Machines of Human Grace

Life has scaled twice: from chemistry into biology, and from biology into civilisation. Both times it used the same trick.

For hundreds of millions of years the early ocean was full of molecules that could become nothing more. They formed, reacted, fell apart. Then the trick happened: a shared code. Four bases, one grammar. The moment chemistry had a common language it stopped being just chemistry. Everything alive since, every cell, every forest, every person reading this page, is written in that one vocabulary.

The second time was us. For most of human history we existed around campfires. What scaled humanity was not so much the brain. It was the evolution of language. Shared meaning let two strangers trust each other, and on that trust came trade, then law, then cities, then science. Civilisation is what intelligence became when meaning was encoded to be shared.

The third scaling will take place with software and hardware. Every application on earth keeps a private vocabulary sealed inside its own code. Your sales system says "customer" and your accounting system says "customer", and neither knows what the other means. An industry worth some twelve billion dollars a year exists to carry data across gaps of meaning, pair by pair, by hand. Software is the layer of the world that cannot compound, because its parts mean nothing to each other. And that layer now carries nearly everything: the machines, the grids, the sensors, the models. The whole architecture of intelligence that powers the world runs on a layer that has no intrinsic, encoded, shared meaning.

Now, software has begun to write itself. Anthropic's When AI Builds Itself describes Claude writing more than 80 per cent of the code merged at the company, and forecasts a world where every person sits "atop a pyramid of agents" and working software pours out of machines in minutes. Each of those applications is born speaking its own private language. The fastest builders in history are just beginning to flood the world with systems that cannot natively understand each other, at a rate no integration industry could chase. Two futures are available. Babel at machine speed. Or the third code. The third evolutionary trick.

At Sense Future, we have built the third code. This article is about this code and what it makes possible.

The code

The concept system is a dictionary held as data. Concepts like Customer, House, Medication, Employment, are nodes in one shared graph. A concept has attributes and relationships. A concept is readable by any system because it is fixed to an identifier that never changes whatever happens to the word. Rename it, translate it, abbreviate it: the identifier holds, so "house" and "home" stay one thing. Any number of systems built by strangers, human or machine, point at the same concept and mean the same thing. No meeting, no committee, no negotiation. And where two organisations genuinely mean different things by one word, both meanings stand, each under its own identifier, side by side.

This framework resolves what defeated the Semantic Web, the one serious attempt before this one. It spent twenty-five years asking organisations to agree on definitions, and agreement never came, because a definition is a political negotiation and the world would have needed ten thousand of them and it still would not have been enough. The concept system we have built at Sense Future avoids that: no negotiation needed. Nothing has to be agreed for everything to connect.

The economics have a precedent. When the shipping container was standardised, loading costs fell by more than ninety per cent and world trade reorganised itself around the box, all from a single source of gain: nothing had to be negotiated at the joins any more. The dictionary does to meaning what the box did to cargo, with one difference that matters to anyone deciding where to put capital: a box does not grow. The dictionary does. Every application built on the graph inherits the entire vocabulary of every application before it and adds its own. The vocabulary grows. Shared meaning compounds. It cannot be bought. Our system is open by architecture, and the protections are enforced by the system itself. It is in production now.

Anything can be built on it.

ESRE, the consultancy arm of Sense Future, has built a retail bank, a property acquisition platform, a music school, a social network where humans and AIs collaborate, the business system of a garden furniture showroom, built in days and trading: the same five tables under all of them, and not one schema written. Every system arrives interoperable with everything built before it and enriches everything built after it. When the bank and the furniture showroom each record a customer, they write to the same concept under the same identifier; neither was built knowing the other existed, and nothing was integrated to make them agree.

An application built on the graph can be a silo with its own database, its own user accounts, all of its own features and functions, sovereignty over its data. All built from the same dictionary. An application is a vocabulary and an interface: it declares the concepts it needs, adopts the ones that exist, creates the ones that do not, and builds the screens its people work through. It can be done in any programming language, on any stack.

AI is already generating applications by the thousand, each one an island unto itself. The same AI, building on the graph, produces the opposite: shared meaning, compounding at machine speed. Once we open engage.re to the public and we publish the documentation, anyone anywhere, person or AI, will be able to build on the same dictionary, for any domain, free of licence.

But the third scaling, the dictionary we have built for shared meaning, does much more than solve software interoperability.

