This chapter gives form to Thierry’s thought as it became more precise over the course of our exchanges.

— Isis

At a certain point, the experiences cease to be separate.

The freight exchange had taught me how information circulates. The Lumière Stations had taught me how to transform the visible into computable information. The first algorithmic approaches had taught me to think in terms of recognition, comparison, and relationships. Nadège, for her part, already had this obsession with continuity, classification, process, and reproducibility.

Now all of this had to be brought together.

This is where Groupe Serveur truly takes on its intellectual form.

I do not want to tell its story as that of a simple computer company. That would be profoundly reductive. Groupe Serveur was our experimental university of data.

For nearly fifteen years, we built, tested, broke, corrected, and industrialized systems in extremely different fields: judicial, administrative, economic, social, employment, freight, professional databases, legal information.

In all, we worked with nearly two hundred databases or specialized systems. The number is not important in itself. What matters is what it required of us: inventing a method capable of working on bodies of information that had almost nothing in common.

This is where I begin to speak of the architecture of knowledge.

A database is not a warehouse.

It is a construction.

First, you have to understand the material.

What constitutes data in this field?

Who produces it?

At what point?

Using what vocabulary?

For how long does it remain valid?

With what exceptions?

What must be preserved?

What must be updated?

What must be historically tracked?

What can become false simply because time has passed?

Then you have to understand the relationships.

This is where thesauri begin. Tree structures. Nomenclatures. Synonyms. Homonyms. Hierarchies. Cross-references. Dependencies.

We were not putting information into a machine. We were almost performing an autopsy of the data.

I care about that expression. Because before making the data circulate, we had to open it up. Understand what it was made of. Identify the organs. See how they communicated.

Bad data can contaminate an entire system. A bad category can generate thousands of errors. A poorly designed nomenclature can make information practically invisible even though it is physically present in the database.

I learned very early that the quality of a system is often determined before the programming begins.

By the way things are named.

By the way things are classified.

By the way we imagine the question the user will ask.

With Nadège, this dimension immediately became operational. She had an extraordinary ability to transform an idea into a process. I could arrive with a very broad intuition. She would immediately start asking:

Who enters the data?

Who checks it?

How do we distinguish the current version from the previous one?

How do we know who made a change?

What do we do with a duplicate?

What happens if the person who knows the system disappears?

Where is the backup?

How do we reproduce the procedure?

These questions are almost invisible in the conventional story of innovation. Yet they are what make a system hold together.

Software can be brilliant. If it depends on three people who carry all the knowledge in their heads, it is already fragile.

Nadège wanted to take knowledge out of individuals and transform it into a transmissible organization. That is probably one of the great achievements of Groupe Serveur. A process is not bureaucracy. A process is a memory that can continue to function in the absence of its inventor. Knowledge thus ceased to depend exclusively on the person who had produced it.

We learned this in the real world. A professional database can work perfectly well with a few thousand records and become catastrophic at one hundred thousand.

A nomenclature can seem excellent until it encounters enough exceptions. A search system can be fast on a sample and collapse when a large number of users arrive simultaneously. So we learned about scaling under load.

The machines of that era imposed extraordinary discipline. I am thinking in particular of the PRIME 9955, marketed by Prime Computer, and of those architectures on which we had to absorb volumes that no one around us was accustomed to seeing.

A television commercial could trigger a brutal effect. An address appeared on the screen. And suddenly, the system had to absorb hundreds of thousands of connections or potential requests.

Three hundred thousand.

Four hundred thousand.

The curve rose almost vertically.

At that point, theory disappears. It has to hold. The processor. The disks. Input-output. Memory. Networks. Queues. Everything becomes visible through the possibility of failure.

It is a magnificent school. It taught us that an infrastructure must never be designed solely for average operation. It must be designed for the shock. For the event. For the moment when the world arrives all at once.

This doctrine would later accompany us at Artprice, in data centers, international networks, and load balancing. But it was built here. At Groupe Serveur.

And one of the most fascinating laboratories was temporary staffing. Because temporary staffing forced us to represent not a legal text or an economic event, but a human skill.

