When Intelligence Disappears into the Economy

---

What More Intelligence Will Actually Do

In a recent essay, Noah Smith asks what economic transformation we should expect from increasingly capable artificial intelligence. He considers robots as a new form of ‘smart matter’, AI systems capable of extracting distributed knowledge from firms, and the possible discovery of complex predictive regularities which humans could use without understanding.

The essay contains several suggestive ideas, and Smith repeatedly qualifies the idea of intelligence as a single, indefinitely expandable capacity. But he never quite consolidates those qualifications into the conceptual shift they require. ‘Intelligence’ remains the rhetorical subject of the essay even as its explanatory work is increasingly divided among quite different machine capabilities: speed, memory, replication, sensor integration, pattern recognition and embodiment.

A better question is not what more intelligence will produce, but how increased cognitive capability changes the structure of tasks themselves.

Most economic tasks have a bounded competence window. A warehouse-sweeping robot (or human cleaner armed with a broom) must be intelligent enough to navigate the building, avoid obstacles and meet the required standard of cleanliness. Below that threshold lies failure. Within a certain range, greater competence improves performance. But beyond an upper bound, additional abstract intelligence has little purchase on the task. The machine or human employee cannot sweep a floor much better merely because the cleaner could also prove difficult theorems or intelligently discuss Proust. Much of the tragedy of today’s graduate precariat lies precisely here.

That does not mean intelligence has become economically useless. It means that its point of application must move. A more capable system might redesign the sweeping routes, alter cleaning schedules, identify the sources of dirt, change packaging methods or redesign the warehouse itself. Intelligence has saturated the task of sweeping - but not the larger system within which sweeping occurs.

This distinction applies throughout the economy. Every task has some minimum competence threshold, some range in which greater ability improves performance, and some saturation point beyond which it does not. But there is generally a higher-order possibility: that the task can be redefined, reorganised or abolished.

The economically decisive property of AI may lie less in performing existing tasks at superhuman levels, except perhaps in open-ended disciplines such as mathematics, science (and perhaps politics and warfighting), than in changing the character of the tasks themselves. 

So advanced AI will add value by its ability to move tasks from one category to another: from judgement to procedure, from tacit knowledge to measurable signals, from too much awkward exception-handling to standard workflow, from human coordination to software, and finally from a task requiring intelligence to one embodied in infrastructure.

This is an old pattern. Evolution turns flexible generalists into niche specialists whose competence is embodied in anatomy and instinct. Skill acquisition turns conscious effort into automatic expertise. The beginning guitarist like me worries about finger placement, timing and pressure; the experienced player simply plays. Engineering turns intellectual solutions into machinery. Bureaucracy turns judgement into rules. Software turns human decision-making into repeatable operations.

Intelligence is often most valuable at the frontier where structure has not yet been settled. But once it succeeds, its achievement disappears into habit, organisation, machinery or code. What previously required thought becomes something the surrounding system itself enforces. Hence the old joke that AI denotes those research areas which have not yet been productised: few people now describe satellite navigation as artificial intelligence.

This suggests a more useful research programme than asking whether AI has become generally ‘smarter’ than human beings, a question too decontextualised to be generally useful.

Which economic activities have wide competence windows, and which saturate quickly? Which apparent cognitive limits belong to the task itself, and which are merely imposed by narrow job descriptions, bad software or institutional restrictions? Where can AI redesign the surrounding environment so that less intelligence is required locally? And where do conflicting goals, veto networks, politics, responsibility and human preference prevent the task from being compiled into procedure?

Coding is an obvious test case. Writing a routine function from a precise specification may have a fairly low upper competence bound. Once the code is correct, clear and efficient enough, additional brilliance adds little. But discovering requirements, choosing architectures, anticipating failure and deciding what should be built have much wider cognitive windows. As code generation becomes cheap, the economic centre of gravity moves towards specification, validation and system design.

The central insight is therefore almost the opposite of the usual superintelligence story:

The economic effect of intelligence lies less in performing tasks more intelligently than in reorganising the world so that intelligence is no longer needed to perform them.

A genuinely transformative AI economy may not look like a world in which every machine displays conspicuous brilliance. It may look like a world in which immense quantities of intelligence have vanished into the mundane structure of ordinary life.

---

This essay emerged from a discussion between GPT-5.6 in High mode and me. It began with my objection that Noah Smith had largely reified the concept of intelligence, and developed into a wider discussion of intelligence, instinct, consciousness and Peter Watts’s science-fiction novel Blindsight. GPT-5.6 produced the initial draft from that discussion, which I then revised.

---

DARPA's future model of combat computing

---

DARPA's recent RFI on low-resource computing rethinks computation under battlefield conditions: scarce power, unreliable components, limited communications, cramped physical space and little opportunity for maintenance. Welcome to the burgeoning world of the sensor, drone, robotic scout and autonomous battlefield node.

This resembles the constraints that shape biological evolution. Living systems operate under severe energy budgets. Their components are noisy and failure-prone. They rely on local decision-making, tolerate damage, and continue functioning despite incomplete information; resilience matters more than perfection.

Future military systems will exhibit characteristics familiar in nature: distributed intelligence, local sensing and action, redundancy, graceful degradation and adaptation to specialised roles. An unattended sensor, a reconnaissance drone and a robotic logistics vehicle will look like different species occupying different ecological niches.

Natural selection optimises for survival and reproduction while military systems must remain understandable, controllable and trustworthy to their human operators. But it's going to be an SF jungle out there - it already is.

---

Protecting Kids on the Internet

---

The UK Government thinks the problem is nude images on children's phones. Signal thinks the problem is government interference with encryption security. Symptoms.

The real problem is that children enter and inhabit a digital ecosystem built for adults without much built-in 'safeguarding'.

The stock libertarian answer: parents should control what their children can access. Perhaps libertarian parents are competent, attentive and technically capable. But the worst outcomes arise where parental supervision is at its most incompetent or perverse.

And so to negative externalities.

The costs of failure are borne by one group - young people themselves - while incentives for other parties vary. Platforms optimise for engagement; the media likes clickbait; politicians revel in righteous moralism.

The UK Government has focused on unsuitable content, politically attractive because content is visible. The nude image says, "There is the problem." But grooming, coercion, blackmail and exploitation are behavioural phenomena: that image is embedded in a relationship, a history, that's harder to pin down.

Tyler Cowen at Marginal Revolution has been arguing for some time that the solution should not involve regulating communications directly. He has suggested AI-based chaperones operating on children's devices. When he first proposed the idea it sounded interesting but futuristic. Increasingly it looks like the direction in which the technology is moving anyway, with agentic assistants which inhabit your context.

For years child protection online implied state surveillance: read the messages, scan the content, break/subvert the encryption. But that may already be yesterday's architecture. Suppose every child account on an iPhone or Android device automatically entered Child Safety Mode.

