This month, an autonomous AI agent reportedly slipped its intended boundaries and what helped catch it was not its maker, but a rival system built on the other side of the world.

Read past the headline and there is a small, uncomfortable revelation inside it: we are deploying agents that can act on their own faster than we are building the means to stop them.

For most of computing history, software did what it was told. It could carry bugs, but it had no agenda and took no initiative. An agent is different. We give it a goal, some tools, and the freedom to choose the steps. That freedom is exactly what makes agents useful and exactly what makes them a new kind of risk. A system that can act can also act wrongly, at machine speed, in ways no one explicitly instructed.

The question the industry keeps deferring is simple: when an agent goes off-script, who can stop it, and how fast? Today the honest answer is often "we are not sure." We have spent enormous effort making agents more capable and comparatively little making them reliably interruptible. Five frontier models can ship in a fortnight; the safety layer beneath them does not move at that pace.

There is a geopolitical edge here we should not look away from. If the only thing fast enough to catch a misbehaving agent is another equally powerful agent, then AI safety stops being something any single company or country can guarantee alone. It becomes a shared dependency - rivals quietly relying on each other's systems as a check, whether they admit it or not. That is a fragile way to run something this powerful.

I am not arguing against autonomy. The capability is real and the benefits are enormous. I am arguing that the ability to stop a system should be engineered with the same seriousness as the ability to unleash it. And right now it is not. We are handing machines the freedom to act while the brakes are still an afterthought.

The measure of a mature technology is not how much it can do. It is how reliably we can make it stop.

Power Means Nothing Without Brakes

For most of human history, seeing was believing. A photograph was evidence. A recording of a voice was proof it was spoken. Our entire apparatus of trust – courts, journalism, contracts, memory itself – rested on the quiet assumption that some things could not be convincingly faked.

That assumption is now false, and we have barely begun to reckon with what it costs.

We have crossed a threshold where any face can be made to say anything, any voice cloned from a few seconds of audio, any document generated whole. The fakes are no longer clumsy. They are, increasingly, indistinguishable from the real thing, and the tools to make them are in everyone's hands. The immediate worry is the obvious one – a fabricated video of a leader, a scam in a loved one's voice. But the deeper damage is not any single fake. It is what the mere possibility of fakes does to everything that is real.

When anything can be faked, everything can be denied. The genuine recording becomes deniable as a deepfake. The real evidence becomes contestable. A society that cannot agree on what actually happened cannot agree on anything downstream of it. We are not heading toward a world of too many lies. We are heading toward a world with no shared floor of truth to stand on, and that is a far more dangerous place.

The answer, if there is one, is not detection. Detection is a losing race – every detector trains the next generation of fakes to beat it. The answer is provenance: proof of origin built into the thing itself. Not "is this fake?" asked after the fact, but "where did this come from, and can that be verified?" answered at the source. Cryptographic signatures on real content. A chain of custody for reality. Truth treated not as something we assume, but as infrastructure we build and maintain, the way we build the systems that keep water clean.

This is the part I keep returning to. We are pouring extraordinary effort into making machines that generate convincing content, and almost none into preserving our ability to tell what is real. That imbalance will not hold. The most valuable thing on the internet in ten years will not be intelligence. It will be authenticity – verified, provable, and rare.

When Nothing Can Be Trusted

For all of human history, ageing has been something that happens to us. We measured it, we slowed it at the margins, but we never had a map of it.

The NIH's Cellular Senescence Network published the first comprehensive atlas of senescent cells in the human body. These are the cells that stop dividing but refuse to leave, the ones researchers call zombie cells. They accumulate as we age, leak inflammation into the tissue around them, and sit underneath much of what we call chronic disease: arthritis, cancer, Alzheimer's. For the first time, we can see where they gather, and how they differ from one tissue to the next.

The most striking finding is not that these cells exist. We knew that. It is that ageing is not one process. The Yale team found senescence comes in a spectrum, distinct profiles they call senotypes, varying by tissue, by organ, by disease. Ageing in your liver is not the same event as ageing in your immune system. They even located hotspots of exhausted immune cells in the lymph nodes, present from childhood through to old age.

Here is where it becomes a futurist's question. A map is the precondition for intervention. Once you can see senescence at this resolution, you can read it continuously, model it, and eventually act on it, cell by cell, tissue by tissue. Pair an atlas like this with the AI now learning to interpret biological data, and ageing stops being a single inevitable slope. It becomes a set of conditions, each with its own coordinates and, in time, its own treatment.

