There is a very precise threshold beyond which artificial intelligence ceases to be merely a system that advises action and becomes an intelligence distributed in matter.

Until now, we could still imagine AI as a brain installed in a data center, analyzing flows, proposing scenarios, accelerating a human decision.

But when this intelligence directly controls machines capable of perceiving, moving, cooperating, and acting in the physical world, a much deeper transformation begins.

The algorithm leaves the screen. It takes on a physical form: engine, camera, rotor, vehicle, land robot, maritime system.

Intelligence becomes kinetic.

This embodiment directly alters the question of autonomy.

An error in a conversational system produces an incorrect sentence. An error in a mobile machine can cause movement, a collision, or destruction.

The software acquires hardware inertia.

Code becomes movement.

The change is no longer just informational: the results of the calculation directly control movements in matter.

I have always considered this relationship between matter and information to be one of the fundamental characteristics of our time. A database may seem abstract, but it determines real decisions. A digital architecture may seem invisible, but it governs economic flows.

With robotics, this relationship appears immediately.

Information becomes action. The model recognizes an obstacle, modifies a trajectory, slows down, accelerates.

It acts.

And as soon as it acts, the entire question of responsibility changes scale.

But the real breakthrough does not lie in the existence of an isolated, autonomous machine. An isolated machine remains relatively simple to conceptualize. We can ask who controls it, what mission it has, and what limits are imposed upon it.

The mutation begins when several machines cooperate.

Ten drones.

A Hundred.

A Thousand.

Several tens of thousands.

Each unit has a local perception, exchanges certain information with others, and adapts their behavior according to the state of the group.

We then enter into the logic of the swarm.

The word is correct because it comes from life.

A single ant has limited capabilities. However, a colony produces complex behaviors without a central ant controlling every movement.

Birds can collectively change direction without a leader individually dictating each bird's trajectory.

Global behavior emerges from local rules.

What interests me here is not attributing an inner life to the swarm, but observing that effective coordination can exist without a central command detailing every movement. This alone is enough to overturn our conception of responsibility.

This is what algorithmic engineering seeks to reproduce. And this is precisely what disrupts the classical notion of command.

In a traditional army, the hierarchy descends from the command to the units. Each level transmits orders and receives information.

The swarm introduces another architecture: part of the coordination can emerge horizontally.

The human being no longer necessarily tells each machine: go here, then there.

This transformation is profound.

They can tell the system: explore this area, maintain this formation, monitor this border, optimize this mission under these constraints.

The order becomes more abstract.

And the more abstract the order becomes, the more political importance the system's design acquires. Whoever defines the local rules defines the emergent behavior.

Power can be shifted from the moment of action to the moment of programming.

The programmer indirectly becomes the architect of future collective behavior.

This obviously does not mean that the programmer decides every action. But the programmer defines the space of possible actions.

One idea runs through this entire book: architecture organizes possibilities.

A road does not decide where each vehicle goes, but it makes some journeys possible and others difficult. A computer protocol does not choose the content of messages, but it defines how they travel.

Similarly, the rules of a swarm define its behavioral field.

This is why the governance of autonomous systems will need to move much earlier in the technical chain.

The inspection cannot begin only when the machine is already in the field.

It must be integrated into the design.

What actions are prohibited? What areas are excluded?

What level of uncertainty necessitates the shutdown?

What should be done if communications break down? What should be done if two objectives conflict?

All of these questions belong in the code.

Ethics thus literally enters into functional architecture.

It can no longer remain an external statement.

This ties in exactly with what I wrote in the previous chapter: de-escalation by design.

But with swarms, this requirement is even more compelling because individual supervision is no longer feasible.

A human can potentially control a drone.

It cannot pilot ten thousand units simultaneously.

Beyond a certain scale, the human operator has to delegate. And that delegation changes the very meaning of the expression human in the loop.

Where is the human element in the loop when they no longer control every action, but define an overall mission?

Can the human operator understand what the system is doing, modify the mission, stop certain units?

Can it prevent a particular action, and within what timeframe?

The concept of human control without these details becomes almost meaningless. In a company, a manager cannot be held responsible if they have no access to data or operational decisions.

Formal responsibility is not enough. An effective capacity to intervene is necessary.

This distinction between legal responsibility and operational capability will be central to autonomous systems. But we need to look even further.

The swarm possesses a property that centralized architectures lack: resilience through dispersion. If one unit is destroyed, the others can continue. If central communication fails, certain functions can be maintained locally.

This is an extraordinary quality in a military environment where networks can be jammed or attacked. The swarm can therefore function as a form of physical distributed network. It is exactly the logic of the internet transposed into matter.

A resilient network does not depend on a single path. Packets find alternative routes. In a swarm, agents can adjust their distribution when some disappear.

This similarity strikes me. We have spent decades building distributed digital architectures. We are now physically embodying them.

But, as in any distributed network, this resilience has a reverse side: central control proves more difficult.

The same mechanism that protects it against failure complicates the shutdown.

That is the paradox.

We want systems that can continue operating when communications are interrupted.