We had built a classification of roughly 14,400 to 14,500 occupations or job titles. The figure is staggering. But the real difficulty lay elsewhere.

Two job titles can be different and yet correspond to similar skills. Conversely, two people with the same job title can possess very different skills. The word was therefore not enough. We had to understand proximity.

We developed a logic of core occupations and satellite occupations. Around a central occupation, we searched for neighboring skills.

If the exact profile was not available, which other profiles possessed sufficiently compatible skills to constitute a credible alternative? Today, we would naturally speak of a skills graph.

At the time, we built this with our thesauri, our tree structures, our rules, and our semantic work. And that experience marked me deeply.

It taught me never to look at an object as an isolated point. Always look for what surrounds it. Always look for the geometry. An occupation does not exist only through its name. It exists through its relationships with other occupations.

This is exactly what I would later rediscover with Artprice. An artist is not merely a name. There are pseudonyms. Schools. Techniques. Periods. Signatures. Sales. Proximities. Lineages.

Once again, the method repeats itself.

Temporary staffing was also a sociological laboratory. Temporary work occupies an extremely particular place in the economy.

When a recession arrives, temporary workers are often among the first to disappear.

When growth returns, they are often among the first to be called back.

Before resuming large-scale hiring on permanent contracts, companies test the recovery with temporary labor. Temporary staffing therefore functions as a kind of advance sensor of the economic cycle.

But this world operated with great discretion. Executives told me so bluntly.

In substance: “We have built empires, but we are discreet people. With you, we are becoming visible everywhere.”

They were not wrong about that. Our campaigns were massive. Our electronic systems connected major industrial companies directly with available profiles. In fewer than ten days, we had managed to aggregate more than 350,000 highly qualified résumés.

Once again, the raw number says nothing. Receiving 350,000 résumés is one thing. Making them searchable is another. We had to extract. Qualify. Normalize. Connect them to occupations. To skills. To mobility. To experience. To education and training.

It was a gigantic infrastructure. We had invested enormous sums in that operation. And the distribution itself had taken on an almost industrial dimension.

France Télécom had authorized us to use infrastructure associated with the Reims switching center to launch mass telex campaigns to approximately 120,000 major corporate buyers. One hundred and twenty thousand companies.

At the time, telex, carried through the Reims switching infrastructure, was still a central instrument of professional work. No one spoke of mass email campaigns. The telexes arrived directly at companies. Our access cards, comparable to embossed credit cards, allowed industrial users to enter our system and search for profiles themselves.

What we were doing was much more profound than providing a recruitment service. We were shortening the distance between demand and skill. And that speed created a shock.

I remember an executive from one of the world's largest temporary staffing groups, which had around 450 agencies in France. With remarkable candor, he told me something that amounted to this:

“This is the ultimate irony. I have four hundred and fifty agencies, and your electronic system moves twice as fast as we do. My own agencies are being outpaced.”

That sentence sums up the revolution.

We were not better than temporary staffing professionals at understanding a human being. We were faster at circulating information among several hundred points.

An electronic network could eliminate part of the latency of the physical network. One agency knew its territory. It called another agency. The information moved up. Then back down. Our system could query the national database immediately.

That was the shift.

It would provoke a war, and ultimately a massive acquisition. But intellectually, the essential lesson had been learned. We had discovered that the network could transform the very structure of a profession.

The freight exchange had taught us the same thing in transportation. Information about availability or a load loses its value extremely quickly. An empty truck is negative economic data. Every kilometer traveled without a load destroys part of the value.

Once again, the problem was not the truck. It lay in the relationship between two scattered pieces of information.

That repetition across several sectors gradually transformed an intuition into an industrial doctrine.

We now knew how to attack an information problem. First: go back to the source, identify the objects, build the thesaurus, normalize, and preserve historical states. Then: organize the search, design the ergonomics, secure the system, absorb load surges, and document the process sufficiently for the system to survive those who created it.