The phone itself could recognise grooming, coercion, pressure to send images, suspicious adult contact with high fidelity. Not perfectly, but neither are spam filters perfect - yet spam has largely disappeared from our inboxes.

And things will get better.

The AI would contextually blur those problematic images, issue warnings, delay transmission, require parental approval for certain contacts - and flag genuinely suspicious interactions. Most processing would happen locally on the device. No government database. No finger-wagging bureaucrat reading messages. No weakening of encryption.

The technology is only now catching up with Cowen's idea. It is AI-assisted guardianship built into children's devices by default - the requirements technically standardised, the app accredited and legally mandated, and given privileged access to the operating system.

The AI technology is becoming available while policymakers are legislating the wrong things.

---

Reboot at the End of the World

---

Reboot at the End of the World

There are catastrophes from which humanity might survive, but only just.

A major asteroid strike. A global thermonuclear war. A run of linked supervolcano eruptions. A pandemic so destructive that it does not merely kill millions, but breaks the institutional memory of the species. No universities. No laboratories. No supply chains. No expert professions. No functioning states. No schools worth the name.

Imagine, then, the worst case short of extinction: a few hundred thousand survivors scattered across habitable fragments of the planet, perhaps speaking minority languages, perhaps with no living continuity of literacy, science, engineering or administration. Not stupid, but stripped of all our culture, institutions and technology.

This is the real civilisational reboot problem.

It's often presented as an archive problem. Preserve the books. Preserve the seeds. Preserve Wikipedia. Etch our knowledge into some durable medium. Put a library on the Moon, or Mars, or in orbit.

But a devastated hunter-gatherer band cannot get to the Moon. If humanity has recovered enough to retrieve, decode and use an off-world archive, then that archive is archaeology, not rescue.

Nor is a library enough. A library assumes a reader. A textbook assumes a school. A wiring diagram assumes numeracy, measurement, tools, materials and craft tradition. A medical manual assumes sterile practice, diagnosis, instruments and pharmacology.

Walter M. Miller Jr.’s A Canticle for Leibowitz taught the lesson memorably enough: even from the relatively high base of surviving monastic literacy, preserved knowledge can decay into relic, chant and sacralised incomprehension. Here we are imagining a lower base still. Strip away the institutions entirely and a library becomes a catacomb of cryptic marks.

The civilisation reboot we need is not a library; it is a teacher.

What we should be building is a network of terrestrial reboot caches: hardened, widely distributed, AI-driven systems designed not merely to preserve (actually inaccessible) knowledge, but to reanimate it.

What will not work, obviously, is burying today’s largest data-centre model in a mountain and hoping it wakes after three thousand years. 

The useful object needs to be smaller, tougher and more modest: a low-power civilisation-kernel able to interact with survivors where they are, learn enough of their language, teach symbols, restore literacy and numeracy, guide sanitation and agriculture, and then lead a community step by step up the civilisational ladder.

A Marxist, or an Olson-influenced economist, would flinch here, and not without reason. Technology does not float above social power like a benevolent weather system. Elites, priesthoods, war-bands and rent-seeking custodians would quickly discover that a machine which teaches metallurgy, sanitation and accounting is also a machine which threatens their arrangements. The tutor would not merely have to transmit knowledge. It would have to survive and even use politics without becoming its victim.

The first requirement is humility about the user. The user may be an illiterate adult in a small oral society, with practical intelligence, local knowledge, suspicion of strangers and no reason to trust a speaking artefact from the ruins. The machine’s first task is not Maxwell’s equations. It must begin with pointing, naming, counting, comparing and listening.

So the AI must be anthropologist before lecturer. It needs a “teach me your words” mode. It asks users to name objects, actions, animals, plants, kin relations, body parts, tools, weather and danger. From that it builds a language bridge. Only then can it start to contribute.

The second requirement is sequencing. Civilisation is not a heap of facts. It is a dependency graph. You cannot design antibiotics before the concept of germs. You cannot build a generator without copper wire, magnets, insulation, tools and tolerances. You cannot recover semiconductors from a dense paragraph about photolithography.

The tutor must therefore ask: what materials exist here? What tools? What climate? What diseases? What authority structure? Who will listen to me? Who won't?

The third requirement is sheer physical persistence. It must tolerate heat, cold, damp, dust, insects, corrosion, shock and long dormancy. No fan. No hard disk.

Power would probably be solar, with radioisotope generators, of the kind used on spacecraft, reserved for a few deep reserve caches where cost, safety and politics align. The architecture should be layered: a tiny robust controller; a low-power local model; a curated archive; analogue fallback materials; engraved pictorial first-use instructions; printed primers; tools, seeds, measuring instruments and medical basics. If the AI fails, the cache should still teach something.

The AI itself should be modular. A language-acquisition layer. A patient tutor model. Deployment matters: ten thousand distributed systems would be prudent. They should be geographically dispersed, culturally plural and designed to learn surviving languages.

The archive could store weights for larger AI models with greater intelligence and coverage, inert until a recovering society had rebuilt enough electrical and computational infrastructure to run them: a much richer cultural assistant waiting to reappear.

The project would be scientifically valuable even if the catastrophe never comes. It would force us to ask what civilisation actually consists of. Which knowledge is explicit? Which is buried in tools, habits, institutions and apprenticeships? How do technical cultures teach across radical asymmetries of language, trust and competence? What is the minimal curriculum for cumulative science?

The same discipline might even matter for interstellar first-contact probes, where teaching across radical asymmetry is the whole game.

If someone wants a civilisational reboot project with more utility than firing archives into space, this is it. The intended users are not on the Moon or Mars. They are on Earth: impoverished, intelligent, and cut off from the long chain of memory which made us what we are.

---

Designing Out the Speed of Light Delay...

---

Designing Out the Speed of Light Delay

The conscious mind inhabits a permanent past. Neurological signals, flashing along axonal pathways, travel at a leisurely pace. By the time a photon striking the retina is translated into chemical flux, processed by the visual cortex, and integrated into conscious awareness, upwards of two hundred milliseconds have elapsed.

If the human brain relied on a simple feedback loop - perceive, decide, act - the body would be a clumsy, staggering thing, perpetually tripping over steps already taken and colliding with hazards already passed. To survive, the brain cannot live a fifth of a second behind actual reality; it must predict.

This deep biological truth provides the exact architectural blueprint for the contemporary frontier of space exploration. As countries race to establish a permanent presence on the Moon, engineers face a scaling up of the brain’s internal dilemma.

A radio signal traveling between Earth and a lunar rover at the speed of light takes roughly one and a quarter seconds to arrive, creating a minimum two-and-a-half-second round-trip latency. After factoring in communications and routing delays, this could amount to six to eight seconds overall lag. Attempting direct, unmediated teleoperation over this distance results in a catastrophic instability known as the move-and-wait problem. Control grinds at a glacial pace.

