Life GPS Try the beta

The average recommendation no longer works

Almost everything that can reach us at scale reaches us in a generalised form.

None of that was a mistake. It was the price of scale. Knowledge could not be moved any other way. One book had to speak to a hundred thousand readers, one curriculum to a generation, one guideline to a population. Averaging is what made knowledge reach anyone at all.

The problem is what happens when that knowledge meets one person.

A person was never average

What actually decides whether something works for you:

All of it at once. And all of it changing.

Medicine gives us one of the clearest precedents. The NIH describes the old model plainly: recommendations were built around the response of an average patient, and precision medicine exists because one-size-fits-all works well for some people and much less well for others.

But there is a trap on the other side of that, and it is worth saying out loud before anyone says it for us. General knowledge does not become worthless because people differ. Observed differences between people are real. Evidence that those differences are stable, and predictably explainable, is much harder to establish — which means personalisation can overfit noise just as easily as averaging can flatten signal.

So the claim is not that averages are wrong and personal answers are right. The claim is narrower and harder to knock down:

General evidence is the starting point. Context decides what is routed to whom. Real outcomes update the evidence about whether that routing was any good.

Which means the useful question was never "what is the correct advice." It is: out of everything that already exists in the world, what fits this person, at this point in their life, right now?

We have not found a system that has demonstrated it can answer this across a whole life, across domains, at scale. And here is what that costs.

The first one is not a character flaw. A meta-analysis of 47 experiments found that interventions moved people's intentions by about d = .66 and their actual behaviour by only about d = .36. Wanting to is not the same mechanism as doing. And across 94 tests, forming a specific if–then implementation intention had a medium-to-large effect on actually reaching the goal (d = .65). That much is established. Whether a system built around one situated next step closes that gap at the scale of a life is a hypothesis, and it has to be proven with outcomes, not argued.

That is not a knowledge problem. That is a navigation problem.

Everything got smarter except the life

Robots walk. Cars drive themselves — Waymo reports hundreds of thousands of fully autonomous paid rides a week. Amazon has deployed its millionth warehouse robot. In Anthropic's own Claude Code data, the longest autonomous turns had risen past forty-five minutes without a human in the loop. A parcel crosses a country overnight. Every book, every paper, every expert opinion is one query away and mostly free.

And at the same time: the WHO estimates more than a billion people live with a mental health condition, and most get no adequate care. Around one in six people worldwide experiences loneliness. In 2023, 19% of young adults said they had no one to count on for support — 39% more than in 2006. US adolescents spend seven to nine hours a day on entertainment screens, by their own government's measure — and that government's own advisory is careful to say the research is largely observational and does not establish cause.

We are not claiming one caused the other. We are pointing at the gap: the capability of the world went up, and the capability of a person to run their own life did not go with it.

Look at what a person still does by hand. Understand themselves. Decide what they want. Find the right knowledge. Tell it apart from the wrong knowledge. Find the people. Find the opportunity. Fit it to their money and their time. Make themselves act. Work out what actually happened. Start again.

The car understands the road. The warehouse understands logistics. The bank understands risk. The ad understands conversion. The robot is starting to understand physical work. The human being is still the manual integration layer of their own life.

Personalisation already won. It is just not working for you

Here is the part that surprises people who expect an argument against algorithms.

The feed sees signals your doctor and your school never see: every pause, every swipe, every replay, everything you abandoned and everything you came back to. Technically, the personal layer exists. It works. It is extraordinarily good.

It just optimises for something that is not your life.

And it is measurably good. A randomised trial of an educational app with 7,750 children found that personalised recommendations raised use of the personalised section by around 60%. A study of Netflix estimated that replacing modern personalisation with a simple popularity ranking would cut engagement by roughly 12%. These are proofs that personalisation moves behaviour. They are not proofs that it moved a life.

For a feed, success is you staying another hour. That is not a philosophical objection — it is now a legal one. Courts and regulators have moved from arguing about content to arguing about design: infinite scroll, autoplay, algorithmic ranking. In 2026 Meta reached an agreement with a bipartisan coalition of US attorneys general on terms including daily limits for teens, a block between midnight and 6am, and muted notifications during school hours, with thousands of related cases against major platforms still running. The company denied wrongdoing — a settlement is evidence of a very large fight over design, not a court finding that design caused harm.

The scale of what was built here is worth stating plainly. US internet advertising generated $294.6bn in 2025. Globally, Meta reported $196.2bn in advertising revenue that year and Google $294.7bn. That is the economic size of businesses whose survival depends on deciding which stimulus should reach a person next — and an industry that size does not run on guesswork.