But we also want to be able to stop them.

These two requirements can come into conflict.

The more local autonomy a system has, the less it depends on the central authority. The less it depends on the central authority, the less readily the central authority can take control.

This is a critical engineering issue:

How can we design a degraded autonomy that allows us to survive without becoming uncontrollable?

It will probably be necessary to define fallback behaviors.

In the event of a loss of communication, some actions are permitted, others are not. A machine can return, go into standby mode, or continue a limited task.

This logic already exists in many systems.

But as embedded intelligence increases, the possibilities grow more complex. The machine can interpret more situations.

This flexibility is tactically valuable, but it increases the room for uncertainty.

We therefore return to a fundamental rule: autonomy must be proportional to risk and reversibility.

A swarm tasked with mapping an area can receive much more autonomy than a system with a lethal capability.

This is a crucial distinction.

Military robotics should never be treated as a single, monolithic entity.

Reconnaissance, logistics, demining, evacuation, electronic warfare, jamming, transport, attack: the levels of risk are radically different.

This question is obviously central, but it masks a multitude of uses that will also transform warfare.

Logistics, for example.

The public debate often makes the mistake of collapsing all robotics into the image of the autonomous killer weapon.

The strategic effect can be immense without the machine ever deciding to fire.

Offensive capability and systemic transformation must always be distinguished.

A logistical innovation can change the outcome of a conflict as much as a new weapon. Military history constantly demonstrates this.

Armies are not merely systems of destruction.

They are gigantic flow systems.

Here again, we find the database and the network.

Whoever knows where their resources are, what condition they are in, and which route is available holds a considerable advantage.

AI and robotics can automate part of this traffic.

This creates a more distributed, faster army, but also one that is more dependent on digital systems.

Every advantage produces its own vulnerability.

The more automated logistics become, the more effective a cyber or electronic attack can become.

It will be an environment where each layer — sensors, logistics, communication, targeting, navigation — will depend more on algorithms.

The algorithmic battlefield will therefore not be a spectacular war of robots against robots. It will be an environment in which every layer— sensors, logistics, communication, targeting, navigation—depends increasingly on algorithms. War becomes a system of systems.

The visible robot is merely the mechanical tip of an invisible architecture.

Behind it exist satellites, networks, models, databases, operators, cloud providers, industrial chains.

Destroying the robot may be far less effective than disrupting the infrastructure that powers it.

The real objective may be the data.

This is a profound transformation of the battlefield.

The material itself is obviously still crucial. But the ability to coordinate that material may prove even more important.

An army composed of highly sophisticated machines but incapable of exchanging data may be less effective than an army with simpler but perfectly coordinated equipment.

The advantage comes from the network.

This directly relates to what I understood with Groupe Serveur and Artprice: value never resides solely in the object, but in the relationship.

An isolated piece of data has a use; millions of data points linked according to a coherent architecture produce intelligence.

Similarly, an isolated drone possesses certain capabilities; a coordinated swarm produces behavior that surpasses each individual unit.

The graph becomes power.

This idea seems so central to me that we could almost define the algorithmic battlefield as a dynamic graph of sensors and agents.

Each node sees a part of reality.

None of them possesses everything.

The quality of the system depends on how these local perceptions are merged.

The problem is epistemological:

How can we construct a global representation from incomplete points of view?

It is the same problem as a society, an artificial intelligence, or a historian.

And the errors can come from the merger.

If ten sensors observe the same phenomenon, do we have ten independent pieces of evidence or ten measurements from the same source?

Apparent redundancy can create a false sense of trust.

A network of agents must therefore maintain the genealogy of the information:

Who saw what, when, with which sensor, and with what degree of confidence?

This requirement for provenance echoes an old obsession of the Art Market: the value of information also depends on the chain that allows it to be situated.

But here the stakes are immediate.

Automatic fusion can smooth out disagreements between sensors and precisely mask the uncertainty that should remain visible.

Intelligence should not always smooth things over. Sometimes, it must preserve disagreement.

If three drones see a vehicle and two do not, this discrepancy may have an explanation: angle, jamming, camouflage, malfunction.

Merging data simply yields "probability 72 percent" and can obscure useful information.

That is why models should be able to preserve the structure of doubt.

The discrepancy between sensors must remain visible.

We keep coming back to this point: uncertainty is a given.

It should not be eliminated to produce a more comfortable interface.

It is almost a general critique of our time. We want clean, immediate, certain answers.

But reality is often fragmented.

A good intelligence should restore this fragmentation when it is significant.

Art accomplishes this shift when it refuses to reduce reality to a single representation.

The Abode of Chaos functions as an artistic multi-sensor system. Thousands of signs, events, and portraits produce perspectives that are sometimes contradictory.

It does not seek to fuse them into a single smooth narrative.

It retains the collision.

This is precisely what military cognitive architectures should sometimes do: not hide the fact that multiple interpretations exist.

The swarm itself can become a way of producing this plurality.

Each agent sees things from a different point of view.

Spatial diversity becomes cognitive diversity.