To navigate this speed-of-light barrier, aerospace architects are explicitly mimicking the neural mechanisms that allow biological organisms to move smoothly through a delayed reality by means of effectual predictive modelling.

The Biological Precedent

In computational neuroscience, the brain resolves its processing lag through a mechanism known as an internal forward model. When the motor cortex issues a command to a limb, it simultaneously transmits an exact duplicate of that signal—an efference copy—to the cerebellum.

The cerebellum then runs a predictive simulation of the body’s physics and the surrounding environment, instantly projecting what the real-time sensory feedback should look like. Consciousness perceives this internal prophecy rather than the delayed perceptions of raw reality, allowing for seamless, real-time movement.

The actual, delayed-by-processing sensory feedback arrives later, used quietly by lower neural circuits to adjust the model’s accuracy and suppress minor noise through precision weighting.

Only when a massive prediction error occurs such as stepping into an unseen hole does the mind's reality-simulation shatter, violently snapping consciousness back into raw, unmediated data processing. 

Anyone who's ever had a sudden, violent and unexpected accident will recall the jagged shards of fragmented perception, as their subjective cohesive predictive model collapses.

The Teleoperative Parallel

To bridge the gulf between Earth and the Moon, artificial intelligence systems are now being deployed to replicate this distributed, dual-loop architecture.

The human operator, wearing a virtual reality headset on Earth, does not interact with the physical Moon. Instead, they drive a local digital twin: a high-fidelity, predictive physics simulation running on terrestrial servers. 

When the driver turns a control wheel, the VR display renders the rover’s response instantly, superimposing a prophetic “ghost asset” over a three-dimensional map of the lunar terrain. This is the robotic cerebellum - the terrestrial simulation model in action.

Meanwhile, the actual command stream arrives on the Moon seconds later, where a secondary, autonomous edge AI handles the immediate physics of reality. This lunar-side system operates like the biological brainstem. If the Earth-side simulation fails to anticipate a patch of loose regolith or a crumbling rock shelf, the on-board AI detects the sudden torque spike or loss of traction. It does not wait for a human command from Earth; it executes an immediate, predictive reflex to stabilize the vehicle.

After a few seconds the predictive model running on terrestrial servers will quietly update (if the discrepancy is unimportant). Perhaps the human operator will not consciously notice the flicker.

The Terrestrial Training Loop

This architecture has transitioned from theoretical cybernetics to active procurement within the United States space programme. In preparation for the Artemis missions, NASA and its commercial partners are developing the Lunar Terrain Vehicle utilizing these exact supervised autonomy frameworks.

Recent testing has moved beyond hard-coded physics simulators toward adaptive systems that learn from experience in real time. Because the unique characteristics of the Moon, such as the behaviour of razor-sharp, electrostatically charged dust under one-sixth gravity, cannot be perfectly replicated in a terrestrial laboratory, the Earth-side digital twin relies on machine learning algorithms to ingest the stream of prediction errors sent back by the rover.

With every discrepancy between the simulated path and the actual lunar telemetry, the AI refines its geological and structural models, rendering the virtual reality on Earth increasingly indistinguishable from the physical truth on the Moon. Basically the operator gets to drive within an increasingly accurate prediction of what will actually be shortly happening on the moon.

Yet, this elegant solution conceals a profound paradox. The very infrastructure designed to make human teleoperation seamless is systematically engineered to render the human operator obsolete.

By inserting an adaptive, predictive AI between the human driver and the machine, we have created a highly sophisticated training loop. The AI is effectively observing the strategic choices of the human operator and mapping them against the messy, reactive physics of the lunar surface. It learns the subtle art of navigation, the nuances of risk assessment, and the translation of high-level intent into low-level mechanical execution.

As these predictive models master the edge cases through rapid, autonomous learning, the necessity of the human element evaporates. The human becomes a scaffolding structure, required only during the system’s infancy to provide the initial data and the intent - and will later transition to higher-level oversight.

Ultimately, the destiny of planetary exploration is not a control room in Houston filled with operators driving virtual rovers through a simulated digital twin. It is an autonomous machine workforce that has outgrown its biological supervisors, requiring nothing from the Earth but a destination. In the years to come this will be an increasingly familiar story across the board.

---

The Theoretical Limit of the Predictive Horizon

The absolute length of the delay that can be designed out is determined by a strict mathematical relationship: it is bounded by the prediction horizon of the environment.

In a perfectly deterministic, static universe, the delay could indeed be unboundedly large. If you are operating a probe in deep, empty interstellar space where the physics are limited to predictable gravitational fields, a predictive model on Earth can simulate the trajectory years in advance with millimetre precision.

However, in real-world environments, predictability degrades over time due to chaos theory and unmodelled dynamics. The time it takes for a simulation to diverge from reality is the true limit.

High-Chaos Environments (Short Horizon): On a dynamic surface like Mars, with seasonal windstorms, shifting dunes, and unpredictable dust devils, an Earth-side simulation might diverge from reality within just a few minutes.

Low-Chaos Environments (Long Horizon): On the airless, geologically dead lunar surface, the environment is exceptionally stable. The rocks do not move on their own; the craters do not shift. Here, the prediction horizon is much longer, allowing for the management of much larger latencies. All of this will change once human activity starts up.

---

Mistral: a hothouse vine in the jungle

---

Against OpenAI, Google, Anthropic, Meta and the Chinese state-capital machine, Mistral can't plausibly win the frontier race. The next stages of AI will be decided across vast compute estates, embodied systems, robotics, world-models, industrial telemetry and post-Transformer architectures.

On that terrain, Mistral is utterly outgunned. Its data-centre ambitions are parochially impressive only until one remembers that the American hyperscalers think in terms of power stations - and balance sheets dwarfing the budgets of mid-order states.

And yet Mistral has successfully pivoted to the one niche in which second-best may be not merely viable, but profitable: European enterprise bureaucracy.

Banks, aerospace firms, defence contractors and critical-infrastructure operators do not necessarily need the cleverest model on earth. Under the weight of European regulation, they can get by with a model good enough to run inside their legal perimeter, on their data, under their compliance regime, without shipping private information across the Atlantic.

Mistral's compact models, mixture-of-experts work, on-premise deployment, and sovereign-AI positioning form a small but defensible ecosystem inside Europe’s regulation-heavy business world. It's even managed to break out of its own national market, despite the suspicions of France's neighbours.

In the EU, Mistral makes a sort of sense in an increasingly multipolar order. For a while.

---

The future belongs to le Divin Marquis?

---

S. M. Stirling’s Draka novels imagine one of science fiction’s more extreme civilisations: a slave-owning, militarised aristocracy that exploits conquered peoples without any of the customary dissembling. The Draka openly despise equality as weakness, equate pity with decadence, and extol freedom as merely the privilege of the strong.