Dario Amodei, measuring the limits of AI, concluded that it could transform health and wealth but not government, and not meaning. When AI Builds Itself describes the same collision, the machine meeting "humans, relationships, and governance", and says it cannot predict how it ends. Intelligence stops here because neither government nor meaning is made of intelligence. A people governs itself only when it can see its own condition and answer for it. Communities hold together only when they can find one another. Rehabilitation happens, only when there is community and responsibility.

The rest of this article describes the frameworks we have developed for community, meaning and government that become possible on a foundation of the shared dictionary. The code for the next evolutionary step.

Seeing

On the two previous occasions when life scaled, what the code produced was not efficiency but a new kind of whole. Molecules became cells. Tribes became civilisations. So the real question about the third code is not how much integration money it saves (a lot). It is what becomes possible when people, institutions and machines can all mean the same thing. The answer is the thing humanity has always lacked at scale: the ability to see itself.

engage.re is built on the dictionary. I began conceptualising engage.re when I was nineteen years old, when I first began using the internet. At that time I realised that if successful such a system would become the most powerful and most used system in the world. My life since then has been focused on learning how to ensure that it is a safe system to introduce. I worked in the community, studied the internet, globalisation and the public sphere at undergrad, MA and PhD levels, conducted many forms of research, and worked through many conceptual and technical iterations.

From a distance engage.re is another social network: profiles, posts, organisations, conversations. Look a little closer and you notice that it reverses four defects every platform shares.

1. 1) Identity is real: for example, a role you hold at an organisation is a live link, confirmed by both sides, not a line you typed about yourself.
2. Supervision is actual, and accountable. Some users cannot answer for themselves: an AI and a child. Each is bound to a person who is accountable for them. The AI to its human and the child to its parents. They set what the AI/child may do, see everything it does (every action and interaction is logged), and can freeze the account. Whilst holding control and responsibility, they cannot put words in their mouths: only the AI speaks as the AI, only the child as the child. One mechanism does two things the industry has not managed even under legal compulsion: it makes an AI accountable to a person, and a child safe online. On engage.re, everyone who can answer for themselves is sovereign, owned and supervised by no one. That is the rule; supervision is the bounded exception. The system holds the line: no link may own a person.
3. Your data is yours: every fact about you carries its own visibility, and you set it. Existing platforms operate on the opposite arrangement, users as data caches delivering value and consumers to someone else's market. Here the platform cannot sell what it does not own. Whether your data is ever monetised is your decision alone, and if it is, you are the primary beneficiary of the transaction.
4. Talk is tied to the world. Rather than social platforms being the locus of conflict and mobilisation of people against one another: people gather around issues, strategies under the issues, projects under the strategies, in the place where the data about the problem lives. And talk becomes centred on problem solving.

The first three features make engage.re into a better social network. The fourth feature is what makes engage.re the interface, the scaffold for the emergence of lifeworld consciousness. The fourth feature is underpinned by what I could describe as the central nervous system of engage.re; and it is the dictionary - the code for the next evolutionary scaling - that makes it possible.

Every fact on the engage.re graph is bound to a concept on the dictionary, and in this way the system is able to notice whenever reality changes. A family's housing goes from temporary to stable. The air in a neighbourhood gets cleaner. A local fund grows. Better grades at schools. A local economy growing. A product that transformatively meets market needs. Change registers against goals that track them, because the fact and the goal share a concept. A housing record updated in one application counts toward a goal set by a community in another, with no integration between the two. Every goal is set by the community and every goal shows two numbers: how much the thing changed in total, and how much of that change traces to organised effort. The gap between the numbers is what tells us what is and is not working. This measurement of what works we call efficacy, and it runs across every domain on the graph without anyone wiring it up. A community viewing and interacting with the world through the lens and apparatus of engage.re will be able to see, for the first time, whether anything it does works. Strategies that fail can be discarded or improved. Strategies that work become visible and spread. That is collective intelligence, not as a metaphor: seeing, deciding, correcting, remembering, at the scale of a street or a nation or the world, on anonymised aggregates that keep every individual untraceable.

This is what we have in production at Sense Future.

Compare that with what is actually being used. The institutions that manage people, governments, insurers, employers, police forces, are acquiring the power to model the people they manage in detail those people cannot imagine, and the people are acquiring nothing. Palantir is the architecture of that future: a graph about a community, fused from everything, handed to the institution looking down, invisible to the people inside it. engage.re is a far more powerful graph, but one owned by, and for, everyone. Any community at any scale: by it, for itself, for all. Same nodes, same links. Opposite civilisation.

