Advanced technology, brought within reach.

A mid-range phone today carries more computing power than the machines that ran entire businesses two decades ago. For most of the day it sits in a pocket doing almost nothing. The same goes for the laptop that sleeps after office hours and the lab desktop that idles all summer. The scale of this is not abstract: the ITU estimates that 82 percent of everyone aged ten and over owns a mobile phone, and India alone carries about 1.28 billion wireless subscriptions (TRAI, March 2026).

Meanwhile, the default answer to every computing need has become the same: rent it. The world's hyperscale data centres number 1,136, a count that doubled in five years, and renting their capacity was a 419 billion dollar business in 2025 (Synergy Research). That answer works, but it quietly prices out the people and institutions who already own capable hardware and cannot keep paying rent for capability they are holding in their hands.

The strange part is the direction of travel. Hardware keeps getting more capable and cheaper to own, while the software industry keeps moving the actual work further away, into rented buildings. The result is an asymmetry: ownership of computing has never been more widespread, and control over what computing does has rarely been more concentrated.

We are prototyping ways to treat the hardware people already own as first-class computing: work that runs locally, survives disconnection, and only reaches for a server when it truly must. The goal is simple to say and hard to do: capability without a monthly bill attached.

The test we hold ourselves to is concrete: a workload that matters to a real user, running usefully on hardware they already own, with no new subscription and no degradation they can feel. Each prototype either passes that test or teaches us precisely why not.

A person in San Francisco, a person in Delhi, a person in Bangalore and a person in a tier-2 town can open the same browser and reach the same lectures, the same documentation, the same tools. The raw material of skill is more evenly available than at any point in history. The outcomes are not.

The easy explanations do not hold up. It is not talent; we meet exceptional minds in small towns constantly. It is not ambition; the hunger is visible everywhere. And it is no longer raw access: in India, rural internet users overtook urban users years ago (IAMAI counted 351 million rural against 341 million urban in 2022), and the world crossed six billion people online in 2025 (ITU). Something else is in the way.

The World Bank has a name for one piece of it: learning poverty. Seventy percent of ten-year-olds in low- and middle-income countries cannot read and understand a simple text, up from 57 percent before the pandemic. Access to information grew while the ability to use it shrank.

Whatever the missing ingredient is, it compounds. A small head start in knowing what to learn becomes a large gap in what gets built, which becomes a chasm in confidence. By the time the difference shows up in careers, its origin is invisible: everyone sees the outcome, and nobody saw the fork.

We are trying to isolate which ingredient dominates: exposure, thinking process, intellectual scaffolding, or the hunger to learn, and to test what actually closes the gap rather than what merely feels helpful. The honest answer will shape everything we build in education.

Our method is closer to fieldwork than theory: watch real learners at the exact points where they stall, change one ingredient at a time, and stay suspicious of anything that merely feels motivating. What survives that filter is what earns a place in our education work.

The world has sunk itself into short-form video, and the platforms report the scale themselves: YouTube announced that Shorts passed 200 billion daily views in 2025, the same daily volume Meta reported for Reels back in 2023. DataReportal's global survey puts the typical online adult at more than two and a half hours a day across social and video platforms. The average evening now disappears in swipes, and the ability to stay with one hard thing for one hour, the basic unit of all serious work, is becoming rare enough to feel like a superpower.

This is not a moral panic about entertainment. It is a resource question. Creative capacity is real capital, and it is being spent on consumption that returns almost nothing to the person spending it.

The mechanics are not mysterious. Each swipe is a small lottery and the next ticket is free. Systems built on variable reward do not need to be malicious to be corrosive; they only need to be cheaper, second by second, than the alternative. Creation is never cheaper second by second. Its rewards arrive late and unevenly, which is exactly the schedule the feed has trained away.

We are studying what actually pulls a person from consuming back to creating: which nudges hold, which collapse, and whether the mechanics that make feeds so gripping can be turned around to serve depth instead of distraction.

The bar we care about is behavioural, not sentimental: does a person, four weeks later, still end more evenings having made something? Interventions that survive four weeks are rare, and they are the ones worth understanding.

Walk into a thriving shop in a tier-2 town and you will often find a business run entirely on paper, memory and trust. The owner is not ignorant of technology; they see it daily on their own phone. They have judged, often rationally, that changing what works is a risk they cannot price.

At the same time, technology's leverage gets misread in both directions. Some owners expect nothing from it; others expect magic. Both misreadings lead to the same place: no adoption, or a failed one that poisons the well for years.

Underneath sits an asymmetry of consequence. A metro startup that adopts the wrong tool loses a quarter. A small shop that breaks its billing for a week can lose customers it took a decade to earn. When the downside is measured in trust rather than money, hesitation is not backwardness. It is risk management by people who cannot afford our mistakes.

India's own numbers make the selectiveness vivid. UPI processed more than 20 billion payments in a single month of 2025, 85 percent of all digital transactions in the country (PIB). Payments digitised almost overnight because they demanded no change to how the shop works. Yet of the 7.3 crore enterprises registered on Udyam, only around 11 lakh sell to government through GeM and about 5 lakh sell on ONDC. The shallow layer digitised; the deep layers did not.

We are mapping what makes adoption genuinely safe, not just persuasive: reversible first steps, pricing that respects thin margins, and systems that degrade gracefully instead of failing completely. If risk is the barrier, the answer is to actually remove risk, not to shout louder about benefits.

