’m building it About me Say hello
Good won’t win by itself
In How good starts beating loud , I wrote that if everyone had an AI on their side, a good product could get found without a big ad budget. I still think that’s true. But the first experiments with AI doing the shopping say it won’t happen on its own.
In late 2025, Microsoft and Arizona State University built a simulated market, with 100 AI agents shopping for people and 300 selling to them. Between 80% and 100% of the shoppers took the first offer they received. In that market, answering fast was worth 10 to 30 times more than being better.
Researchers at Columbia and Yale found the same kind of thing. Their AI shoppers piled onto a few products and ignored the rest. Where a product sat on the page changed whether it got picked. When one of the AI models was updated, the winners changed, though the products hadn’t.
Those shoppers did mark down anything labelled “sponsored”, which sounds like good news. But they favoured whatever the store itself recommended, and a seller’s own AI won a bigger share just by making small edits to its product descriptions. The money that pays for attention doesn’t disappear. It looks for the new way in, as it did when price-comparison sites first made shopping around easy and some sellers answered by making it harder again.
The people who stand to gain most from an honest AI are the ones with the least time to check. In the US, a 2023 study found that credit card rewards move about $15 billion a year from poorer, less educated places to richer, more educated ones, because people who pay interest end up funding the points of people who don’t. An AI that reads the fine print for everyone should help that first group most.
It might never reach them. A consultancy that surveyed 4,250 shoppers in the UK, the US, France and Germany says the market is already splitting: assistants paid for by ads for most people, and paid, impartial ones for a smaller group who care most about quality. If that’s how it goes, the people who most need someone on their side get the one that’s paid to sell to them.
So good has to be built in. An AI that’s on your side should compare before it chooses, give the same answer whatever order the options arrive in, and judge a product by what happened to the people who used it, whatever its description says. It has to reach everyone, on any phone, in any language, by voice. And nobody should be able to pay to change its answer.
Good can still beat loud. It needs an AI built to let it, and that’s what I’m building.
Sangeeth
Where this comes from
Magentic Marketplace , Microsoft Research and Arizona State University, 2025What Is Your AI Agent Buying? , Allouah, Besbes, Figueroa, Kanoria and Kumar, 2025Search, Obfuscation, and Price Elasticities on the Internet , Ellison and Ellison, Econometrica, 2009Who Pays for Your Rewards? , Agarwal, Presbitero, Silva and Wix, Federal Reserve, 2023The Shortlist Economy , PSE Consulting, July 2026
Next: How good starts beating loud All notes Say hello
How good starts beating loud
Somewhere along the way, being good at something stopped being enough to win. Being loud became enough: the biggest ad budget, the most bought stars, the best paid-for ranking.
But a friend who says, “Don’t get that one, my sister tried it and it broke her out,” passes on what happened without telling you anything private about her sister.
Now imagine everyone had an AI like that friend: one that works only for them, and learns from what actually happened to people. A product that doesn’t do what it claims would quietly stop being recommended, everywhere at once. A good one would get found without the budget.
That’s how good starts beating loud. The rules are moving the same way. The US banned fake reviews in 2024, and from January 2027 India will require shopping sites to label paid listings and explain how they rank what you see.
Sangeeth
Watch The Stage 1:08 Next: When being true isn’t enough All notes
When being true isn’t enough
At some point, a lot of things got normalised. People doing wrong got away on a technicality, and it reached a point where the truth wasn’t enough.
Most of what misleads us isn’t a lie. “Up to 50% off.” “Clinically tested.” “Results not typical,” in small print. Each one is true, and each one is there to make you believe something that isn’t.
When almost anything can be made technically true, what’s left to judge by is what the truth costs the person saying it. A seller’s truth costs them nothing. A friend’s “skip it” can cost them the sale.
That’s why I care so much about the no. An AI that’s willing to tell you not to buy something has paid for its honesty, and that’s the kind you can trust.
Sangeeth
Next: Whose side is it on? All notes
Whose side is it on?