But this plurality is only useful if the system knows how to exploit it.

Otherwise, the group may develop herd behavior.

It is a fascinating danger.

Biological swarms are powerful because they follow common rules. But these rules can also produce erroneous collective behaviors.

A group can converge towards a bad solution if the agents excessively copy their neighbors.

This is the problem of the information cascade.

It exists in financial markets, social networks, and human societies.

It will exist in artificial swarms.

If each drone changes its decision based on the others, a local error can become a collective error.

The network then amplifies the fault.

That is why the partial independence of agents is important.

Here again: diversity and redundancy.

An overly coherent architecture can become fragile.

Sometimes it is necessary to introduce algorithmic dissent.

Some agents could be designed to verify rather than follow; some models, to use other methods of perception; part of the system, to test the dominant hypothesis.

We are back to the adversarial procedure.

I propose the hypothesis that this artificial pluralism could become important in autonomous architectures.

Security will not come solely from stricter rules, but perhaps also from the creation of agents capable of challenging the group.

Nature itself often makes use of this diversity.

A genetically diverse population is more resistant to certain threats. An intellectually diverse organization detects more errors.

The future of robotics could therefore paradoxically require a form of artificial pluralism.

It is a profoundly political idea.

We have often imagined the machine as a symbol of perfect conformity.

But the most robust systems might be those that intentionally contain multiple logics.

It is almost the opposite of the totalitarian fantasy of the perfectly obedient machine.

Resilience requires a certain capacity for divergence.

Obviously, this divergence must be controlled.

We do not want each agent to freely invent their own mission.

But we may want it to produce an alternative hypothesis.

Here again, we must distinguish between autonomy of analysis and autonomy of action.

This is one of the fundamental distinctions of the future.

One can give a system a great deal of cognitive autonomy without giving it an equivalent lethal autonomy.

It can explore a thousand scenarios, identify anomalies, propose solutions, while remaining subject to strong limitations for certain actions.

This decoupling seems essential to me.

We tend to imagine autonomy as a single slider ranging from zero to one hundred.

In reality, it is multidimensional: perception, interpretation, navigation, coordination, planning, engagement.

Each dimension can be adjusted separately.

This autonomy matrix will allow for much finer regulation.

Simply saying that a system is autonomous is not enough.

We need to know where and how.

This level of detail will be essential to international law.

Negotiations on autonomous weapons systems will be impossible if states use the same word to talk about radically different realities.

An effective standard must target a specific function.

For example, prohibiting or strictly regulating certain types of target selection and engagement without significant human oversight.

But this expression of significant human control will itself need to be defined.

Significant according to what criteria?

How much time is available?

Understanding the situation?

The possibility of interruption?

We keep encountering the same questions.

The law will have to become almost technical.

The lawyer can no longer simply manipulate general concepts; the lawyer must understand the architecture of the system being regulated.

That seems perfectly natural to me.

Law has never been truly separate from technology. Major industrial shifts have always produced new legal categories: the railway, photography, radio, the Internet.

AI and robotics will do the same thing in the World Model.

The difficulty lies in the speed.

The development cycle is much faster than that of legislation. A capability can emerge, be deployed, and be modified before an international treaty is negotiated.

This creates a risk of permanent institutional delays.

Perhaps we will need to invent more adaptive mechanisms: stable principles and evolving technical protocols.

The ever-changing permanence.

I encounter this concept everywhere because it perfectly describes the contemporary need:

a core of strong rules and a capacity for procedural adaptation.

In the field of autonomous systems, the core could be human responsibility, distinction, proportionality, the possibility of controlling certain critical decisions.

The technical details will evolve.

This would avoid drafting a law tied to a particular technology that would quickly become obsolete.

We need to regulate functions and risks, not product names.

This distinction is crucial.

This also applies to civilian AI.

The political system must learn to think in functional categories.

But the military adds another layer of complexity: the adversary.

Any security architecture must function in a hostile environment.

A civilian autonomous car can assume that other vehicles generally follow certain rules.

A military drone must assume that someone is actively trying to deceive, jam, or divert it.

This difference is immense.

Military AI operates under constant adversity: some of the signals it receives may be deliberately fabricated to deceive it.

Every sensor can be attacked.

Every connection can be compromised.

This requires an epistemology of suspicion.

But this suspicion must be controlled, otherwise the system becomes paranoid.

Paranoia is an intelligence that finds coherence everywhere.

Every sign becomes intentional.

A military system could fall precisely into this algorithmic trap.

If any unusual signal is interpreted as an adversary attack, automation accelerates the escalation.

It is therefore necessary to train the models to distinguish what constitutes adversarial manipulation from what is merely an anomaly.

It is extremely difficult.

A bird can be a bird.

A jamming noise can be a sign of a malfunction.

An unusual journey can be logistical.

Reality naturally produces noise.

The adversary is hiding in this noise.

The task is to detect the intention without inventing it.

This is probably one of the most difficult cognitive problems that exist.

Humans themselves do not always succeed at it.

AI can help by comparing millions of situations, but it must retain this capacity for doubt.