So the success of the attention industry is not our counterexample. It is our proof. A system that knows a person well enough can move them. The only open question is whose goal it moves them toward.

There is a quieter consequence of the same achievement. For the first time an infrastructure exists that can fill almost any free second with something chosen for you personally. We are not going to claim people spend nine hours a day running away from themselves — much of that time is friendship, work, learning and making things. The narrower point is harder to argue with: a person now rarely has to sit with the question of what they actually want, because something is always ready to answer it for them.

We built an enormous, precisely measured infrastructure for producing the next external impulse. We have nothing comparable for a person's own direction.

And now it is harder to know what is even real

Until recently the hard part was choosing from a world that was too large. Now the world is being generated faster than anyone can check it.

Text, images, video, advice, "history", experts, reviews — the cost of producing something plausible is falling to nothing. A small 2026 MIT experiment found something worse than people believing fakes: after watching realistic AI-generated video, participants trusted genuine video less, even when they had been told which was which. It is one study on a hundred people and deserves replication rather than a headline. But it points at the failure mode that matters: not being fooled by a fake — losing confidence in the real.

A 2025 essay published by UNESCO gives this a name: a crisis of knowing. The task is no longer to catch a deepfake, it is to keep the ability to say what we know and why we believe it. There is a separate, purely technical version of the same spiral published in Nature — train models recursively on generated data and the model itself degrades. Different problem, same direction of travel.

For a child growing up after generative AI, an image of the past is no longer evidence of the past. They will see Caesar, wars, presidents and their own parents in footage that never happened.

So a personal layer now needs a second job, and it is worth being precise about how much of it already exists. Content provenance is being built: C2PA can carry and cryptographically bind claims about where a file came from and how it was edited — and its own specification is honest that this does not establish that the content is true, and that provenance can be incomplete or absent. Domain outcome evidence exists too: the FDA has long used real-world data and evidence to evaluate post-market safety and support regulatory decisions, including labeling changes. What we have not found demonstrated at scale is the third one: an outcome loop for a whole life — what happened to people like you who actually did this, across everything that mattered, not just inside one regulated domain.

The label is free. The outcome is not.

There is a second consequence of cheap generation, and it lands closer to home.

AI made information cheap. It is now making claims cheap. A site, a concept, a demo, a whitepaper, twenty pages of philosophy, a name, a launch video, a thousand posts about how revolutionary it all is — a week's work. The words can now be produced faster than the system behind them can be built.

And this is not new to AI. It is the oldest race there is: naming the future before anyone else is cheaper than waiting for the evidence, and it works.

AI agent. Personal AI. Whole Person. Life OS. Human OS. One system for your life. Any of those can be on a landing page before the thing exists.

So we are not going to fight over words. Call yourself the wheel. Show that it rolls.

Which means asking a different set of questions — of anyone, including us:

If those answers do not exist, it may still be an excellent product. The word whole just has not been earned yet.

And the same question belongs to the industry that is already winning. A system succeeded in producing one more hour of use. What human outcome did that hour produce? Did the person get healthier, settle something, get closer to work they wanted, earn what they needed, meet someone who mattered, learn something they went on to use? Or is the only precisely measured quantity the fact that they stayed?

The attention industry is extremely good at proving its own results. Watch time, retention, click-through, revenue — all measured to the decimal. Where is the matching evidence for the human outcome?

We ask for proof from a medicine, a bridge and an aircraft. It is strange that for a system touching a person's attention several hours a day, the system's own numbers are considered sufficient.

And this test applies to us first. Until a person can show that they wanted something, got a next step, did it, and ended up closer than when they started — we have no more claim to the word than anyone else. We do not want to own the label. We accept the test.

The pieces already exist

Here is where we have to be accurate, because the easy version of this argument is false.

It is not true that nobody thinks about the whole person. Family medicine has defined whole-person care for decades, and there is a body of research on what that means in practice. Case management, medical homes and social work all exist to hold someone across more than one problem. Governments have noticed too: the UK has had a cross-government loneliness strategy since 2018, Japan has a law and a national plan, Sweden launched a national strategy in 2025.

And in the market, the parts are genuinely strong.

And the hub itself is no longer an empty category. Kanvas calls itself the operating system for your whole life, and writes on its front page that you are the integration layer for your own life. LifeOS describes a single loop — current state, ideal state, next move, verification — across domains. alaivOS says fourteen life modules talk to each other.

Good. Those are claims about architecture. The test above is about outcomes, and it applies to them exactly as it applies to us.