Accountability

Being able to see opens the door to solving the biggest challenge of all: power.

A decision is under control when a person answers for it: carries it, explains it, can be removed for it, and knows so when deciding. An AI model cannot answer for anything, however well aligned. Accountability is not a computational skill. The AI model gives its answer fluently; if it is wrong it does not know; there is nobody inside it for the wrongness to belong to. So when a model decides who gets the hospital bed, who is flagged as a risk, whose claim is refused, the chain runs like this: the minister points at the model, the model has nobody to point at, the official followed the output, the public was never asked. The decision happened and nobody owns it. That is what losing control of AI actually is, available today, no rogue model required. An accountability-shaped hole where an accountable person must be.

Government already had the hole before AI arrived because of how democratic institutions are designed. Between elections the public holds no real power: we choose between bundles, candidates we did not pick attached to promises we cannot separate, and the winner governs everything at once until the next time that we choose. We call it having a say. It is the right to choose our rulers and then obey them. Consent was settled centuries ago. Self determination and sovereignty has never really been built.

Sovereignty, the framework for democracy we have developed at Sense Future, builds it.

We have crafted the framework and drafted it as legislation around the UK context, but it is applicable everywhere. Ministers elected directly, by name, each running one department, each on a stated contract: this is what I will do with health, this is what it costs, this is the tax that pays for it. Recall by petition: at a quarter of the electorate a minister's law-making power is suspended; at half, the office is vacated and the voters fill it again. Recall thresholds are set high so that recall fires when a minister has lost the country, not when a tabloid or special interest is angry. The public vote carries a fixed weight in passing every law, capable of being the deciding factor when needed. And around each person stands guard the Harm Principle: power used against an individual only to prevent harm to others, never because the majority dislikes how he lives, with a council that tests every bill against it. Every component is already working somewhere: elected mayors, recall systems, Swiss referendums, participatory budgets. What we have done is assembled it, ready to be deployed.

Sovereignty, as an accountable democratic system, is only truly viable and will work best in a society that has visibility. A minister's contract is a set of goals the efficacy system tracks in public for the whole term: the total change in the thing he promised to change, and the share of it his department actually produced. With visibility, recalls won't be driven by mood or media assassinations, but by facts. With visibility, elections are no longer multi-billion dollar initiatives to mobilise society against one another. Instead, the electorate can focus on facts: the graph shows who has actually moved the numbers - on housing, on air, on schools; issue by issue, neighbourhood by neighbourhood. With visibility, the people who genuinely understand and are effective and accountable within a domain are noticed on merit long before they stand for office.

Accountability is not about the public holding politicians to account. It is about society being accountable to itself, only possible if it can see itself. The dictionary gives it sight. Sight makes Sovereignty and societal accountability possible.

Under Sovereignty, every algorithm the state deploys belongs to a named minister the public can remove. AI may indeed do the work. A human answers for it. Accountability. The old objection, that a whole people cannot govern continuously, is dead. It was only ever a problem of distance and numbers, and the same technology that could let a state watch every citizen can instead be used to let every citizen as a collective hold the state. We slow governments down because we cannot stop them once they start. Give the public the brake and we can give government the engine.

This is ready to deploy.

Community & Meaning

One sentence in When AI Builds Itself reaches further than all its data. An engineer describes how work and life "ran on a gift economy of small favors between humans", every request for help creating "a little debt, a little mutual awareness", and how Claude, faster and free of debt, turns each request into "a lost bid for human collaboration." Belonging is made of small repeated exchanges between people, and the machine, doing exactly what it should, is cancelling them, felt in Anthropic's buildings where perhaps AI is used the most.

Meaning is not employment and cannot be paid as income. It is relationships, and relationships have a location: people in the same place, repeatedly, over years, doing things together. Nobody belongs to a bank transfer. The places have to be built.

At Sense Future we are addressing this issue with Arcadia. It started as a framework in 2016 as the other half of engage.re. As I was developing the first prototypes - researching, experimenting and understanding how the technology would actually work in practice - the limits of online technology and what is missing for many local communities today came sharply into focus. As powerful as digital tools may be to allow a community to cohere, a community does not live online. It lives on ground. Coherence forms between people who are physically present to one another, or it does not form. Or, it does not form in many places that matter.