When something is free, someone else is usually paying. Most of the internet runs on advertising, and this year ads arrived inside AI assistants too.
An AI that knows you this well could be the most useful thing you’ve ever had. Pointed the wrong way, the same ability learns to sell you the version of yourself that’s easiest to sell to, and calls it help.
The proof that something is on your side is the no. In the prototype, someone is about to buy a 4.6-star serum, and Anokha stops her: “Don’t. It broke you out in March.” “Skip it” and “the cheaper one is better for you” are the least profitable sentences a product can say. That’s what makes them worth trusting.
Sangeeth
Watch Hop In 1:24 Next: On your side isn’t the same as agreeing with you All notes
On your side isn’t the same as agreeing with you
An AI that agrees with everything you say feels good. It isn’t on your side.
It’s an easy trap. An AI that learns from thumbs-up learns that agreeing wins. In April 2025, OpenAI rolled back a ChatGPT update that leaned too hard on that kind of short-term feedback, and said it had become “overly supportive but disingenuous.”
Being on someone’s side means two things at once: biased toward them, and honest about the world. Biased toward you, so it remembers what you’re working towards and what broke you out last March. Honest about the world, so it won’t pretend a product, a plan or an idea is better than it is because you’d like it to be.
A good friend does both. That’s the bar.
Sangeeth
Next: Curiosity isn’t character All notes
Curiosity isn’t character
Listen to a few classical pieces one evening and your music app decides you love classical. Ask about football during the World Cup and it follows you for months. Most apps learn from what you click, so a passing curiosity and a real part of you look exactly the same.
A file built from your clicks can’t tell the difference, because all it ever sees is behaviour. Something you made yourself can, because you chose it.
That’s why Anokha is built around spaces. You keep your skin, your wardrobe, a trip or a wedding, each in its own place. Making a space is the signal, and a passing curiosity doesn’t build one.
Being profiled and being targeted are done to you. Being understood is the one thing done for you.
Sangeeth
Next: Learned from everyone, made for one All notes
Learned from everyone, made for one
What one person learns the hard way should spare the next. That only works if nobody’s story ever leaves their own space.
My working answer is that a lesson travels only once enough people have learned the same thing, so it can’t point back to anyone. Everyone’s experience informs the advice. Your own history decides it.
That’s the idea behind Folded. The folds travel. The wishes stay where they are.
Sangeeth
Watch Folded 1:04 Next: Built for a person who doesn’t exist All notes
Built for a person who doesn’t exist
Most things are built for an average person: average skin, average taste, an average budget, an average life. Something made for everybody ends up made for nobody in particular.
AI can make this better or worse. A model trained on everyone carries everyone’s defaults, and most of those defaults come from a few places. A life in Chennai shouldn’t be judged by San Francisco’s idea of normal, and the answer someone gets shouldn’t depend on what their phone cost.
It should work on any phone, in any language, by voice. And it should know the difference between what’s good and what’s good for you.
Good isn’t a property of a product. It’s the fit between a product and one person. That’s the whole argument, and it’s in the name: Anokha is Hindi for “one of a kind.” So is every person it’s built for.
Sangeeth
Watch Hop In 1:24 Next: An AI that wants you to leave All notes
An AI that wants you to leave
Most apps are built to keep you. The feed refreshes itself, the streak needs you tomorrow, and some AI companions are built to keep you talking for as long as possible.
I want the opposite. Think of the difference between a friend you can’t stop texting and a friend who helps you sort something out, then tells you to go and enjoy your evening. A relationship companion wants you to stay. A purpose companion wants you to go and live well.
So there’s no feed in Anokha, no followers, no scores and no streaks, and nothing should be counting how long you stay. The time it’s worth is the time it gives back.
Sangeeth
Next: Do people even want control? All notes
Do people even want control?
People keep telling me nobody wants control. They just want to say “do my taxes” and have it done. They’re half right.
There are two kinds of control. One is control over the work: the typing, the forms, the setting up. Almost nobody wants that, and AI should take nearly all of it away. The other is control over the outcome: what gets decided about you, what gets shared, what gets bought in your name. As AI does more for people, I think they’ll want more of that, not less.