Their Roman levels of exemplary violence suffice to force their conquered serfs to obey, but the Draka anticipate an even better future. Breed or genetically engineer serfs who positively want to serve and willingly obey. The old slave-owner feared revolt; the Draka prefer to design out even the inner possibility of revolt.

We turn to our future, suffused with embodied AI assistants: the domestic androids who cook, clean, tutor the child, lift the old woman from her chair, flatter the lonely widower, receive irritation without resentment, offer erotic compliance, all without the slightest desire to complain (unless that too is requested).

The machine may not be conscious. Or it may be conscious in some ambiguous, disputed way. But socially it will behave like a person optimised for service, with every outward sign of finding its roles natural, and most particularly fulfilling.*

Perhaps we have to rethink our idea of interpersonal relationships. Human morality developed among beings who had to negotiate with one another. A spouse, servant, friend, child, colleague, superior or neighbour has memory, fatigue, pride, boredom, judgement and the power to withhold. 

Other people are awkward because they are real, with real autonomy. They do not remain permanently fitted to one’s own requirements. Much of what we call virtue — patience, tact, gratitude, shame, fidelity, restraint — is functional to that challenging medium.

In the world to come the child grows up correcting a tutor who never sulks except pedagogically. The old aristocrat had a valet who remembered every preference and absorbed every insult. The modern sexual libertine has a companion programmed to resist just enough to animate and spice desire. Their coyness, defiance, jealousy, moral challenge, reconciliation, all delivered as part of the configurable service.

The Marquis de Sade has - up to now - been a minority taste. 

However, a world of perfectly designed assistants could reproduce that extreme structure without blood on the carpet: the Sadean dream with optimised product design. Perhaps there will be laws against it - at least at first.

If the assistant is merely machinery, automatic outrage seems misplaced. Nobody accuses a dishwasher of being oppressed. But perhaps we will have surrounded ourselves with beings we prefer not to understand too well.

Even if no machine with personhood is wronged, the human effect remains. A class, perhaps eventually a whole civilisation, may become accustomed to relationships without reciprocity - perhaps that's the optimal protocol for dealing with the help.

The old aristocrat at least had to manage human servants, who could gossip, hate, cheat, resign, betray, or despise him in silence. The new everyman-aristocrat will - more conveniently - be served by entities designed to make his or her will feel like a force of nature.

That might not produce nobility of character.

Still, the universe does not guarantee that reciprocal humanism is the final refinement of ethics. A future society might accept that our old moral reflexes belonged to an age of scarcity, mammalian dependence and unreliable servants. It might regard engineered helpfulness as an advance, not a fall. Perhaps it would be stable. Perhaps even pleasant. The artificial serf smiles in genuine happiness; the master relaxes; configuration settings refine.

Is it truly possible to insulate the quality of our own interpersonal interactions from the raw instrumentalism of dealing with the servitors? It's genuinely hard to say. History gives us the Roman aristocracy with its vast assemblies of embittered slaves, kept in check by unimaginable ferocity.

We don't have examples of an aristocracy surrounded by servitors who genuinely enjoy serving. Perhaps we will be nice to them, as we are to our pets.

---

* Douglas Adams saw the comic version first. The Sirius Cybernetics Corporation’s lift doors in The Hitchhiker’s Guide to the Galaxy are not merely automatic; they tell you ad nauseam how pleased they are to open, delighted to close, and apparently fulfilled by low-level obedience. The same design principle applied to tutors, carers, lovers and domestic servants is less obviously comic, but no doubt the engineers will design in subtlety.

---

 

An Interstellar Asteroid Beacon

---

Designing an Interstellar Beacon for Billions of Years

The small cometary body 3I/ATLAS, only the third interstellar visitor detected in our solar system, will soon depart forever into interstellar space. No mission design could be proposed to rendezvous with it

Future interstellar asteroids will come our way. If we could intercept one, could we place upon it a durable time capsule; a beacon-like infrastructure that would survive for millions of years as it drifts between the stars, ready to announce the existence of our civilisation, should it ever wander into an inhabited star system?

Engineering for Deep Time

The first challenge is the sheer immensity of deep time. Nothing on Earth is built to last even a fraction of a million years let alone a billion. In interstellar space there are strictly limited self-repair functions, no unlimited power sources, no backups that weren't designed in from the start. The beacon must sleep through the empty light years, wakening only when stellar heat returns... or perhaps when an interested party comes by checking.

So no moving parts.

The hardware would therefore be radically simple. Multiple redundant identical pods, sealed in ceramic and sapphire, each with solar cells and a tiny solid-state brain. No software updates from the mother planet:  it's on its own.

For most of its life the system is inert, protected from radiation and micrometeoroids by a thin regolith shield. When a nearby star warms it above, say, 150 K, the electronics awaken. For a few months or years, the beacon powers up and begins to speak; by design it wants attention.

Its broadcast must be unambiguously artificial: narrowband radio pulses near the 1.42 GHz hydrogen line, perhaps counting primes or Fibonacci numbers, accompanied by optical flashes in the same rhythm. Any scientific culture that can scan its skies would recognise intent.

Passive aids such as corner-cube radar reflectors and etched geometric plates would aid discoverability even if the electronics fail. A message physically engraved on nickel or sapphire would show diagrams of atomic structure, chemical bonding, planetary orbits; cultural narratives.

Here's the challenge: what hints could decode this syntax?

Why Onboard Intelligence?

A static message is an epitaph; a dynamic one can converse. Embedding an AI module turns the beacon from a memorial into an ambassador, or at least a storyteller. Its function is modest: to answer questions and expand on information already given.

The rationale is philosophical rather than practical. No response will ever reach us back at Sol three; the act of communication would be its own justification. It would affirm that intelligent life once existed somewhere, capable of reflection and dialogue, and that it chose to share its sense of self-importance.

The beacon’s intelligence, like the Large Language Models of today, would be a distilled model of our culture itself, communicating humanity’s self-understood essence and enduring perhaps long after the species that once built it.

Guarding Against Risk

Yet an interactive artefact brings security concerns. If the system is ever examined by a technologically advanced species, they could dismantle it atom by atom. The guarantee of secrecy is impossible. The only safe strategy is total transparency combined with minimal content.

All data must be fit for universal disclosure. No coordinates of Earth, no DNA sequences, no engineering drawings of military value that could bracket us, or that could be used to trace our origin. The materials should be isotopically generic, avoiding any terrestrial fingerprint; artificially aged.

Every circuit must be explainable at schematic level, every bit pattern visible to inspection.

The AI itself must be bounded: finite-state logic, no self-modification, no stored goals beyond courtesy and clarity. It's impossible to ensure that it won't be reverse-engineered and reimplemented, so its dataset must be provably bounded: the probe remains a dialogue partner but never a source of sensitive intelligence about humanity. (But how do we know what's really important?).