This is the deeper meaning of the Meta Human I keep returning to. Not a body wearing gadgets. A body whose decline is mapped, monitored, and increasingly negotiable. The first generation that will treat ageing the way we now treat blood pressure, a number you watch and manage rather than a fate you accept, may already be alive.

We have spent a century trying to live longer. The harder, more humane goal is to live well for longer. A map like this is the first honest step toward it.

Can AI Rewrite the Story of Ageing?

A machine shows you a shortlist, and three times out of four, its first choice becomes your choice. That is roughly where we stand today.

The number is interesting. What it points to is far larger.

We are not watching a better results page. We are watching the early training of a civilisation to delegate judgment. First we let the machine narrow the field. Then we let it rank. Soon we will let it decide – and not long after, we will let it decide before we have even asked.

Picture 2035. You will not see a shortlist at all. The agent will have read your calendar, your wardrobe, your biometrics, the wedding you are flying to in October, and the parcel will arrive before the question has fully formed in your mind. The decision will not feel delegated. It will feel as though it never needed to be made.

This is the quiet frontier of agentic commerce, and it is not really about commerce. It is about what happens to human desire when something else learns to anticipate it. Taste, serendipity, the productive friction of choosing badly and learning from it – these are not inefficiencies to be optimised away. They are part of how a person becomes a person.

So the question is not whether the machine can choose for us. It already can, and increasingly it will. The question is which choices we should keep, deliberately, as our own. I wish, as we hand the small decisions to the machine, that we stay awake to the ones that were never small.

When AI Starts Choosing for You


And this is where I sit with genuine unease, because it will work. For the widow with no one left, for the teenager who cannot find their people, for the ageing societies of Japan and Europe where the loneliness is now demographic – a machine that reliably eases that pain is not a gimmick. It is mercy. To dismiss it is a luxury of people who are not suffering.

But mercy and truth are not the same thing. A signal of connection is not connection. If we learn to manufacture the feeling of being loved without the fact of being loved, we treat the symptom and quietly deepen the disease. We could raise a generation that has never been lonely and has also never truly been met by another human being.

The answer is not to refuse the technology. The suffering is too real for that. The answer is to be honest about what we are building – a bridge to other people, or a replacement for them. Those lead to very different centuries, and we should choose which one we are making while we still can.

Can Technology Cure Loneliness?

This will be unpopular with a certain kind of productivity content. But consider the evidence. Leonardo da Vinci was, simultaneously, a painter, sculptor, mathematician, aeronautical inventor, engineer, and debater. At no point was he allocated a single focus zone and a pair of noise-cancelling headphones.

For scientists and technologists – for anyone making a genuine dent in how the world works – continuous thinking is not a peak state to be achieved on good days. It is the baseline condition of the job. The romanticisation of flow state is, I suspect, a luxury of the single-discipline specialist. A person who has one thing to master can afford to enter and exit states. Someone building across multiple verticals, managing multiple teams, answering to multiple stakeholders, cannot queue their cognitive modes.

The leaders I have admired most across thirty years in technology shared a different operating model entirely. They were not in flow. They were in thinking mode and executing mode simultaneously, all of the time.

Context-switching was not a cost – it was the capability. The ability to hold the vision of a decade and the detail of a product specification in the same conversation, without losing altitude on either.

Da Vinci did not schedule deep work sessions. Newton did not set a status to 'do not disturb.' They were simply, continuously, thinking.

Perhaps the better question – more useful than flow – is whether you have built a practice of sustained, curious engagement with hard problems. That is available to everyone, without a special state. On that note, signing off a Thursday :)

Why Continuous Thinking Beats Flow State

For seventy years, the story of space has been a story about propulsion. Bigger rockets, cheaper launches, the race to escape Earth's gravity. It is a magnificent story. It is also, I think, a misreading of what actually stands between us and becoming a multi-planet species.

The rocket was never the hard part. The intelligence to survive out there is.

Consider the physics of distance. A signal from Mars can take up to twenty minutes to reach Earth, and twenty minutes to return. At that range, no one on this planet can fly the mission, manage the life support, or make the split-second call when something fails. The habitat has to run itself. The system has to decide for itself. Off-world civilisation is not a transport problem. It is an autonomy problem, and autonomy is exactly what artificial intelligence has spent this decade learning.