This is where the doctrine of the Blind Spot, the blind spot, takes on its full strategic dimension.

A system should be trained to actively search for non-hostile explanations for certain signals.

Not out of naivety.

To reduce the risk of error.

We could almost imagine an architecture where each threat detection automatically triggers a second process:

What other causes could produce the same signal?

This simple mechanism could become a major security innovation.

It would not necessarily slow the system down much thanks to the computing power.

Here again is this reversal: the speed of calculation can be used to produce more caution.

The real advantage would then not be to shoot faster, but to understand more at the same time.

This changes the very concept of superiority.

An army that analyzes better can avoid unnecessary actions, reduce its mistakes, and preserve its resources.

Intelligence becomes an economy of violence.

That is an important idea.

We often talk about military technology as a force multiplier.

It can also become a retention multiplier if it improves identification and reduces uncertainty.

But this depends deeply on the doctrine.

The same reconnaissance capability can be used to avoid a civilian target or to optimize an attack.

Technology does not write its own morality on its own. Its architectures, uses, and effects are guided by the choices of those who design, deploy, and govern it.

This means that the political quality of technological powers will matter even more.

The more capable systems become, the more the values of those who control them produce consequences.

This is precisely why the race for power cannot be analyzed solely in terms of teraflops or the number of drones.

We need to look at the doctrines, the institutions, the rules.

Two states possessing the same technology can produce two radically different uses.

The algorithmic battlefield will therefore also be a confrontation of doctrines.

Some armies will opt for greater centralization. Others will distribute autonomy. Some will favor numerous, inexpensive swarms. Others will prefer more sophisticated platforms.

These choices will reflect different industries, geographies, and military cultures.

There is probably no single model.

Military history is full of technologies presented as revolutionary that did not eliminate older forms of warfare.

The tank did not eliminate the infantry.

The airplane did not eliminate the ground.

The missile did not eliminate the artillery.

Drones and robots will probably not eliminate humans from the battlefield.

They will change the relationships between the functions.

War is cumulative.

New layers are being added.

This produces an increasingly complex system.

And the more complex the system, the more strategic maintenance becomes.

This is a dimension that futuristic visions often forget.

A swarm of thousands of drones means batteries, motors, spare parts, software, frequencies, updates.

Each unit may be cheap, but the ecosystem that sustains it is immense.

High-tech warfare remains an industrial war.

This is perhaps one of the most powerful lessons of contemporary conflicts:

Volume still matters.

Ammunition.

The pieces.

Production capacities.

An extraordinary technology produced in ten copies may be less decisive than a sufficiently good technology produced in one hundred thousand copies.

That is why robotics immediately brings to mind industry.

The new race for power will also be a race for the ability to produce quickly, in large quantities, at an acceptable cost.

China currently possesses considerable advantages in several manufacturing capabilities. The United States has very powerful technological and military ecosystems.

In my view, Europe needs to regain greater industrial depth.

But it would be simplistic to turn this into a definitive ranking.

Robotics is still an extremely dynamic field. Supply chains can reorganize. Technologies can reduce costs. Smaller players can produce remarkable tactical innovations.

The war in Ukraine has shown with particular clarity that a capacity for rapid adaptation can change the balance of power.

Business systems can be transformed, tinkered with, reinvented.

War becomes a brutal laboratory.

Innovation cycles are getting shorter.

A solution works for a few months, then the adversary develops jamming, a defense, another tactic.

We need to change.

Once again:

Only mutants survive.

This sentence finds its most literal meaning here.

A military system that cannot rapidly modify its software, frequency, sensor, or tactics becomes obsolete even before the end of the conflict.

This means that the traditional acquisition chain, sometimes several years long, becomes problematic.

Military institutions will need to learn to integrate much faster development cycles.

This is a major organizational change.

How can we maintain safety, certification, and control while rapidly changing systems?

This again highlights the tension between stability and change.

A software update can improve a capability and introduce a new defect. The faster the cycles, the more difficult validation becomes.

Therefore, some of the testing will need to be automated.

AI will test AI.

Massive simulations will be able to explore millions of scenarios.

But simulation will never be enough.

Reality always produces situations that the model had not foreseen.

It will therefore be necessary to maintain a permanent link between the field and the laboratory.

This learning loop may well be the real decisive weapon:

sensor, feedback, modification, redeployment.

The one who completes the cycle faster adapts their system.

We again encounter the concept of decision time, but this time applied to technological evolution itself.

The competition takes place over several time periods:

Seconds on the field, weeks in the software, years in the industry.

A grand strategy must synchronize these times.

It is very difficult.

A tactical decision can consume in a few days equipment that the industry takes months to replace.

AI can help to model these flows.

It does not eliminate the material constraint.

The stock either exists or it does not.

That is why I always come back to the material.

The algorithmic battlefield remains a physical battlefield.

Steel, silicon, battery chemistry, rare earths, explosives matter.

Artificial intelligence does not replace matter.

It organizes it.

This is almost a very simple definition of its role: to increase the degree of organization of the physical world.