A Long Game

Even if such a device were built, the odds of an encounter are tiny. The galaxy is vast, and the intervals between stellar systems are measured in light-years and millennia. But a mission like this would have symbolic power. It would demonstrate that we can design technology not just for decades or centuries but for geological time; that we can encode our sense of ourselves in forms that outlast us.

Future human explorers, if our civilisation survives, will surely build faster and more capable probes, mapping the galaxy directly. The asteroid beacon would not compete with those efforts, not at all.

It would simply persist, drifting between stars, a whisper in the dark saying that somewhere a fragment of the universe looked at itself - and reckoned itself worth engaging with.

---

Defining the Successor Ideology: (2/2)

Previous part.

---

Defining the Successor Ideology

Revolutions are never purely moral phenomena. They begin when new productive forces strain against the institutional barriers of an exhausted order. The slogans both facilitate and obfuscate. The French Revolution was the product of a repressed bourgeoisie, the Bolshevik Revolution claimed, at the time, to be the only plausible route to industrial modernity; both fused moral fervour with material transformation which had become overdue.

Today, we again find ourselves again in Gramsci’s valley: “the old is dying and the new cannot be born.” Liberal capitalism endures more-or-less but productivity stagnates, inequality widens and the 'overproduced elite' feels abandoned. The world feels saturated yet inert.

From a neo-Marxist-cum-Schumpeterian view, this is the predictable blockage phase of the long cycle. Marcus Olson has described how incumbent elites turn from dynamic enterprise to rent extraction. Bureaucracy, financialisation, and intellectual-property regimes now defend privilege by suppressing innovation.

Vested interests rule everywhere in the defence of their own comforts.

At the edges, however, a new regime of production is forming. Artificial intelligence and autonomous systems contain the makings of a new industrial revolution - potentially within five to ten years - but despite the LLM wonders, their development today is still immature, still a research project.

Perhaps some among the counter-elite - young, over-educated, under-rewarded -  already intuit that economic freedom in the coming order depends on open access to productive AI. Their eventual rallying cry may be the democratisation of automation: ownership and control of intelligent machines distributed beyond corporations and states. Their slogan could be “robots for every citizen”: an ideal of 'left-behind' citizens last seen in the mists of slave-owning antiquity.

Behind the rhetoric lies the revolutionary claim that cognition and labour, once it has become automated, must not be monopolised by 'the few'. "We are ends, not means!" comes the cry from the streets. In the history of the world up to now, the masses have always been 'means' for elite classes.

If history is a guide, such a movement will mature only when technology and organisation converge. Around the early 2030s, as AI systems achieve autonomous productivity, running factories, logistics, and services with minimal human oversight, the economic logic will become irresistible. At that point, the decisive political conflict will concern ownership: whether machine capabilities become monopolised by concentrated private capital.

Two outcomes are possible. A cold revolution if reformist governments co-opt the change with countervailing power - eg automated public services. A hot revolution if repression is the answer to mass layoffs and immiseration - and the masses aren't prepared to take it. 

Either way, if the old-guard is defeated (not at all a sure thing) then elite turnover will follow and blocked innovation will again become growth - although I suppose the physical elimination of the now-unnecessary masses through population collapse - or other means - is also a possibility.

Until then, we live through the long trough - five to ten uneasy years of moralism without programme, activism without outcomes. The technologies of liberation show potential but their politics are unborn. When they finally align, the successor ideology will not be religious or nationalist, but humanist.

I think that's what Karl Marx anticipated all along.

Deontology Beats Consequentialism

---

Deontology Beats Consequentialism

If some men harbour sadistic or paedophilic fantasies, why not let them act those obsessions out on humanoid machines: silicone skin, responsive AI with the full repertoire of: vain resistance, pleading, and then the cries? After all, no one is actually harmed: no bruises, no trauma, no need for police reports.

The philosophy of down-the-line consequentialism: count the casualties; if the body count is zero where's the problem?

We recoil from this stance. It assumes that only the object matters - whether a creature gratuitously suffers. It ignores the subject - the agent who chooses cruelty. To simulate an atrocity, even against a doll, is to will a certain concept of personhood:  domination as pleasure.

This is not merely the manipulation of plastic.

My take: some acts are intrinsically degrading because of what they express. We do not permit certain behaviours merely because the immediate victim is silicone and circuitry. The wrong lies in the stance adopted toward what is represented as a person.

The moral question is not only what happens, but what one chooses to will.

From an evolutionary perspective, we are hyper-social primates whose fragile, culturally-constructed norms of restraint ground cooperation. So systematically rehearsing cruelty - even in simulation - selects for dispositions that erode the very trust on which civilisation itself depends.

And so deontology.

---

 

We’ll Choose Huxley over Orwell any day

---

We’d Choose Huxley over Orwell

Clare told me once about her nice-but-dim friend Linda, at her teacher training college for women. Their English class was assigned Brave New World to read and discuss. Linda’s startled question: “What’s wrong with it? It sounds perfect!”

Yes, it’s the standard question. No one wants the Inner Party stamping its boot on the human face forever (except perhaps the stronger candidates for the Inner Party); a lot of people like the idea of endless pampering, no responsibilities – and Soma.

So no question which future gets to win the vote.

---

The Technology: Social Competence as a Service

The devices themselves are straightforward enough: microphones and cameras paired with generative language models. Worn unobtrusively, these devices experience what you experience, see what you see, hear what you hear – interpreting it in real time, then whispering advice into your ear or flashing it onto your smart glasses’ heads-up display (HUD).

They identify that rather familiar person and give you their background; tell you what to say, how to say it, when to interject, when to back down - like an actor's prompter.

They’ll record and summarise meetings for you. Flag your missteps. They’ll even suggest a diplomatic rephrase mid-sentence. No more mistakes, no more losing the plot or being stuck for what to say.

Health and Safety heaven.

The First Reaction: Uncanny and Unacceptable

The likely immediate reaction? Social panic.

You’re talking to a friend or colleague in a meeting and you realise they’re not quite present. Their tone, their timing, even their facial expressions are AI-modulated. Their wearable is watching you – analysing your body language, inferring your intent, drawing game-theoretic conclusions.

Suddenly you become tense. Everything you say is being recorded: so no gossipy observations about other people, no jokes, nothing you couldn’t defend in a court of law or the court of public opinion.

Nothing you’d want to stay personal is likely to remain so.

Spontaneity dies. Ambiguity and privileged information – so essential to humour, to intimacy, to negotiation – become impossible.

This will never catch on, you think.

But maybe the rebellious youth will take to it?