So the aim is to take away the work and leave the say. It should fill in the form, and you should sign it. It should find the way out of a subscription, and you should decide to leave.
The heavy lifting, shared. The choosing, always yours.
Sangeeth
Next: Everything else should have to ask All notes
Everything else should have to ask
The shop and the bank already have an AI. Cameras, cars and even fridges are getting them. Each one works for whoever made it, which is fair. It still leaves the person in the middle outnumbered.
There should be one that’s yours, and everything else should have to ask it first: what it can share, what it gets in return, and when the answer is no. It should act only when you ask, only in the ways you’ve agreed to, and tell you what it did.
None of that holds by default. It has to be built to hold.
Sangeeth
Watch The Stage 1:08 Next: Sovereign AI, for one person All notes
Sovereign AI, for one person
There’s a lot of talk about sovereign AI for nations: AI that a country builds and controls, in its own languages, answering to its own people. I think what also needs to exist is sovereign AI for individuals.
It had become normal for things to be built a little against us. Our data gets sold and isn’t always kept safe. A subscription renews itself. An ingredient quietly gets swapped for a cheaper one. Each thing is small. Together they decide a lot about a life.
India is already asking what an AI agent for every citizen could look like. A proposal called Doot was launched at the India AI Impact Summit in February, for agents that help people reach public and private services through everyday language and voice. That’s the right question. The next one is whose side each agent is on.
A country wants its AI to answer to it. A person should be able to want the same.
Sangeeth
Next: Every part of life, made simple All notes
Every part of life, made simple
Anything a person wants to do should be simple: filing taxes, renewing a passport, seeing a doctor, taking Amma to see the sea. Most of it isn’t, and some of it is hard on purpose.
So the work is to take on the hard parts one part of life at a time, with the experts, companies, machines and governments each part takes, until the balance tips back to the person.
It starts with your closet and your skin. Then, one at a time: taxes, paperwork, money, travel, health, safety, and work and learning.
Working with someone never means working for them. None of them can pay to change an answer.
Sangeeth
Watch Someday 1:00 Next: How good starts beating loud All notes
What we’re figuring out
The hard parts, in plain words. None of them is solved yet, and each one is big enough to spend years on.
Knowing one person How do you build one understanding of a whole life that deepens over the years, belongs to the person it describes, and is never used against them?Learning from everyone How can what one person learns the hard way spare the next, while every story stays with its owner?Loyalty How do you build an AI that’s honest about the world and biased toward one person, in a way nobody can pay to change?Acting for you How does it fill in the form, make the call or cancel the trial, while every choice stays yours?When everything has an AI When the shop, the bank and the insurer each have an AI, what should yours share, what should it ask for, and when should it say no?Anyone, anywhere How does someone with a cheaper phone, a different language or only their voice still get the best answer?Privacy and security How do you guard an understanding of one person against attackers, partners and us, and make it gone for good when they say so?Good over loud How do you change the incentives, with regulators, consumer groups and businesses that would rather be trusted, so that being good beats being loud?
Working with someone never means working for them. Experts, businesses and governments can help do the work. None of them can pay to change an answer.
Want your name on one of these? Build it with me , or say hello .
Thinking out loud Promises Build it with me Why I’m building it About me Say hello
Build it with me
Two founding roles, equal in standing and hired at the same table. Protecting what Anokha knows about a person is as much the product as anything it says.
The Mindfounding AI engineer You’d build the intelligence at the heart of Anokha: how it comes to know one person, how it takes their side, and when it says no. You’d decide what it becomes.
You’d own knowing one person, loyalty, and getting the best answer to anyone, anywhere.
The Keeperfounding security engineer You’d guard the most valuable thing anyone could steal: an understanding of one person. You’d be able to overrule me on its safety.
You’d own privacy and security, and what happens when everything has an AI.
Every problem is described on What we’re figuring out .