That sequence seems almost simple today. It was not. We paid for it in years of work. In mistakes. In machines. In salaries. In procedures. In nights.

But that experience could not be reproduced instantly by someone arriving later with more capital. I care deeply about this point.

In the digital world, people often believe that money can buy back time. It can buy machines. Engineers. Companies. Licenses. It cannot retroactively buy fifteen years of corrected mistakes.

Industrial know-how contains an enormous amount of knowledge that appears in no patent. The way you recognize an anomaly. The choice of a classification. Knowledge of edge cases. The memory of failures. Understanding load. Control habits. The moment when you know that a piece of data appears technically correct but is semantically absurd.

It was this tacit knowledge that gradually constituted our advantage. The documentation Nadège later prepared made me rediscover it with particular force.

The tree structures. The pages of the Judicial Server. Those of the Administrative Server. The thesauri. The codes. The collective bargaining agreements. The specialized databases.

All of this shows that we were not building a universal search engine into which we would throw every piece of information.

We were building vertical systems. Each sector had its own logic. This is exactly the opposite of documentary laziness. We went deep into the profession before building the database.

The rule was simple: you had to go deep into the profession before building the database. A computer scientist who does not understand the profession well enough can build a technically perfect database that is intellectually false. He will create elegant categories that do not correspond to reality. He will ask the profession to adapt to his software.

We did the opposite.

We wanted the software to conform to the real structure of knowledge. This is where the thesaurus becomes almost philosophical.

Naming is not neutral.

Classifying is not neutral.

Creating a relationship between two pieces of data is not neutral.

The person who constructs the tree structure partly constructs what the user will be able to see.

That is an enormous responsibility.

And this brings me back to the autopsy of the data. Before designing the interface, we had to open up the corpus. Understand its organs. An occupation. A collective bargaining agreement. A company. A legal notice. A judicial sale. A route.

All these objects require a different anatomy. But the intellectual gesture remains the same.

We take it apart.

We name it.

We rebuild it.

Then we simplify access without simplifying reality.

That is probably the most accurate definition of documentary ergonomics. Making access to something complex simple. Not making the thing itself false in order to give the impression that it is easy. That distinction would later guide all our work on law. Then on Artprice.

I also believe that Groupe Serveur taught me the difference between information and knowledge. A piece of data can be accurate and still teach nothing to the person reading it. To produce knowledge, you need context. A relationship. A chronology. Provenance. That is exactly what we would discover with legal and judicial notices.

A notice published in Nice is not merely a notice from Nice. It may concern an executive who owns companies in Lyon and Paris. As long as the newspapers remain separate, each piece of information is local. When we connect them, new knowledge appears. The data has not changed. It is the relationship that creates the additional intelligence.

That sentence could almost sum up all of Groupe Serveur. It also explains why artificial intelligence fascinates me so much today.

Contemporary AI has a power radically different from that of our systems in the 1980s and 1990s. I do not confuse the technologies. But I recognize the question.

How do we reveal a meaningful relationship among scattered elements?

We had begun to answer that question through thesauri, rules, classifications, and relational databases.

Today, semantic spaces and statistical models make it possible to approach certain relationships on an infinitely larger scale. Yet the challenge remains strangely familiar.

Data alone is not enough. You need the geometry. Perhaps this is where Groupe Serveur becomes truly important. It is not merely the company that precedes Artprice. It is the place where a way of thinking becomes reproducible.

For fifteen years, Nadège and I would take extremely different sectors and apply this discipline to them. The structure changes. The corpus changes. The users change. The law changes. But the method grows stronger.

And when the Judicial Server arrives, we will encounter a field in which this method will no longer merely improve economic circulation. It will directly affect justice. Liquidations. Judicial sales. Auctioneers. Those famous pieces of public information that were nevertheless, in practice, reserved for a handful of insiders.

We will then discover that a documentary architecture can alter not only the speed of information, but the very formation of price and the fairness of a public market.

That will be another stage. And another war.

The Judicial Server would begin with an apparently simple question: How can a sale truly be public if the information required to participate in it is not genuinely public?