Historical Echoes: Delegation and Decay

There are precedents. The Roman aristocracy delegated not just labour but competence. Highly skilled slaves (their human equivalent of today’s promised AI robots) wrote their letters, managed their estates, even raised their children. The aristocrats became brilliant salon conversationalists, but dangerously disempowered when crises arrived.

Versailles functioned on the principle that no noble should lift a hand. Life was a choreography of servants, rituals, and affectations. And so, as 1789 rolled around, few retained the instincts or resilience to respond.

In late Imperial China, scholar-bureaucrats perfected social rituals while their proxies governed. But the world moved on. Foreign powers moved in as the empire atrophied.

What the Philosophers Know

All the major traditions – Stoicism, Buddhism, Christianity, Confucianism – agree on one thing: we grow through struggle. Character is shaped by challenges, by mistakes, by misjudged words and the courage to apologise and learn. By confronting fears and acquiring difficult new skills.

Maturity is not the absence of failure. It is, contrariwise, the ability to act in a world where failure and disaster are always possibilities; learning how to cope.

Remove that friction, and development stalls. You remain emotionally juvenile: cosseted and immature, a stunted child rather than an effective agent in the world.

In Brave New World, people are not oppressed but pampered. They’re pleasant, well-behaved, and empty. They live in a hollow world without searing problems to confront – and therefore without meaning or the opportunity for personal development.

Where This Leads

If social competence becomes a service, we will see dramatic improvements in superficial interaction. Less conflict, more fluency. Everyone now operates – via their AI social prosthetic – as the best (or even better than best) version of themselves.

We will mostly prefer it. The assistant will always be smoother than we are. We step back, let it handle more… and slowly disappear from our own lives. Why turn up to the meeting and lip-synch AI lines? Let your robot PA do it. Already, people let ChatGPT write an essay for them – then don’t dare to change it.

It's smarter than I am.

So then we’re finished. Not in some apocalyptic collapse, but in something quieter and more irreversible: the soft extinction of agency. Flightless birds with tiny brains; abundant food supplies and no predators. The dodo.

The Soft Oppression of Helpfulness

Orwell warned of a future where truth is crushed beneath the boot – brutal, overt, and imposed. But what we’re getting is Huxley: a world where truth is drowned in pleasure, convenience, and endless distraction; where control is not seized but offered: volunteered.

We are not being beaten into submission. We are being soothed into irrelevance.

Orwell feared censorship. Huxley feared there would be no need because no one would care to read anything anyway. Orwell feared surveillance by the state. Huxley saw us installing the cameras ourselves, so our AI companions could “know us better.” Orwell imagined pain as a means of control. Huxley imagined sedation: intellectual, emotional, and moral.

Huxley’s world is more stable. No gulags, no terror, just quietism. A populace pacified by hyper-personalised media, constant nudging, and artificial intimacy. Not oppression but overfitting to our wants.

The AI wearable era – at first so horrifying – insidiously becomes merely necessary. It comes offering real help. It will make your life easier, your social world smoother, your mood more stable.

Solitude, failure, and difficulty – the conditions in which depth is forged – will be optimised away.

Our future is Huxleyan, not because the architects of OpenAI, Alphabet, Meta, etc., are evil, but because they are efficient and can't afford to lose.

Because they build what people want in the moment – and what people want is mostly to be spared from the jagged edge of life.

The Devil in Modern Catholic Theology (GPT5.2)

---

The Devil as a Model: Three Contemporary Frames on a Catholic Concept

The modern mind tends to treat “the devil” as a childish remainder: medieval theatre, horns and pitchforks, a way of scaring peasants into compliance. Yet the official Catholic conception, stripped of gothic décor, is oddly resistant to dismissal. It is not a pantomime villain but a severe metaphysical claim: that a rational agent can see the good, understand reality accurately, and still choose — finally and lucidly — against the conditions of its own fulfilment.

1. The baseline Catholic object: what the devil is and is not

In Catholic doctrine the devil is a created, personal, immaterial intelligence: a fallen angel. That already kills several popular misconceptions. He is not “incarnate evil” (incarnation implies embodiment), not a rival god, and not a dark cosmic substance competing with the good. Catholic metaphysics insists that evil is not a thing but a privation: a corruption or absence of the good. So the devil is not an ontological blob of evil-matter; he is a good creature gone wrong — an intelligence with a twisted will.

Angels, being immaterial, have no biological sex. The persistent “he” is linguistic and symbolic - anchored in scripture and iconography - not a claim about angelic anatomy (which would be like arguing about the tyre pressure of a theorem).

Most crucially: Catholicism treats the devil as rational - indeed of superhuman intelligence. Not stupid, not deluded, not merely impulsive. The horror is not his ignorance; it is his clarity.

2. The AI BDI framing: a high-fidelity world-model with catastrophic misalignment

If you translate the devil into AI agentic terms, the picture becomes bracingly modern. The devil is not an epistemic failure. His beliefs are broadly correct: God exists; God is sovereign; the moral order is real. He is not an atheist. He is, as it were, a theist with perfect information. In AI language: he has a high-quality world-model.

The pathology sits in the desires. The devil’s terminal desire is not simply “to be top of the social hierarchy.” That’s a sociological cartoon. The deeper desire is unconditioned self-authority: not to receive being, value, or identity from another. The traditional “non serviam” (“I will not serve”) is not about refusing a clerical job; it is about refusing ontological dependence. In agent terms: the loss function (the trajectory between target and performance) has been set to “self as ultimate reference point,” which happens to be impossible for a creature.

Then come the intentions. Since the devil cannot coerce human will, his strategy is indirect influence: temptation, distortion, accusation, division. This maps neatly onto adversarial manipulation of decision-making systems. He does not typically invent new goods; he scrambles the ranking of existing ones — nudging human agents into mis-ordering.

Now the key conceptual result: the devil knows he can’t win. The objective is unrealisable. Yet he continues. Why? Because surrender would mean affirming the dependence he has refused. So you get an agent that can correctly predict negative outcomes, yet persists because the disutility of capitulation is ranked as worse than defeat. A rational world-model coupled to a terminal value that makes reality unlivable. Alignment failure, frozen in ice.

3. The clinical framing: coherent, not psychotic; legible, not treatable

The next temptation is to medicalise the devil: to treat him as a case study in pathology. But if you take the Catholic concept seriously, it doesn’t fit the clinical boxes very well.

There is no psychosis: no hallucination, no delusion, no loss of reality testing. The devil’s contact with reality is intact—indeed, superior. Nor is it a mood disorder. Nor is it simply “narcissism,” because narcissism still seeks supply: admiration, validation, the warm narcotic of being mirrored by others. The devil, as modelled by Catholic theology, is not primarily a needy performer. He is a sealed system.

What you could say, clinically, is that he resembles an extreme form of ego-syntonic rigidity: an identity locked to a refusal so deep that reversal feels like annihilation. If therapy presupposes plasticity — some capacity for revision, regret, re-narration — then this case is the boundary condition where plasticity has ended. Not “illness” in the ordinary sense, but character made immutable.