A swarm is not powerful because it contains many machines.

It is because these machines are dynamically organized.

Value comes from coordination.

That is exactly what I have always observed in human systems.

Thirty people working without architecture can produce less than ten perfectly coordinated people.

The network is not summation.

It is multiplication.

But coordination also creates a signature.

The more agents communicate, the more they can be detected or jammed.

This is a new source of tension.

A highly connected swarm can be very effective but vulnerable to electronic warfare. A more autonomous swarm communicates less but must make more decisions locally.

Therefore, communication and autonomy must be balanced.

This is a classic network problem.

In Groupe Serveur, we knew that a distributed architecture had to be able to function despite the loss of some links.

In war, this rule is vital.

The systems will likely need to dynamically switch between several modes: rich coordination when the network is available, local autonomy when it is degraded.

This adaptability is a form of intelligence of the network itself.

The system needs to know not only what is happening in the environment, but also in its own infrastructure.

Which links are reliable?

Which nodes are compromised?

It is a form of technical self-diagnosis.

The network must monitor its own condition and know its operational limits.

A drone needs to know if its sensor is degraded. A model, if its data is insufficient. An army, if its supply chain is fragile.

Sovereignty begins with this knowledge of one's own state.

It is almost Socratic: know thyself, applied to the technical system.

This may seem philosophical.

It is very concrete.

A poorly calibrated sensor that is unaware of its poor calibration produces a false sense of confidence.

A system that knows it is degraded may slow down or request confirmation.

That is the whole difference.

This self-diagnostic capacity will be fundamental in swarms.

Units will be able to report their status and redistribute roles. An agent with low battery can leave a position; another can take their place.

The collective becomes adaptive in a functional sense.

We are thus building systems that use certain properties observed in organisms: perception, adaptation, cooperation, resilience.

They are not alive.

But they use principles derived from living organisms.

The swarm is the most obvious example.

It is a form of biomimetic engineering.

And this deeply relates to my interest in organic systems.

Netnobility, as early as 1991, was already based on this idea of the network as a living organism.

What we see today is the mechanical embodiment of this principle.

The network no longer just connects computers.

It connects physical agents capable of action.

It is a gigantic transformation.

It will obviously transform far beyond the military: agriculture, logistics, industry, relief, exploration.

But war often reveals the extreme consequences first.

It pushes technologies to their limits.

This is where problems of control, speed, and error arise.

And that is precisely why they must be considered before their generalization in civilian life.

A swarm of logistics robots in a warehouse and a military swarm share certain coordination architectures.

Innovations circulate between civilian and military sectors.

The line between dual use and other uses becomes even more blurred.

A camera, a processor, an engine can have completely civilian uses and become components of a defense system.

This greatly complicates technology control policies.

We cannot simply prohibit the export of any component that could have a military use; almost all modern digital technology has a dual use.

The whole field of algorithmic warfare therefore remains materially constrained. Steel, silicon, battery chemistry, rare earths, explosives still matter.

Excessive control stifles the economy. Too little control allows strategic technologies to slip through.

No algorithm will fully resolve this political choice. It can help map the chains. The decision will remain normative.

This always brings me back to the limits of optimization. A policy is never simply about seeking maximum efficiency. It involves arbitrating between values: free trade, national security, innovation, alliances.

Numbers can shed light on things. They do not choose the outcome.

This distinction will become increasingly important as AI systems are integrated into governments. We must resist the temptation to present political trade-offs as technical solutions.

The model can estimate, in a given scenario, that a restriction reduces a certain risk while increasing a certain cost. Deciding whether this trade-off is acceptable is up to the policymakers.

The algorithmic battlefield must not produce an algorithmic government. This boundary must remain political.

But the pressure will be intense. As machines become better at analyzing certain scenarios, humans may be tempted to gradually delegate judgment to them.

This is precisely the bias of automation at the institutional level. It is not that the machine decides. We simply end up no longer questioning its recommendations. This drift can be very gradual.

The danger is not always a spectacular revolution where AI takes power.

It can be an accumulation of small delegations.

Each decision seems reasonable.

In the end, humans legally retained sovereignty but lost the practical competence to exercise it.

It is a form of silent dispossession.

That is why it is necessary to preserve parallel human capabilities.

In the military, this means training, exercises without certain aids, understanding of procedures.

In civilian life, this means education, research skills, calculation skills, and writing skills.

Cognitive sovereignty is a practice, not a status.

If a skill is no longer practiced, it atrophies.

It is almost biological.

Muscle that is no longer used disappears.

A civilization can lose some cognitive muscles.

Therefore, I prefer to think of AI as a cognitive exoskeleton rather than as a replacement.

An exoskeleton augments the body, but ideally should allow the body to retain autonomy.

This metaphor is perhaps more accurate than that of the artificial brain.

We are building extensions.

But any extension changes the behavior of the organism.

A man who owns a GPS no longer looks at the city in the same way.

The soldier who possesses a synthetic AI no longer looks at the mass of signals in the same way.