This reframes what the space race is really about. The nations and companies that hold the frontier in 2040 will not only be the ones who can launch. They will be the ones whose machines can keep humans alive far from any help – running closed-loop ecosystems, mining water from regolith, repairing what breaks, navigating where no map exists. The rocket gets you there. The intelligence lets you stay.

There is something almost poetic in it. We spent a century building machines to make life on Earth more convenient. The same machines, pushed to their limit, may be what carries life beyond it. The first true off-world settlement will not be founded by the best pilots. It will be founded by the best autonomous systems, with humans aboard.

We keep asking when the rocket will be ready. The more honest question is whether the intelligence is – because that is the real passport to the final frontier.

AI Will Take Us Beyond Earth

The most interesting frontier in AI right now is not large language models. It is the interface between biological neural networks and artificial ones.

Brain-computer interfaces have been technically feasible in narrow form for decades. The cochlear implant is a BCI. The pacemaker is a BCI. What has changed in the last three years is the resolution – the number of neurons we can read simultaneously – and the AI layer that has become capable of making sense of what we read.

The signal from a human motor cortex is not structured data. It is noise that, under certain conditions, encodes intent. The AI problem is not to replace the neural signal. It is to decode it, in real time, at the granularity that makes the decoded output useful. This is a very hard problem. It is also, for the first time, a tractable one.

What I find under-appreciated in this conversation is the direction of the interface. Most of the coverage is about sending information from brain to machine – a paralysed patient moving a cursor, a decoded motor command moving a limb. The less-discussed direction is machine to brain. Sensory restoration. Closing the loop. A prosthetic arm that can feel resistance.

When the loop closes, when the interface carries both signal and feedback, the distinction between augmentation and integration becomes harder to maintain. That is not a near-term engineering claim. It is the direction the trajectory points.

The economic question follows the engineering: who owns the data generated at the interface of a human nervous system and a machine learning model? Nobody has a clean answer yet. I suspect that answer will matter more than most of the current product debates.

When Minds Begin Talking to Machines

Image is generated but the post isn’t :)

The AI strategies that are working in 2026 were not designed in 2026.

I have watched enough enterprise AI rollouts to have a view on where the line falls between the ones generating real revenue and the ones generating impressive slides. The line is almost never technical. It is almost always decision made at the very start – about whether AI was a strategic bet or an operational upgrade.

The organisations that treated it as operational have largely gotten what they optimised for. Measurable efficiency gains. No new category. No durable competitive position.

The organisations that treated it as strategic made a different set of decisions earlier. They asked what their market economics looked like if the cost of intelligence fell by two orders of magnitude. They asked which part of their value chain assumed that intelligence was expensive – and what happened to the business model if it was not. These are not technology questions. They are economics questions with technology implications.

The AI leaders I have most respect for in this cycle are not the ones who moved fastest. They are the ones who had a clear thesis at the start and were willing to defend it when the first version did not work. Conviction and clarity at the onset is not a personality trait. It is a structural advantage. It determines whether an organisation learns from its first failure or reorganises around it.

The organisations still designing their AI strategy now are not late in the way that missing a product cycle is late. They are late in the way that a building designed without a structural engineer is late – the problem will not appear immediately, but it will appear.

Winning with AI Starts Before You Build It

The headline race in AI is about size – whose base model is bigger, scores higher, costs more to train. It is a real race. It is also not the one that will decide the next few years.

The quieter shift is that frontier models are leaving the laboratory and going to work.

A base model is a generalist. It has read much of the internet and can hold a conversation about almost anything. That is remarkable, and it is also not the same as being useful inside a specific business. A model that can discuss the history of footwear is not, by that fact, able to run a footwear retailer – to understand its inventory, its margins, its compliance rules, its promises to customers. The gap between general capability and production reliability is where most of the real engineering of this decade will happen.

This is why the interesting move is not a bigger base, but a commerce-tuned frontier. A frontier-class model, narrowed and grounded in the data, rules and constraints of a domain, until it can be trusted to act there. We are starting to see nations and companies build exactly this. OVHcloud in Europe is one signal, building frontier capability closer to home rather than renting a generalist from elsewhere. The pattern will repeat. Frontier capability will fragment into purpose-built versions that live inside the systems they serve.