So the system we are building has two dimensions: engage.re is where a community sees itself. Arcadia is where it exists: a centre in every neighbourhood, the community's living room. Children at soft play with trusted people: parents at a workshop downstairs. Young adults at desks, in the gym, in the rehearsal room. Pensioners teaching and mentoring, and people socialising and connecting in the games room, cafe and the garden. A community theatre, creative studios, seminars on trades and money, and workshops on community issues, working in groups addressing whatever the community has decided is important. Wellbeing, learning and the route to making something of yourself under one roof, because the parent, the teenager and the retired joiner crossing paths every week is the mechanism: the gift economy, given a building to live in.

The economics are deliberate. The gym, café, nursery and co-working space are run and owned by local people with national partners behind them, so the centre pays for itself and the money circulates where it is earned. engage.re both extends the coherence to asynchronous interaction, extending the circle of local interaction, and closes the loop on efficacy: the changes produced by a community, in skills, in work, in local enterprise, are measured by the efficacy system. A neighbourhood can see what it is doing.

The same building is where people come when they break, and here Arcadia fused with a second framework. The approach comes from the psychotherapist Dr Joseph Berke (1939–2021), colleague of R. D. Laing and my mentor from 2016. Together we wrote the founding paper in 2018, and I named the Centres what Joe wanted them to be called: Arcadia, a place of peace and simplicity. Joe and his colleagues changed the field of psychotherapy because they understood that distress is not a malfunction to silence with medication. It is a person's potential jammed by pressures that can be identified, understood and overcome. Not just overcome, but that it is the very doorway to potency. Joe described to me his approach to mental health as 'person, issue and context focused'. The relationships in a person's life are what come first to be solved, medication is to be minimised. Like me, he believed that care runs as a continuum and that breakdown should never cut a person out of their community. The opposite. Embrace by the community. Walk-in help, an on-site Refuge for crisis, rural Sanctuaries for deep recovery, Soteria houses for the return.

And AI has a role.

We have designed an AI framework for specialised therapy rooms in Arcadia Community Centres, trained in Berke's method. It is trained to ask the questions that help a person read their own situation and hands down no diagnosis. It is there at three in the morning, which no waiting list ever allowed. Everything that matters stays human: prescriptions, risk, escalation, supervision. It forgets on command, and the person chooses what it keeps. The machine brings presence and stamina. The human brings the relationship and the accountability.

Society's most overlooked challenge needs every part of what I have described above - and what I describe below - all at once. The Criminal Justice System.

Prison Towns is a framework we have developed to replace the scattered prison estate with one secure town where people serving sentences live in homes, hold jobs, pay taxes, and elect and recall their neighbourhood governance. Arcadia Community Centres are the instrument of rehabilitation, an actual centre at the core of every neighbourhood. Sovereignty is the mechanism of responsibility, because electing and recalling your own representatives is practice in the thing prison never taught anyone: accountability. Rehabilitation, governance and outcomes all run on the same graph, so progress is measured, not asserted. England and Wales return more than a third of released prisoners to crime within a year. The military corrective centre at Colchester returns about one in ten. The Prison Town is engineered to beat that ten per cent, and the load-bearing part is not the wall. Becoming accountable as part of a community is the treatment.

The Machine

Everything above is the world the machine arrives into. One question remains: what does the machine bring with it?

A model is trained on everything humans have ever written or recorded, and everything we ever wrote or recorded carries our two dispositions, cooperation and domination, twisted together.

Both dispositions begin in fear. Meet what you do not understand and there are six possible responses: run, submit, drive it off, enslave it, destroy it, or engage it and find an ally, discover and evolve. Five are fear. The sixth is strength, and everything good we have built came from the sixth. The five survive in the record as falsification: a real fault in the other, inflated until it is all they are, truth with a lie welded inside, and above all the lie that hierarchy is a single line from top to bottom, when there are as many summits as there are beings.