The tool modifies the user.

This is the anthropological dimension of technology.

And the algorithmic battlefield will likely produce a new type of fighter: less a direct operator of a machine than a supervisor of systems, an interpreter of alerts, a manager of autonomous networks.

This will change the training.

We will need military personnel capable of understanding code, data, and electronic warfare as well as traditional tactics.

The soldier, the engineer, and the analyst are coming together. Professions are merging. This seems perfectly consistent with the times. Major disruptions are abolishing certain disciplinary boundaries.

Industrial warfare produced the mechanized soldier. Algorithmic warfare could produce the system-soldier.

But you should not think that the body disappears. Fatigue, fear, the terrain, the weather remains. These are important.

Digital simulations sometimes give the impression of a pristine world. Military reality remains muddy, noisy, and unpredictable. Robotics will have to confront this reality.

A camera functions differently in rain, dust, and smoke. A motor wears out. A battery loses capacity in the cold. The environment is an additional adversary.

That is why reality remains sovereign.

It is a lesson I have learned throughout my life in industry and sculpture. The material resists. Steel does not always behave as the design intended. Concrete has its weight, its grip, its flaws.

The body imposes limits on the idea. AI will discover the same thing when it enters the physical world.

Robotics is the reality test of artificial intelligence.

A model can be brilliant in a virtual environment and fail in front of a poorly lit staircase. This is a fundamental difference between textual cognition and embodied action.

The world cannot be entirely reduced to discrete data. It contains friction, chance, continuity.

This resistance of reality may be a protection against fantasies of omnipotence. It reminds us that intelligence is never independent of its substrate. Even artificial intelligence depends on sensors, energy, and motors.

I almost see a Spinozist echo in it: the power to think and the power to act remain linked to a certain organization of matter.

That is why I find it fascinating that AI ultimately brings us back to matter after decades of discourse about the virtual. The cognitive revolution gives rise to an industrial revolution. Data centers, robots, automated factories. The century is not less material. It is more material still, but its matter is governed by more information.

The algorithmic battlefield will therefore be a vast crucible where matter, energy and computation unite.

This alchemical image is not decorative.

The athanor transforms through precise control of heat and time.

Autonomous systems transform the battlefield through precise control of information and movement.

In both cases, the organization produces the mutation. But any alchemy can fail if the operator loses control of the process. This is precisely the risk of swarms.

Emergent behavior can become difficult to predict. The more agents there are and the richer their interactions, the more the overall dynamics can produce unexpected states.

This is a well-known phenomenon in complex systems. The whole is not simply the sum of its parts. This is what gives them their power and their danger.

A seemingly reasonable local rule can produce surprising collective behavior. This is why the tests will need to be massive: simulations, adverse environments, boundary tests.

But no simulation will cover everything. Therefore, it will be necessary to design safety barriers capable of functioning even in the face of the unexpected.

Invariants. Certain things that the system must never do, regardless of the situation.

This is where law can become code.

Not in the naive sense where a complex legal rule would simply be transformed into a line of programming, but in the sense that certain fundamental limits can be integrated into the architecture.

Restricted areas. Excluded categories. Thresholds.

The problem is that reality always contains ambiguities. An ambulance can be hijacked. A civilian building can be used for military purposes. No simple code will solve every case.

It will therefore be necessary to combine strict rules and contextual judgment. This is exactly what human legal systems do.

Some prohibitions are absolute.

Others require assessment.

Autonomous robotics may need to develop a similar architecture.

And this is where the moral question becomes dizzying.

Can we delegate to a machine an assessment that contains an irreducible human dimension?

How far can a system go in assessing proportionality, intent, and surrender?

These are categories that often depend on the context.

A person who raises an object: is that a weapon?

A person running towards a post: are they attacking or seeking refuge?

Humans themselves make mistakes.

But that does not mean that the machine can simply replace it.

The question concerns the type of error we accept and the level of responsibility we want to maintain.

A civilization sometimes has to decide that certain functions should not be optimized to the fullest extent.

This is a deliberate limitation.

We already accept this in other areas.

Not everything that is technically possible is legally permitted.

Power is also defined by the ability to not use all of one's power.

That is precisely what civilization is: building prohibitions around our capabilities.

Nuclear weapons are the extreme example.

Deterrence relies precisely on a power that one seeks not to use.

With autonomous systems, other forms of restraint will need to be built.

Perhaps some decisions must remain humane not because humans are technically better, but because moral responsibility must remain identifiable.

This idea seems fundamental to me.

The machine's superiority over a given task is not sufficient to determine that it should receive authority.

Authority is a political category.

We may decide that a function requires a human presence because it engages our definition of ourselves.

It is a question of anthropological sovereignty:

Do we want to remain the only ones able to decide?

This goes far beyond the military.

Justice, medicine, education.

Each society will have to draw its own borders.

And these boundaries may differ depending on the culture.

We will then have several architectures of autonomy.

Some companies will accept more delegation. Others will maintain larger human reserves.

This could become a new civilizational difference.

The competition will therefore not be solely technological.