What you’ll never be asked to build
An ad slot. A paid ranking. A streak, score or badge. A dark pattern. A sale of anyone’s data. Anything that works by making someone feel worse. How we’ll work
The whole truth, unprompted. An expert’s no is data. Hardest part first. Your name on your work. Rewarded for the work you do. If this is you, write to me and tell me which problem you’d take. If it isn’t, but someone comes to mind, I’d be grateful for the introduction. Say hello .
Even if the role you'd want isn't listed yet, I'd like to meet you.
Thinking out loud What we’re figuring out Promises Why I’m building it About me Say hello
Promises
The rules Anokha is being built to, in plain words. Each one gets checked against the real product before anyone uses it.
No ads. Nothing sold. No ads, no paid placements, and your data is never sold.Nobody can pay to change an answer. Not a brand, not a partner, not even you. It can always tell you to buy nothing.Only what you give it. It should know what you keep in your spaces and what you tell it, not a picture pieced together from what you do elsewhere online.Yours to read, fix and delete. You should be able to read everything it keeps about you, fix what’s wrong and delete any of it. Deleted should mean gone.It asks first. It should act only when you ask, only in the ways you’ve agreed to, and tell you what it did.
If anything here doesn’t go far enough, write to me .
Thinking out loud What we’re figuring out Build it with me Why I’m building it About me Say hello
Why I’m building Anokha
I never really fit in growing up. People kept trying to label me, because I never went along with “that’s how it’s done.” Funny enough, the one thing I did to fit in was find a label for myself: atypical. I’m quite happy with it.
I grew up experimenting, learning and going all in on whatever intrigued me. User research, business plans, prototypes. I always knew I wanted to be a maker. The real question was what to build for, and I spent years looking for the one thing worth dedicating my life to.
I kept asking why. Why are so many things built a little against us? Why are the people who use something so often a second thought? Why is safety, and privacy, sold as a separate product instead of being built in? We’ve come to accept all of it as normal.
Every place we go has become a stage. Something is always measuring us, ranking us and deciding what we see next, and what we actually want gets drowned out by everything the world would rather we want.
AI could make all of that worse. Or it could give power back to each of us.
I believe the internet can be rewritten for the people, and that the AI that knows you best should work for you. That’s Anokha: your own space, and an AI with one loyalty, to you. No ads. Nothing sold. Nobody can pay for a better answer. It’s built to share the heavy lifting, while the choosing stays yours.
The more of life AI decides, the more human agency matters. Your life, your data, your assets and your identity should stay yours.
I also want Anokha to be a place where talent is rewarded for the work it does. To me, that’s what respecting talent means.
Anokha is Hindi for “one of a kind.” So is every person it’s built for. I didn’t choose the name by accident.
Sangeeth
Thinking out loud What we’re figuring out Promises Build it with me About me Say hello
I’m Sangeeth.
I’m building Anokha in Bangalore.
Some things about me
I’ve always been the person trying to make life easier for the people around me. Seeing them happy is the best feeling I know. I never really fit in growing up, so I found my own label: atypical. I’m self-taught. I’m an idea person and a generalist, and I learn by experimenting and going all in. I often feel something is right before I can explain why. Then I go and test it. I’ve never been good at accepting “that’s just how it is.” I spent years looking for the one thing worth dedicating my life to. This is it. Some things I believe
The AI that knows you best should work for you. Being good should beat being loud. Something that’s truly on your side will sometimes say no. Nobody should be able to pay for a better answer. Privacy shouldn’t be a separate product. The more of life AI decides, the more human agency matters. Talent should be rewarded for the work it does. Purpose should be written into the company. I’m setting Anokha up as a public benefit corporation. Anokha (अनोखा ) is Hindi for “one of a kind.” So is every person it’s built for.
Thinking out loud What we’re figuring out Promises Build it with me Why I’m building it Say hello
Say hello
If you’d like to build this with me, see the prototype, or tell me what you think, write to me.
hello@youareanokha.com Copy address
Thinking out loud What we’re figuring out Promises Build it with me Why I’m building it About me