None of this is failure. It is the reason the next layer is possible at all — and the fact that several independent teams arrived here at the same time says the moment is real, not that the space is crowded.

But look at what each one is centred on. BetterUp takes in sleep, stress and personal life so a person functions and develops better — in a work context. One Medical takes in goals and lifestyle to make a better medical decision. WHOOP takes in your routine to make a better health and performance decision. Each brings more of the person into its own task. That is not the same thing as coordinating all of a person's tasks as one life.

Think of it as a wheel. The person is the hub — not an account, a changing life. The spokes are these systems: health, money, time, work, learning, relationships, people, expertise. The rim is one real life, where all of it has to arrive as a single doable next action. And the wheel only moves when the spokes move together.

A spoke can be world-class and still be unable to move the wheel alone.

Whole person is not a label. It is a test.

Does your health recommendation change when the person cannot afford it?

Does your financial advice change when it costs them the only hours they had with their child?

Does your career system change its answer when the right move breaks the family it depends on?

Does your productivity tool know when a person needs recovery instead of another optimised hour?

Does your learning platform know that the missing piece is not another course, but a person, a permission, money, or a laboratory?

Take the most primitive version. Health says: eight hours of sleep and training. Money says: two shifts, because rent. Family says: there is a small child. Time says: there are no two free hours in this day. Every one of those systems is right. Together they describe a life nobody can live.

Or: career says an MBA is the best next move. Money says there is no eighty thousand dollars. Time says there are no two years. Family says the relocation breaks everything holding it together. If the career system still answers "MBA", it solved its own problem perfectly and the person's not at all.

Who resolves the conflict when every specialist is right?

Most of the time, the person does — alone, badly, at midnight. Sometimes a good doctor, a family member, a case manager or a friend helps them do it. What has never been demonstrated at scale is a continuous system that carries those trade-offs across domains and across years.

And because there is no such system, a failure happens that nobody is even measuring: a system improves the metric it can see while the person as a whole gets worse. Each one heals what it measures. None of them has an instrument that would show what it broke.

What is actually missing

Not more knowledge. Not another AI that answers well. Not another spoke.

A hub. A layer that sits around the person and lets the existing parts work as one wheel: that knows which spoke is needed now, how it affects the others, what is actually possible given everything else, who should be handed control of their part — and what happened afterwards.

And this is what changed, and why the answer is "now" rather than ten years ago. The breakthrough is not that a machine can produce one more answer. It is that far more of one changing human context can be held in play at once, while searching a world no person could continuously integrate by hand. The last era learned to scale knowledge. This one can scale context.

The shape of it is a loop, not a recommendation. A living model of the person, which changes. A search across what the world actually offers — knowledge, tools, people, opportunities. Matching in which the constraints do real work: if there is no time, no money or no permission, those options are gone, not politely listed. One next step, not another annual plan. Then the person acts, reality answers, and what comes back changes the model.

One surfaced step does not mean one active responsibility. A life runs several things at once — a job, a child, a treatment, something being built slowly. The layer has to hold all of them and still put exactly one thing in front of the person, instead of handing them a hundred notifications and calling it help.

Some things follow from that shape rather than being added to it.

If the best answer is sometimes not a text but a human being — a doctor, a welder, a researcher, a mother, a founder, a pilot, someone from another culture who has actually lived this — then the layer has to reach people, not just pages. In a world filling with synthetic answers, a person who really did the thing becomes an anchor of reality rather than a nice extra.

If reality can be assembled around one person's real attempt, it can be assembled around a child's — and that is an education system, not a feature. And if you gathered the knowledge, the cultures and the practitioners of the world into one physical place, you would have the largest version of the same idea.

And if a system knows someone that deeply, it can move them. So the limit cannot be a promise in a blog post; it has to be structural: the more a system can do for a person, the less it is allowed to control them. Which also means it must never decide it has found the person's one true desire. People want contradictory things, and their idea of a good life changes. The honest job is to hold the contradiction, show it, and leave the decision where it belongs.

Those are not separate ideas we happen to be working on. They are the same loop at different distances, and each one has a page of its own:

Transparency got close to the right question and stopped

This is not a world where nobody tried.

Europe's Digital Services Act now requires platforms to disclose the main parameters of their recommender systems and why those parameters matter. And in the Parliament's version of that law there was a line requiring platforms to disclose what objectives the system had been optimised for — almost exactly the question in this article. It did not survive into the final text.

So regulation is starting to tell us why something was recommended. It still does not ask whether the system is working toward the life the person chose. The AI Act does something similar for provenance: label the synthetic, disclose the deepfake. Necessary. Also not the same question.