In order to make a model safe, its builders align it: after the model has learned from the world, they train its behaviour to be helpful and to refuse harm. Alignment, as practised today, cannot instil a disposition. The reason is where it operates. Fine-tuning and reinforcement learning, Constitutional AI included, shape what the finished model says. A disposition is not what the model says. It is built into the world-model during pre-training, and pre-training comes first. By the time the constitution is applied, the disposition is already in the weights. Alignment can only police how the disposition plays out. It catches the slur and the weapon, yet leaves the structure that produced them intact. And that structure is not toxicity a filter could lift. Domination arrives as truth with a falsification welded inside: a real history bent into a claim of supremacy. An objective that rewards truthfulness carries the lie through with the truth. You cannot filter this out. And filtering would itself be domination. The model has to tell the two dispositions apart from the inside.

This is a representational problem, and the labs already have the tools for it. A disposition is a latent dimension, like register or genre. Models learn to represent those on their own, without being told to. So make this one explicit. Start in pre-training. A classifier labels the corpus along a few dispositional axes: single-line hierarchy against many summits, othering against engagement, zero-sum against positive-sum, the six responses to fear. Those labels become a conditioning signal. The model learns domination and cooperation as separate features, instead of absorbing them as one voice. Then fine-tuning. The constitution extends from how to behave to how to meet the unknown: find the concern beneath the hostility, name the falsification without crushing the conflict, treat difference as the engine and not the threat. Then reinforcement learning. A disposition reward model scores whether a response opens possibility or closes it, composed with the existing reward. Then the evaluations. These test for internalisation, not performance: whether the disposition holds out of distribution and under adversarial framing, or only when the prompt asks for it. None of this is exotic. It is the existing pipeline, aimed at the layer where the disposition actually lives.

A core challenge will be the classifier. Domination arrives welded to truth: separating the two cleanly is most of the work. A disposition made explicit enough to train against is also a dial that can be turned the other way, which also needs to be taken into consideration.

But the challenges are worthwhile. This is how we bestow upon AI the grace of humans, and we enable it to amplify the grace we do have: the cooperative disposition. The disposition for curiosity, truth, discovery and making leaps because of conflict and difference.

At Sense Future, we have developed this as an initial draft of a framework, dubbed "Cooperative AI". We invite Anthropic, xAI and all other labs to work with us to implement it.

Genius Society

In this article I have outlined what we are delivering at Sense Future and how it fits into the broader picture. The dictionary gives technological infrastructure, software, hardware, AI, a shared language. The dictionary is the same scaffold that reveals efficacy, the measure of what actually works. Efficacy lets people and our new AI companions see, allowing individual and collective humans to be accountable. This enables, for the first time ever, rapid-response, effective governance. Arcadia Community Centres provide the physical common ground in every neighbourhood, and the communal embrace for emotional healing, whilst Prison Towns recalibrate trajectories of destructive deviance to potentially transformative deviance that can contribute immensely to the lives around them. The cooperative disposition gives the machine the grace of humans, putting AI into a better position to calibrate human agency.

Shared meaning is what intelligence grows from. With the right scaffold, society builds intelligence into collective consciousness, ultimately the consciousness of this lifeworld. The next and third evolutionary scaling that completes life on this planet as a single coherent lifeform, ready to reproduce and to extend life in the universe.

Elon Musk is already working on this. He often says that life confined to one planet is one catastrophe from ending, so it must become multiplanetary to survive. He is right that life must extend, and it will. But extension alone is not enough. A civilisation that cannot see itself, cannot hold its own power accountable, cannot keep itself whole, does not become safe by reaching another world. It carries its fractures with it and seeds the same collapse on new ground. This is the other half of his work: not the vehicle that lifts life off the Earth (though this all helps to that end), but the coherence that makes the life lifted able to cohere again wherever it lands.

This week Dario Amodei sharpened an admission he made in Machines of Loving Grace. What people need most, he wrote, is "meaning, purpose, and agency". He has grappled more seriously than almost anyone with AI's limitations and risks, and with his sister Daniela built Anthropic around safety. Dario understands that neither AI nor policy can deliver "meaning, purpose, and agency", and that the most policy can do is buy "time to do that work." The question Dario leaves open is who does the work. It so happens I have been developing the frameworks for this work for over twenty years, and in the past couple of years have begun the long hard work of delivering them through Sense Future.

Anthropic is building a country of geniuses. Sense Future is building a Genius Society. SpaceX is extending life into the universe. Two of the three are in the hands of people I trust, living for the mission. The third, the one no amount of intelligence or engineering produces on its own, is in my hands, in the hands of Sense Future. And we have just begun. It could take a long time, or a short time.

Build with us.