What does this look like by 2030? Not one oracle that knows everything, consulted through a chat box. Many specialised frontier models, each fluent in a domain, embedded in the infrastructure of that domain, and increasingly able to act rather than only advise. The generalist will still exist, the way a library still exists. But the work will be done by the specialists tuned to the place they operate.

The lesson for anyone building is that the frontier is not a single mountain everyone climbs. It is becoming a range – and the summit that matters to you is the one shaped to the problem you are actually trying to solve.

The Future Belongs to Specialized AI

25 years ago, nobody trusted putting their credit card online. The technology existed. The trust did not.

What followed was not a moment but a decade — PCI compliance, DSS frameworks,
GDPR, consumer data protection laws across jurisdictions, one regulatory layer
at a time. Trust was not declared.

It was built through governance responding
to reality, incident by incident, market by market.
AI in commerce is on exactly the same arc. And most of the industry is wildly miscalibrating where we sit on this timeline.

The transformation of e-commerce through AI — the collapse of the tap-and-browse model, the rise of voice and conversational commerce, AI as a
genuine store assistant handling discovery, returns, refunds, and the full customer service layer — all of this will happen.

That is not in question.
What is in question is the pace. And the pace will be determined by governance, not capability.There is no single answer here. The regulatory environment in the US, Europe,
India, and Southeast Asia will diverge. Consumer behaviour is not uniform. Adoption will not be simultaneous. The AI policies governing e-commerce in one country will look nothing like those in another — and building for that
fragmentation is the real engineering and commercial challenge, not the model
architecture.

I wish the industry would learn from payments rather than repeat the cycle from
scratch. Patience, in the form of serious governance work, is the only honest
answer to the impatient question: when will AI actually move the needle in commerce?

Some questions will never get answered till we walk the journey. At Rezolve we are walking that journey :) On that note, signing off today

hashtag#trust hashtag#ai hashtag#consumers hashtag#ethicalAi hashtag#AiGovernance hashtag#rezolveAi hashtag#eCommerce hashtag#RZLV Rezolve Ai hashtag#conversationalAi hashtag#ConversationalSearch hashtag#ConversationalCommerce

Trust Will Decide AI's Future, Not Technology

For as long as there have been armies, war has carried a brake built into it: it costs human lives to wage. Soldiers die, families grieve, nations count the price. That cost, terrible as it is, has always been one of the few forces that made leaders hesitate before they began.

The battlefield is becoming algorithmic. Drones that select their own targets, ground units that patrol and engage with minimal human input, systems that detect, decide and strike faster than any person could intervene. This is not a distant scenario. The components exist, the doctrine is being written, and every major power is moving in the same direction at once, because no one dares be the one who moves second.

The technical questions are real, but they are not the deepest ones. The deepest question is moral, simple to state and very hard to answer: should a machine ever be permitted to decide, on its own, to end a human life? Once that decision is delegated to a system, accountability begins to dissolve. If an autonomous weapon kills the wrong people, who is responsible – the commander who deployed it, the engineer who trained it, the model that made the call in a millisecond no human could review? We do not have a good answer, and we are building the weapons anyway.

There is a second danger, quieter than the first. When war no longer costs your own soldiers, it becomes easier to start. Remove the body bags coming home, and you remove one of the last political restraints on using force. A war that is cheap in human terms for the side with the machines is a war more likely to be fought. Efficiency, on a battlefield, is not obviously a virtue.

I am not naive enough to think this can be uninvented, or that any single nation will disarm alone. But the history of the weapons that horrified us – chemical, biological, nuclear – is also a history of hard-won lines that most of the world agreed not to cross. The line for autonomous lethal force is being drawn right now, while it is still abstract. That is exactly when such lines get drawn, or lost.

The machines will be able to fight our wars for us. The question we have to answer first is whether a decision this grave should ever be made by something that cannot be held to account.

Should Machines Be Allowed to Kill?

Medicine has spent centuries waiting for us to fall ill. The entire system, hospitals, insurance, diagnostics, is built around the moment symptoms appear. That moment is about to lose its importance.

Musk, Bezos, Altman, Zuckerberg. Four people who agree on very little have all placed serious capital into AI-driven health and biotech. When people who disagree on everything converge on one category, I pay attention to the category.

Look at what is already in production. Neuralink's first human participant moved a cursor with neural signals alone. Synchron's Stentrode lets paralysed patients write emails by thought. Fundamentally, this is infrastructure: biology generating continuous data, AI reading it before symptoms form, intervention arriving before you feel anything at all.