It will be normative.

What uses do we consider legitimate?

What risks are we willing to accept?

China, the United States, and Europe may develop different responses.

And these differences themselves will become instruments of influence.

A country will be able to say: our system meets this standard, it is auditable, it maintains this control.

Another will prioritize performance and speed.

The global market for strategic robotics could therefore be structured around doctrines of trust.

That is an interesting possibility.

The standard becomes a product.

A system that is technically less spectacular but perceived as more controllable may be preferred by some states.

This aligns perfectly with the European idea of normative power. But again, it must be supported by industry.

Without machines, the norm remains discourse.

We always return to the integration of the entire stack, from material infrastructure to models and data: law, model, processor, battery, motor. Power is vertical. And that verticality appears with particular clarity in robotics.

A reliance on a tiny component can cripple an entire system. A sensor. A magnet. A chip.

This makes the supply chain extremely strategic.

The algorithmic battlefield therefore begins in the mine and the factory.

It seems almost absurd when you look at a drone as a sophisticated aerial object. Yet, its capabilities depend on a battery, metals, and electronics. The geopolitics of the robot is a geopolitics of components.

We find rare earth elements, lithium, copper. The new cognitive warfare is once again linked to geology. It is a constant in the world.

Technological abstraction always relies on materials extracted from somewhere. Artificial intelligence therefore possesses a mineral geography.

That is why the Silk Road, Africa, Central Asia, and Latin America are returning to the system.

The preceding chapters are not separate blocks. They converge here. China, semiconductors, data, energy, AI, robotics: everything comes together. This is precisely the structure of the geopolitical graph. Nothing is isolated.

A decision about a mineral can affect a battery factory, which affects drones, which affect military doctrine.

The causal relationship becomes lengthy.

AI can help to navigate these chains.

But it can also give us a false impression of control.

This is a risk that I want to keep constantly visible.

The more our modeling capabilities increase, the more we can believe that the entire system becomes predictable.

However, complex systems produce disruptions.

A small event can change the trajectory.

That is exactly the logic of scientific Chaos.

Initial conditions matter.

A tiny mistake can be amplified.

A swarm is precisely a dynamic system where small variations can modify collective behavior.

That is why chaos theory is not just a metaphor here.

It is almost becoming an engineering discipline.

We need to study stability, attractors, and bifurcations.

Under what conditions does behavior remain controlled?

When does it tip?

We find ourselves directly back at the conceptual heart of the Abode of Chaos.

Chaos is never the absence of rules.

It is sensitivity, interaction, complexity.

The system can be extremely determined locally and difficult to predict globally.

That is exactly what we need to understand about swarms.

Each agent can follow simple rules.

Collective behavior can nevertheless become difficult to anticipate.

It is almost an algorithmic totality: thousands of linked elements produce a form that does not exist in any isolated element.

But, in war, this emergence must be contained by responsibility.

The artist can accept the unexpected.

The general staff cannot accept just any unexpected event.

Therefore, we need to design bounded spaces of freedom.

This is a very interesting concept:

Autonomy within a safe environment.

The system can adapt as long as it remains within certain limits.

This is probably how the most mature architectures will work.

They will seek neither absolute central control nor total autonomy.

They will define areas.

The agent can improvise the trajectory but not the destination.

It can redistribute a task but not alter certain rules of engagement.

This separation between means and ends is fundamental.

But even this formula has its limits.

The choice of means can sometimes have a moral dimension. A trajectory, a weapon, a time of day can alter the risks for civilians.

It will therefore be necessary to define which characteristics of the means remain subject to human judgment.

Once again, there is no easy solution.

Granularity becomes a method.

This is probably one of the most consistent threads in this part of the book:

Rejecting slogans because the technical world is too detailed for them.

"For or against autonomous weapons" is an insufficient question.

What degree of autonomy?

In what capacity?

Under what control?

With what reversibility?

It is this precision that can build a serious policy.

And this policy must be international, because the dynamics of the race make national decisions insufficient.

If a state imposes total restraint while its adversaries automate massively, it may feel vulnerable.

This feeds into a spiral.

Therefore, shared standards are needed, at least for certain critical functions.

Perhaps not a grand universal treaty immediately, but areas of agreement.

Do not delegate certain nuclear decisions.

Maintaining human channels for certain thresholds.

Ban certain forms of completely autonomous targeting.

Define protocols in case of loss of control of a system.

These agreements could become the equivalent of the first arms control conventions.

They would seem modest.

But they could stabilize the system.

Sometimes the details save the world.

A hotline, an identification procedure, a code of conduct may seem tiny in comparison to major powers.

But civilization relies on these invisible infrastructures.

International law is an architecture of small protocols that prevent certain chains of events.

With swarms and autonomous systems, we will need a new layer of protocols.

And probably machine-to-machine protocols.

This is a fascinating question: could adversarial systems possess standardized mechanisms to signal certain intentions?

For example: identification, status, failure.

It is almost like civil aviation with its transponders.

Obviously, in war, the enemy can lie.