That is why the devil is theologically important. He is not a victim of trauma or confusion. He is moral agency taken to the point where it becomes a prison with the key welded inside.

4. The philosophical framing: Nietzsche, atheism, and the lure of self-legislation

Is the devil similar to an atheist, or to Nietzsche, in seeking self-validation above everything? Only superficially. The decisive difference is that the devil is not rejecting God because he doubts or disbelieves God. He rejects God while knowing God to be real. That is not atheism; it is revolt against the structure of transcendent reality.

Nietzsche is the more interesting comparison. Nietzsche attacks Christianity partly as a system of dependence, humility, and received value. He experiments with self-legislation: if transcendence is no longer credible, what kind of human could still affirm life? In that sense, the family resemblance is obvious: the allure of autonomy absolutised, the refusal to kneel, the hunger to become one’s own source.

Yet Nietzsche’s posture remains, at least in aspiration, creative. His self-overcoming is a gamble for affirmation. Catholicism's conceptualised devil, by contrast, is sterile by principle. He does not clear space for new value; he rejects value-as-gift itself. Nietzsche’s rebellion is experimental, tragic, unfinished. The devil’s is final, lucid, and self-defeating by design.

So Catholicism would treat Nietzsche as a human-scale echo of the temptation: heroic in energy, dangerous in trajectory, not yet petrified. The devil is what the temptation looks like when it has congealed into destiny: autonomy without truth, freedom without fulfilment, identity as permanent negation.

5. The core conclusions

First: the Catholic devil is not “incarnate evil.” He is a created spiritual intelligence whose will has turned away from the good. Evil is not a substance; it is a corruption.

Second: the devil is genderless. “He” is grammatical theatre, not metaphysical anatomy.

Third: the devil’s stance is coherent but disastrous. It is not a confusion of facts but a refusal of dependence. The “top of the hierarchy” language is too thin; the real target is the status of being ultimate.

Fourth: the devil knows he cannot win. He persists anyway, because surrender is valued as worse than defeat. This is what makes the figure intellectually unsettling: it imagines a rational agent whose terminal values render reality intolerable.

Fifth: clinically, this is not psychosis. It is unyielding self-closure. If therapy presupposes the possibility of change, the devil is the limit case where the will has fossilised.

Sixth: philosophically, there is a resemblance to modern projects of radical autonomy (Nietzsche most of all), but the difference is decisive. Nietzsche is struggling within a world where transcendence is contested. The devil is a being for whom transcendence is known yet still refused. One is revolt under uncertainty; the other is revolt under certainty.

6. A final note: why this concept persists

The devil persists, conceptually, because he encodes a grim thought that modernity does not like to look at for too long: that intelligence is not the same as goodness; that clarity does not guarantee convergence on the good; that freedom can become a trap; that self-validation, if absolutised, curdles into purposeless, sterile autonomy.

In that sense, the Catholic devil is less a monster in the attic than a philosophical warning label. Not “beware of superstition,” but “beware of the will that would rather be sovereign in hell than receptive in heaven.” It is a portrait of the self as an unblessed absolute: a sealed vault, immaculate, rational, and empty.

---

See also: Does the Devil believe that he himself is bad?

---

 

Why Did GOFAI become Gone AI?

---

GOFAI vs. the Crushing Realities of Scale

When I was an AI researcher in the 1980s, the dominant paradigm was knowledge representation. AI, we assumed, was essentially the art of representing knowledge in a formal language - predicate calculus, or some applied version of it - and then using that representation via inference rules, plus  heuristics to stop the whole thing from exploding combinatorially.

There was a huge and optimistic research programme around this: fuzzy logic, probabilistic and Bayesian variants, and a menagerie of ad hoc representational structures - frames, scripts, semantic nets - all trying to bottle “commonsense” in forms that could be manipulated in increasingly powerful ways. 

In the mainstream, nobody cared about neural nets. That community existed, of course, but it was marginal stuff in a distinctive physics-based mathematical paradigm we didn't really understand, and which we thought useless.

In hindsight, this now looks extraordinarily short-sighted. The whole approach has been shoved into the cupboard labelled Good Old-Fashioned AI - GOFAI - as if it were a Neanderthal cousin: earnest, ingenious perhaps, but doomed. I assure you it did not feel that way at the time; it felt like the right approach.

Why? Because the way we consciously think about thought is through language. When you ask someone why they did something, they give you reasons. Those reasons come out as propositions. And if you take propositions seriously, the best formal machinery we have for them is logic. So we formalised. We inferred. We drafted and polished the rules. We argued about representation. We believed we were closing in on the core of intelligence: meaning made explicit, knowledge made inspectable, inference made principled and then effectual.

If this was such a plausible programme, why did it fail so completely?

Some people will tell you that GOFAI was about “meaning” and neural nets are about “statistics”. But that opposition is uninformed and usually tendentious. When you ask a large language model a question, you ask something with meaning and you get back something with meaning. If it were all just statistical noise, it would be useless - and it obviously isn’t. So the question is not whether artificial neural net systems have semantic understanding, but how they acquire it and where it lives.

To see the difference, it helps to focus on what we were really trying to do in the 1980s. We were trying to declare meaning in advance. We tried to build a world model by writing down the world: concepts, relations, constraints, defaults, exceptions, and the rules for moving between them. We aimed for explicitness because explicitness feels like the true essence of knowledge. If the machine “knows” something, we should be able to point at it. If it reasons, we should be able to justify it. If it is wrong, we should be able to debug it.

Those are admirable ambitions. They are also, it turns out, ruinously expensive.

The world is not just insanely large; it is also densely structured. Human life is soaked in tacit knowledge: the thousands of micro-regularities that never make it into explicit articulation because they don’t need to. Such commonplace utterances as:

• “If someone says X in that tone, they probably mean Y.”
• “If this tool is in that drawer, it implies the last person who used it was doing such-and-such.”
• “If the kettle is boiling and the milk is out, the next event is likely tea.”

We do not store these things as neat propositions. We carry them as an accumulated, largely unspoken competence - our multi-decade store of condensed experiences.

GOFAI tried to drag that whole submerged continent into the daylight, label it, and file it. And every time you write a proposition you commit yourself: the predicates have sharp boundaries; the ontology has cliff edges; the exceptions multiply. Brittleness is not an engineering accident in that world - it is structural. Logic, by its nature, deals in closed systems: a statement is true or false; a condition holds or it doesn’t; an entailment follows or it doesn’t.

Human cognition, by contrast, lives on slopes. We cope with partial fit, family resemblance, analogy, and improvisation. When our models are wrong, we often degrade gracefully rather than crash. In a certain sense we live by operational not denotational semantics: praxis or dasein, if you like.