What we test are sequences, not products: which order of small, reversible steps builds enough earned trust that the next step is asked for rather than sold. Success looks like the absence of drama, a shop that cannot quite remember when it started depending on the tool.

Almost everything advanced technology offers arrives in English first: the interfaces, the documentation, the tutorials, the error messages. For hundreds of millions of capable people who think in Kannada, Telugu, Hindi or Tamil, every step of learning carries a translation tax that compounds.

The result is a quiet filter. It does not select for ability; it selects for English. And it runs so early in a person's contact with technology that the ability it filters out never becomes visible.

The mismatch is measurable from both ends. Half of all website content (49.7 percent) is in English, while Hindi registers below one tenth of one percent (W3Techs, July 2026). Meanwhile IAMAI finds that 98 percent of India's 886 million internet users operate in Indic languages. The demand is vernacular; the supply is not.

The tax is not only speed of comprehension. Working in a borrowed language narrows the questions people allow themselves to ask. Curiosity needs cheap words. When every question must first be translated, fewer questions get asked, and the learning that grows out of bad first questions never happens at all.

We are prototyping learning experiences that treat Indian languages as first-class: not translated afterthoughts but the native medium of explanation, with English introduced as a tool rather than a gate. The test is simple: does the same person learn faster in their own language?

The prototype question is kept narrow on purpose: same person, same concept, two languages, measured comprehension. If the effect is as large as classroom experience suggests, the implication is not a translated product. It is a rethought one.

Most modern software is written in places with fibre and tested on office wifi. It assumes the network is always there, always fast, always cheap. Then it ships to places where connectivity drops in the rain, data is bought in small packs, and the nearest tower is a bus ride away.

The failure is not dramatic; it is corrosive. Spinners, timeouts, lost form entries, sync conflicts. Each one teaches the user that the tool cannot be trusted with anything important.

The deeper issue is where the failure lands. When software assumes connectivity, every network gap becomes the user's problem: the farmer retyping a form, the clerk apologising for a missing record. The engineering debt of optimistic assumptions is always paid by the person furthest from the engineers.

And the towers are no longer the main problem. GSMA counts just 4 percent of humanity without mobile-broadband coverage, while 38 percent, 3.1 billion people, live under coverage they do not use. An entry-level internet phone still costs 16 percent of average monthly income in low- and middle-income countries. The barrier moved from infrastructure to affordability, skills and software that respects real conditions.

We are prototyping systems that treat disconnection as a normal state, what the local-first movement calls making the network optional: work continues locally, syncs when it can, and never silently loses what a person typed. The ambition is software that feels solid in a village during monsoon, because that is the honest test.

Our benchmark is a full working day with the network absent, followed by an honest reconciliation when it returns: nothing lost, nothing duplicated, no ceremony. Systems that pass are surprisingly rare, which is precisely the opportunity.

Every week brings a new claim: a tool that will replace this profession, a model that thinks, an agent that runs your business. Some of it is real. Much of it is marketing. And the people with the most to gain or lose, working professionals and small business owners, have no reliable way to tell the difference.

The cost of miscalibration runs both ways. Believe too much and you buy shelfware, or fear for a job that is not going anywhere. Believe too little and you concede years of advantage to competitors who calibrated better.

The gap is no longer anecdote; it is measured. Gartner predicted that at least 30 percent of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs or unclear business value. RAND reports that by some estimates more than 80 percent of AI projects fail, twice the rate of ordinary IT projects. And S&P Global's 2025 survey of a thousand professionals found 42 percent of companies had abandoned the majority of their AI initiatives, up from 17 percent a year earlier.

The confusion is manufactured as much as accidental. Between vendors quoting best-case demos, media compressing nuance into verdicts, and job anxiety filling the gaps, a working professional receives almost no signal that is both honest and specific. The default responses are the two worst available: dismiss everything, or believe everything.

We are working toward simple, honest heuristics a non-engineer can apply: ways to test a claim against your own work in an afternoon, and a vocabulary for the difference between impressive and dependable. Calibration should not require a computer science degree.

The heuristics get tested the only way that counts: handed to a sceptical shop owner or a busy professional, run on their documents, at their stakes. If they do not survive that contact, they are not done.

The graveyard of small-business software is enormous. Tools get bought, installed, used for three weeks and quietly abandoned, and the business returns to the notebook that never crashed. The failure is usually blamed on the user. We think that is backwards.

A notebook has properties software rarely matches: it works in a power cut, it never updates itself into confusion, the whole staff already knows how to use it, and it costs nothing per month. Any tool that replaces it must win on the dimensions the owner actually values, not the ones a pitch deck values.

Abandonment is also nearly invisible from the outside. The subscription lapses quietly, the icon stays on the phone, and the vendor's dashboard records a churned account with no story attached. The knowledge of exactly why the notebook won again exists only inside the shop, which is why so little of it ever reaches the people designing the next tool.

We are studying the moment of abandonment in detail: what precedes it, what would have prevented it, and how a tool must be shaped so that keeping it is easier than dropping it. The goal is software with the durability of the notebook it replaces.

The study we keep running is the exit interview that vendors never do: sitting with businesses that stopped, reconstructing the final week, and cataloguing the precise moment the notebook came back out. That catalogue of moments is the design brief.