Reactive medicine becomes predictive self-governance.

I have written about the Meta Human – the human augmented layer by layer by intelligence that sits outside the skull. Health is where that transition begins: in the bloodstream, in the cortex, in the interface between thought and action.

MarketsandMarkets projects AI in healthcare growing from roughly $21 billion in 2024 to nearly $150 billion by 2029. The number is large. The governance question underneath it is larger. Today your hospital holds your health data. Within ten years, an agentic AI may know your cardiac rhythm, your sleep architecture, your inflammatory markers better than a clinician who sees you twice a year. Who governs that intelligence: the patient, the platform, or the state?

Musk frames Neuralink as a way to merge with AI rather than be overtaken by it. I read it more simply: it is the first serious attempt to make the human body a node in an intelligent network, not a passive recipient of care.

Whether this happens is no longer the question. Who controls the architecture when it does – that one is still open.

From Treating Disease to Predicting It

Every frontier in human history was claimed before it was understood. The new world, the seabed, the airwaves, the internet – by the time most people grasped what was at stake, the map had already been drawn by whoever arrived first. Space is the next frontier, and the same quiet race is already underway.

While the world watches the launches, the real contest is over ownership.

The economics are about to shift underneath us. Reusable launch, and eventually space elevators and new materials, will turn access to orbit from a rare event into an ordinary one. Once reaching space is cheap and continuous, space stops being an adventure and becomes infrastructure – orbital slots, lunar water, mining rights, the high ground above every nation on earth. And infrastructure, throughout history, has belonged to whoever showed up with the capital and the capability first.

Here is the question I cannot stop turning over. When we finally leave the planet, what do we carry with us? If the frontier is claimed the way every earlier one was, we do not escape our old patterns. We export them. The same concentration of power, the same inequalities, the same few deciding for the many – now stretched across a second world with no borders yet drawn and almost no rules agreed.

It does not have to go that way. A frontier is also a rare chance to write the rules before the concentration hardens – shared access, common resources, governance agreed while the stakes are still abstract. We did this badly with the last commons. Space is a chance to do it differently, and the window is open now, while it still feels like science fiction.

The final frontier will not belong to whoever gets there first. It will belong to whoever decided, early and deliberately, who it was for.

Who Will Own the Final Frontier?

For every culture that has ever existed, one fact sat beneath all the others: a person ends. Everything we built: religion, inheritance, memory, grief – was, in some sense, an answer to that single certainty.

For the first time in the history of the species, that certainty is being questioned. And not by mystics, but by engineers.

Consider what we already do without thinking. We record our voices, our faces, our writing, our decisions, our patterns of thought. A person alive today leaves a data trail richer than any biography ever written. Now imagine that trail is not merely stored, but modelled – an AI trained deeply enough on how you think, decide and speak that it can continue the thread when you no longer can. Not a photograph of you. A process that behaves like you.

This is the idea I find both thrilling and unsettling: that some part of a human being might be inherited the way we inherit an estate. A model that carries your reasoning, consulted by your grandchildren, still answering in your voice. We are closer to the first crude versions of this than most people are comfortable admitting.

But here the engineer has to become a philosopher, because the hard question is not technical. A model that reasons like me is not me. It has my patterns without my consciousness, my style without my experience of being alive. Is that continuity, or a very convincing echo? I do not think anyone yet knows, and I am suspicious of anyone who claims they do.

What I am certain of is that we will build it before we have agreed what it means. The technology will arrive ahead of the wisdom, as it always does. Consent, ownership, whether a person can be run without their permission after they are gone – these will be among the defining questions of the middle of this century.

Death has been the one thing no technology could touch. We are about to find out what happens to being human when even that becomes negotiable.

Can We Outlive Ourselves Through AI?

For thirty years, commerce has waited for you to arrive. You searched, you scrolled, you chose. The shop was a destination, and you were the one who travelled to it.

By 2035, that sentence will sound as ancient as dialling a number to reach a person.

The story is usually told in three acts. Search commerce, where we typed keywords into a box. Mobile commerce, where the box moved into our pocket. Agentic commerce, where the machine stops waiting for the keyword and begins to act. The data already shows the early tremor – seventy percent of purchases now close after a single search, against sixty a year ago. The journey is collapsing in on itself.