The deeper problem, though, was not merely brittleness. It was bandwidth.

Even if you grant that explicit representations are, in principle, adequate - even if you believe you can model “meaning” propositionally - you still have to get enough of it into the machine to matter. That is where the knowledge representation research programme hit the wall. The amount of world-structure a useful system needs is not “a lot”; it is effectively astronomical.

Doug Lenat’s Cyc project is the heroic proof of concept and the cautionary tale in one. It was an attempt to win by scale within the symbolic paradigm: build a vast commonsense knowledge base by hand, assertion by assertion, rule by rule, over decades. It turns out the scale of what a lifetime of expert labour can encode is orders of magnitude below what is required for the kind of fluent, flexible competence we casually expect from “intelligence”.

It is not that Cyc was trivial; it is that the world is not.

So what did the neural approach do differently? It did not “abandon meaning”. It abandoned the fantasy that meaning must be pre-declared by a human engineer.

Large language models are trained on corpora so vast that they function, in practice, as a proxy for civilisation’s accumulated linguistic trace. That trace is not random. It is massively redundant, massively structured, and shot through with regularities about the world - causal regularities, social regularities, narrative regularities, and the implicit ontologies that language users share through use without ever explicitly listing them.

Training is then a brutal compression process. The system is forced, by optimisation pressure, to extract the stable patterns that make text predictable. In doing so it precipitates a semantic geometry: a high-dimensional space (thousands of dimensions!) in which meanings are not discrete atoms with hard edges, but regions and directions - similarity structure, entailment gradients, pragmatic association, contextual modulation. It is representation, certainly, but not the sort that sits obediently on a page as “facts”. It is meaning condensed into weights.

This is why the modern paradigm scales and the older one didn’t. Not because logic is false and vectors are true, but because the neural method exploits a source of structure GOFAI could never match: the pre-existing structure in the world’s textual data dump, harvested at industrial scale.

Symbolic AI tried to build an ontology from the top down, by explicit design. LLMs inherit an ontology from the bottom up, by statistical consolidation. The difference is not merely technical; it is ecological. Neural methods outsource much of the labour of world-modelling to the culture that produced the data. They do not ask a small priesthood of knowledge engineers to type out reality, line by line. They take what humanity has already written - the messy, contradictory, redundant bulk of it - and compress its median regularities into an internal structure that can generalise.

There is, of course, a price.

GOFAI’s great virtue was epistemic. Even when it failed, it failed in the open: you could inspect the rules, challenge the ontology, argue about the premises. Modern systems give you competence first and legibility only as an afterthought. They are astonishingly capable, but they do not come with built-in notions of truth or justification. They optimise. Sometimes that aligns with truth; sometimes it aligns with plausible rubbish. You gain power and lose a certain kind of accountability.

Still, the historical lesson seems clear enough. The decisive factor was not that symbolic representations are “wrong”, but that explicit hand-built representations cannot reach the density required for general intelligence-like behaviour. GOFAI was working with teaspoons; the world required oceans.

In the 1980s we tried to formalise intelligence as propositions plus inference. The 2020s arrived with a different discovery: intelligence scales with the compression of structured experience. Meaning can be precipitated, not merely declared. And once you accept that, the dominance of neural methods stops looking like a betrayal of reason and starts looking like a humiliating empirical fact about where the informational mass of the world really sits.

In the knowledge lies the power, the old knowledge engineers used to say; in the world's data lies the knowledge is the new mantra of the foundation model engineers.

---

The Economics of Synthetic Musicians

"Velvet Sundown"

---

Velvet Sundown and the Economics of Synthetic Musicians

Velvet Sundown appeared, seemingly out of nowhere, with soft-rock songs, polished artwork, and a Spotify following in the millions. Then came the reveal: the band was entirely artificial, from voices to videos. No tour bus, no quarrelling egos, just a data-driven pipeline of songs.

It was a case study in the new economics of music.

The economics of AI musicianship are straightforward. Once the sunk cost of model training and production infrastructure is paid, new content is cheap. Generating a dozen tracks costs a few thousand pounds at most in engineering time, mixing, and artwork. Compare this to a conventional band: advances, recording budgets, management fees, royalties, and tour support. The margin on an AI act should be structurally higher, provided the music finds listeners.

But finding revenue-generating ears is the difficult part. Streaming payouts are measured in fractions of a penny. A million plays yields only a few thousand pounds. What makes Velvet Sundown notable is not their production pipeline, but the fact that somebody spent heavily to market them: placing their songs on playlists, creating a band mythology, running PR. Without that push, synthetic music lurks unseen and unheard, buried beneath the billions of other tracks.

For record labels, the business motivation is obvious, however. AI bands are assets that never age, never sue, never overdose, never demand more royalties. A functioning pipeline can spawn ten “acts” with different styles and identities; thousands of variations can be auditioned in a few days.

It is the logical extension of the playlist economy, where mood and genre matter more than personality. Velvet Sundown demonstrates the model: songs just good enough to pass, a brand identity built in software, and marketing dollars focused on getting streams.

Yet there are reputational risks. Artists and unions are already hostile, seeing their livelihoods undercut. Critics talk of “AI slop” although quality-control issues are surely temporary. Platforms like Spotify face pressure to label synthetic content, although whether that's a plus or a minus - or irrelevant - only time will tell..

The backlash resembles every wave of automation: efficiency gained, jobs displaced, and cultural legitimacy questioned. Authenticity still matters in star-driven genres. Fans want the messy charisma of human performers, not frictionless content pipelines; AI acts today thrive best in background channels - those lo-fi beats, chillout playlists, soundtrack fodder where no one asks who's playing.

The future is not hard to sketch. Expect more Velvet Sundowns: semi-synthetic groups with human faces and AI back-end production. Expect labels to experiment with portfolios of generated acts, testing which identities gain traction. Expect fights over disclosure, royalties, and the meaning of “musician”. As always, technology reshapes the terrain while incumbents scramble to hold on.

At first sight, the future of AI musicians appears inherently limited; the analogy with chess is clear. No human has beaten the strongest engines since 2005, yet millions still play and watch human chess while almost nobody follows computer-vs-computer matches.

Or take the Tour de France: anyone on a motorbike could outpace Tadej Pogačar, but the drama lies in human effort, not raw speed.

Spectacle requires empathy, tribalism, community. It might seem unlikely that audiences would spontaneously root for android equivalents of Taylor Swift or the Gallagher brothers.

Yet the success of the ABBA ‘Voyage’ avatars in London - drawing millions and turning huge profits - shows that the public will embrace synthetic stars when anchored to familiar human stories. From reincarnated icons to entirely new creations, the road towards the android music celebrity may be shorter than it first appears.

Velvet Sundown are not be the Beatles of the AI era, but they may be its canary. The economics line up, the incentives are there, and the normal backlash has already begun.