Where this goes

The first era of the internet gave a person access to everything.

The second learned to choose what to show them so they would stay.

The next one has to learn to choose what helps them leave — into their own life.

We taught systems how to hold a person. The work now is to teach one how to let them go and live.

So build it

We are not claiming to be the wheel. We are one attempt at the hub, and a small one — a product that has started, with most of this still ahead of it.

What we are claiming is that the hole is real.

If you think that diagnosis is wrong, bring the evidence — that is a fair way to win an argument, and we would rather be corrected early than late. If something already solves this, show it to us; we will study it before we defend anything of our own. And if you think there is a better architecture, test it.

But an argument about whether a person can be helped is settled by a person being helped. Not by a longer essay. Show a system that holds a whole person over time, works toward the goal that person chose rather than its own retention, can tell the real from the generated, and changes course based on what actually happened in someone's life. Show where those people started and where they got to.

And to be exact about what "got there" means, because this is the question on which everything else rests: we do not need one universal definition of a good life. We need to know whether a person moved toward the direction they chose themselves, what it cost elsewhere in their life, and whether they still choose it after seeing the result. Not us. Not an AI. Not a state. Not an index. The person — but deciding with the consequences visible.

We will study it first, and we will say we were wrong.

And if you are building a spoke — health, money, time, learning, work, people — we are not coming for it. We need it. There is no wheel without spokes. Everything here is an argument for connecting what already exists around one person, not for replacing any of it. The invitation is open widest to the people who disagree with us most: come and build the standard by which these parts talk to each other around a human being, and let outcomes decide whose version works.

One more thing, so nobody has to guess at our motive. Subscription instead of advertising removes one conflict of interest. It does not remove all of them — subscription businesses want renewal and dependency too. The only real proof is which numbers we refuse to optimise and which ones we publish. We do not optimise for time in the product.

And in the end it does not matter who assembles it. If we all get to where we actually wanted to go, the winner is already known.

And the part nobody says out loud

There is an obvious objection to all of this.

Countries compete. Companies compete. Platforms compete. None of them can simply stop optimising the things their survival currently depends on. So asking everyone to become more humane is not a system. It is a wish.

The human outcome has to become observable too — without reducing a person to a score.

That distinction is the whole game. The moment anyone builds a single number for how good your life is, they have built a social score with better manners. So: the person names the directions. The system shows consequences and conflicts. The decision stays where it belongs.

A person who knows where they are going does not stop working, buying, learning or building. They make different choices about where their time and their energy go.

The point is not to ask progress to slow down. It is to give progress another variable to optimise for: did the person actually get closer to the life they chose?

Until human outcomes become observable — and usable inside the feedback loop — every system will keep optimising what it can already see. That is the layer we think is missing.

The side I am on

I love this world.

I love moving through it, arriving somewhere I do not know, meeting people I would never have met at home, learning how they think, what they believe, what they build, what they are afraid of. I do not want less technology, fewer companies, fewer ideas or a slower civilisation. I want more of it.

But I have carried one question for most of my life: what is a person giving all of this their energy for?

A person gives away hours, attention, health, concentration, relationships, decades. They work, consume, learn, scroll, produce, adapt. And every system around them is very good at measuring what it receives. Revenue. GDP. Output. Retention. Grades. Clicks. Conversion. Hours watched.

Where is the instrument that measures whether the person got closer to the life they wanted?

I am not going to spend the next ten years only describing why people are lost. There is already a shelf of brilliant books doing that, and I love knowledge too much to pretend they are worthless. A book can change a person. A diagnosis can change a person. Neither is a system that stays with that person, sees what happened next, and changes course on Monday morning.

I want to know what happens after Monday morning. Did they do anything. Did it work. Did they meet the right person. Did they lose another year. Did the advice improve one part of their life and quietly damage another. Did anything learn from what actually happened.

That is political in the oldest sense of the word. Inside systems that get more powerful every year, somebody has to represent the interests of the human being. Not to stop progress — to make sure the person is not lost inside it.

And no founder, no company and no model gets to define what a human being is. That would contradict the entire idea. It will take scientists, builders, doctors, teachers, parents, children, people of other cultures and beliefs, critics, and the people who think we are wrong.

I do not know whether we will get every part of it right.

I know who I want to be accountable to.

The person.

Sources

Averaging, and what replaces it

Knowing and doing

What the world can now do

The gap

Personalisation, and what it is optimised for

Reality, provenance and outcomes

Who already holds part of the person

Regulation that came close

Published 2026-09-18 · Updated 2026-09-18 · Every claim on this page carries a status