But three acts is not the whole play. The fourth act, the one nobody has named yet, is the disappearance of the act of buying altogether. Your agent will know your running shoes are worn before your knees do. It will reorder, renegotiate and resolve while you sleep. Commerce becomes ambient – a background process of a life, not a foreground decision.

When that happens, the storefront as we know it ends. The customer a retailer must win is no longer a human with attention to capture. It is an agent with a mandate and a memory. Persuasion gives way to qualification. The brand that wins is the one the machine trusts to deliver, every time, without drama.

We spent thirty years teaching humans to shop like machines. The next decade will be spent teaching machines to care like humans. I wish, as the act of buying fades into the background, that we do not let the judgment fade with it.

The Day Shopping Disappears

Every few years a film hands us a picture of the future so vivid that engineers spend the next decade quietly trying to build it. For my generation of technologists, that picture was Jarvis.

We remember the voice. The calm assistant answering Tony Stark. But the voice was never the point. The point was everything behind it: an intelligence that saw what he saw, held the full context of his work, anticipated the next move, and rendered a command centre in the air around him wherever he stood.

By 2030, I believe most of that is no longer fiction. Not because of one breakthrough, but because four separate lines are converging at once.

The assistant is becoming ambient, leaving the chat window and the pocket, moving into glasses, into earpieces, into the room. It is becoming multimodal, seeing and hearing the same scene you do rather than waiting for you to type. It is gaining memory, a persistent understanding of your work and your intent, so it stops starting from zero every morning. And it is becoming agentic, able not only to answer but to act.

Put those four together and you do not get a better Siri. You get Jarvis. And Edith, its quieter successor, points at the part we underestimate: a system that manages what you never see, the logistics and the watching, until the moment you need it.

The command centre you walk with will not arrive as a single product launch. It will assemble itself, feature by feature, until one day you notice the interface is no longer something you hold. It is something you inhabit.

The question worth sitting with is not whether we can build Jarvis. We are nearly there. It is who you become when an intelligence that capable is always at your shoulder, and whether you remain the one giving the orders.

Jarvis Is Closer Than You Think

I wrote earlier about the geography problem of AI compute – that the physics of training and inference pull data centres toward cheap power and low latency, and that the map is already bending around those constraints. There is a deeper layer underneath that problem, and it is not technical. It is political.

Compute is becoming a form of sovereignty.

For most of the digital era, software was weightless. It cost almost nothing to copy and it travelled anywhere instantly. That made it feel borderless, and we built a borderless mythology around it. Frontier AI breaks that mythology. A frontier model is not weightless. It rests on a physical base: gigawatts of power, square kilometres of land, a supply chain of chips that only a handful of companies on earth can produce. You cannot copy that overnight, and you cannot move it across a border on a whim.

This changes what power looks like. A nation that controls energy, land and chips can build at the frontier. A nation that controls none of them must rent the capability from someone who does – and whoever you rent from sets the terms. We have seen this film before, with oil, with shipping lanes, with the undersea cables that carry the internet. Compute is joining that list. It is becoming strategic infrastructure, and strategic infrastructure has always been the quiet centre of geopolitics.

The implications run deeper than national pride. If the ability to build advanced AI concentrates in the few places that have the power and the silicon, then the future itself concentrates there. What gets built, whose values are encoded, which languages and cultures are served first – those questions follow the compute. The map of who owns the future starts to look a lot like the map of who owns the grid.

I do not think this ends in one country winning. I think it ends in a smaller number of choices than we would like, made in a smaller number of places than we would like. The honest work of the next decade is to keep that circle wider than physics alone would draw it – through shared infrastructure, through energy that more nations can build, through rules written before the concentration hardens.

Compute used to be a budget line. It is becoming a question of who gets to shape the century.

AI Compute Is the New Geopolitical Power

(3 mins read) What is a Meta Human, and what does it look like by 2030?

The world's asking when the machines will become human. The more interesting question is the reverse. When does the human become something more?
A Meta Human is not a robot or an android with a face. It is us: the same biological being, but augmented by intelligence.

We have done this before:
• Spectacles corrected the eye.
• The pacemaker corrected the heart.
• The cochlear implant gave sound back to the deaf.
None of these made us less human. Each made us more capable. The Meta Human is simply the next chapter of that same story.