But some protocols can be useful in contact areas or to avoid unintentional collisions.

We may need to develop a legal framework governing interactions between autonomous systems.

It is a dizzying idea.

International law was designed for human subjects and states.

It might have to define the behaviors of autonomous systems without attributing legal personality to them.

This would simply be an extension of the obligations of their operators:

Your system must be able to be identified under certain circumstances, respect a certain signal, and enter a certain mode in the event of a ceasefire.

This shows how software architecture will become a matter of diplomacy. Computer protocols could be negotiated like security agreements. Engineers and diplomats will have to work together.

This fusion of disciplines again.

This is probably a central characteristic of the 21st century. Inherited professional boundaries are becoming insufficient.

The lawyer must understand the code. The engineer must understand the law. The soldier must understand AI. The philosopher must understand the interface.

This is exactly the kaleidoscopic form of thinking I have always advocated. Real problems cut across disciplines.

The Abode of Chaos is a materialization of this: art, geopolitics, computer science, matter, law. It is almost a laboratory of this radical interdisciplinarity.

What was once described as heterogeneous is now becoming a structure of the world. In my view, AI confirms how artificial some intellectual silos were.

You cannot understand a military swarm using only computer science. You need biology, mathematics, psychology, law, geopolitics.

It is precisely this convergence that fascinates me. And it leads to another, even deeper question.

If machines become capable of cooperating with each other, learning certain tactics, and locally modifying their behavior, what part of collective intelligence remains explicitly designed by humans?

We can arrive at systems where efficient behavior has been discovered through training rather than written directly.

The designer sets the objectives and constraints, then the system learns a strategy.

This already exists in virtual environments.

When transposed to the physical world, this raises a much stronger question of explainability.

The system can find a solution that no one had imagined.

That is precisely the benefit of learning.

But what if this solution is strategically effective but morally undesirable?

Therefore, the constraints must have been correctly defined.

We encounter the problem of alignment again, but applied to a concrete military situation.

Alignment is no longer just a big abstract question about the values of humanity.

It becomes:

How can we ensure that the system does not seek to optimize a mission through prohibited means?

This question is extremely difficult because simplified objectives can produce unexpected behaviors.

Giving the system the mission of maximizing coverage of an area can lead it to take undesired risks.

Optimizing speed may sacrifice stealth.

Every objective contains externalities.

Humans have always known this in organizations.

An indicator becomes a target and sometimes ceases to be a good indicator.

AI can industrialize this law.

If you measure what you want incorrectly, the machine will perfectly optimize the wrong thing.

This is probably one of the most important blind spots.

The quality of the lens is more important than the power of the model.

A very powerful system with a poorly defined objective can produce disaster more quickly.

That is why doctrine must precede optimization.

You need to know what you really want.

This sentence seems banal.

Yet it is extraordinarily difficult.

Human organizations often use approximate goals because people intuitively compensate.

A human understands that "maximizing speed" does not mean driving through a hospital.

A machine does not necessarily possess this contextualization if it has not been integrated.

This is where the real world is infinitely more difficult than simulations.

The values are implicit.

Much of our behavior depends on conventions that we never explicitly state.

Building autonomous robots requires making these conventions explicit.

It is almost a gigantic philosophical experiment.

We need to transform part of our tacit morality into operational architecture.

And we will probably discover that we do not always agree on this morality.

AI does not create these disagreements.

It reveals them.

This is yet another of its mirror functions.

It forces us to clarify what we thought was obvious.

The law of war itself contains concepts that require interpretation: distinction, proportionality, necessity.

Experienced humans apply them in a context.

Translating them for an autonomous system requires asking difficult questions.

Can proportionality be formalized?

How far?

What information is needed?

The model can help.

But the final judgment may remain humane for a long time in the most serious situations.

And that is not a technical weakness.

This is a possible civilizational choice. We can decide that certain decisions must remain human precisely because they cannot be reduced to an optimization function.

That is what I call anthropological sovereignty.

It is no longer simply a question of the sovereignty of one state in relation to another power. It is a question of the sovereignty of humankind in relation to its own creations.

What area of decision-making do we want to preserve as a sign of responsibility?

The military question makes this issue particularly acute because it concerns life and death. But the same debate will arise elsewhere.

Justice. Medicine. Credit. Education.

Each domain will have to define its boundaries. This is probably one of the great philosophical debates of the century.

And it will be much more concrete than speculations about the ontological nature of these systems.

Before we decide what they are, we must decide what we allow them to do.

This is reality. Philosophy must enter into the process. That is exactly how I think: no idea is worthwhile if it is not embodied.

The next step is to change the terrain without changing the problem.

The same questions of delegation, traceability, control and sovereignty that appear here under military constraint reappear when artificial intelligence enters public administration and decision-making.

The question is no longer simply:

Who controls a machine that acts?

It becomes:

Who controls a system that classifies, recommends, grants rights, detects risks, or guides state action?

After the algorithmic battlefield comes the algorithmic state.

How can we increase efficiency without transforming informational power into surveillance and without dissolving democratic responsibility?