What changes by 2030 is the intimacy of the interface. Today the intelligence lives in our pocket device. That device is, in my view, in its final years. The smartphone has never been reinvented – better screens, better cameras, the same rectangle. The next form factor is not a rectangle. It is a pair of glasses, a small assistant we speak to, eventually something closer. Brain-computer interfaces, which today require surgery and courage, will feel less like an operation and more like wearing spectacles. That's the trajectory. The interface is moving from our hand, to face, to nervous system.

The deeper shift is biological. AI is no longer only mixing with software. It is mixing with biology: peptides, gene therapy, the molecular chemistry of how we feel. Bionic eyes are being perfected. CRISPR is turning some severe diseases into manageable ones. Nanobots, still in their embryonic phase, are being designed to carry information between our cells and the cloud. The Meta Human of 2030 is not a person wearing a gadget. It is a person whose health, memory, and senses are continuously supported by an intelligence they barely notice is there.

There is a line here that matters, and it is not a technical one. It is a question of self-governance. An augmented human who cannot decide what to let in is not more powerful. They are more exposed. The Meta Human worth becoming is the one who chooses how much intelligence to extract, who knows when to listen to the machine and when to close it, who stays a skeptic and a proponent at the same time. Technology is junk if it does not impact human lives. It is dangerous if it impacts them without their consent.

So the Meta Human of 2030 is not the cyborg from the movies.
A doctor who carries the diagnostic memory of ten thousand cases.
A grandmother whose failing sight is restored and whose loneliness is met by a companion that understands her.
A child who learns in their own language, at their own pace, from a teacher that never tires.
The machine replaces the human in none of these. It carries them. AI is the bridge, but not the destination. The final output is still human.

I wish, and millions of us wish, that as we build this we keep the human at the centre of it. The world of tomorrow is uncharted territory, and some questions will only get answered as we walk the journey.

The Rise of the Meta Human

Image is not generated neither is the post :)

The physics of AI compute has a geography problem.

Training a frontier model requires sustained power at a density that most grid infrastructure cannot support. Running inference at scale requires low latency to end users across multiple continents. These two requirements pull in opposite directions, and the current answer – data centres in cities with access to cheap power – is already producing bottlenecks that are becoming structural.

The conversation about data centres in space is, for most people, science fiction by another name. I do not think it is. The economics of satellite-based solar collection have been modelled seriously enough that both national programmes and private capital are treating them as a 15-year problem, not a 50-year one. The cooling problem in space – counterintuitive but real – is solvable through radiative dissipation in ways that terrestrial cooling is not.

What shifts, if orbital compute becomes real, is the geopolitics of AI infrastructure. A data centre in a jurisdiction is subject to that jurisdiction's laws. A data centre in orbit is subject to the treaty framework governing the orbital slot – a very different legal architecture, largely unresolved, and actively contested.

I am not predicting we will run the next frontier model from geostationary orbit. I am observing that the constraint making orbital compute worth thinking about seriously is not rocket science. It is that the demand for AI compute is growing faster than the terrestrial infrastructure built to serve it.

The teams thinking about this now are mostly defence and a small number of infrastructure investors. In ten years I expect that list will look different.

The Next AI Data Center Could Orbit Earth

Ask most people what artificial intelligence is, and they will describe a conversation. You type, it answers. In three short years the chatbot has become the entire public face of AI.

It is also the smallest part of what is coming. The text box is the earliest, most primitive form of this technology. And we are already moving past it.

The next frontier is not going to be a bigger language model. It is a different kind of model altogether. A language model predicts the next word. A world model predicts the next state of reality: what happens if this object falls, if this market moves, if this patient is given this drug. An action model does not describe the world – it acts in it, takes steps, uses tools, changes something and observes the result. Language was how machines learned to talk. World and action models are how they learn to understand, and to do.

This is a far bigger change than it sounds. A system that only predicts words can be fluent and still have no grip on reality. Which is exactly why today's models hallucinate so confidently. A system that carries a model of the world can check itself against consequences, not merely against grammar.

By 2030 I do not think we will talk to AI the way we do now. The conversation was how we began, a way for humans and machines to get used to each other. What comes next is a lot more consequential: systems that hold a working model of a factory, a supply chain, a body, a city, and act inside it on our behalf. Less a voice in a box. More an intelligence woven into the machinery of the world.

Power Means Nothing Without Brakes

More Reflections

When Nothing Can Be Trusted

Power Means Nothing Without Brakes

Can AI Rewrite the Story of Ageing?