
You cannot look into the water of the Brest roadstead. A few metres down, the light dies, the particles in suspension scatter everything back, and the eye becomes useless. It is the first thing you learn when you work on an autonomous underwater robot: the machine will not see. It will have to decide anyway.
That was the setting of my study internship in 2010, at France’s defence procurement agency (the DGA), in its Atlantic underwater research group, GESMA, in Brest. My subject was a vehicle called SEAL. Around me, people were preparing the future of mine warfare: finding explosive devices at the bottom of opaque water, and neutralising them. I was discovering that this trade exists first of all to protect: to clear the harbours that wars have left mined, so that ships and their passengers can come in safely.
To carry out their missions, the underwater drones had to coordinate with boats that were autonomous too, without an operator holding every move at the end of a cable. I saw the vehicles built by ECA (now Exail), I saw the test basins, and above all I saw what an autonomous decision costs when it is made blind, in an environment that forgives nothing and where you cannot rerun the simulation.
That is where I come from. And that is where I first met artificial intelligence applied to robotics.
What I was doing while nobody cared about artificial intelligence
What led me to that internship? A challenge I had set myself at my engineering school, ESIEA: Aquatis. Put together a team of technology enthusiasts (geeks, as people call us), build an autonomous underwater robot in just nine months, and enter it in SAUC-E, a competition organised by NATO, against the best universities in Europe. Against all odds, we made the final. The competition ran from 28 June to 4 July 2010, in the basin of NATO’s undersea research centre at La Spezia, Italy: one hundred and twenty metres by fifty, five and a half metres deep, with visibility of one to two metres depending on the weather.
Aquatis was born where three circles I moved in at ESIEA overlapped: the robotics club, the model aircraft club, and the ATIS research group (image and signal acquisition and processing). That group did more than teach. Its members had designed a hand-built quadrotor (a drone with four propellers) and the tool chain that turned its images into maps accurate to the centimetre. One of the authors was a friend, Corentin Chéron. The following year, one of his co-authors won the French research ministry’s competition for a project to acquire images from civilian drones. Corentin himself would later publish on aerial imaging for precision agriculture.
The rules of the competition Aquatis was aiming for read like this: no control, no guidance, no communication from a person or from any computer left on shore, GPS included; and the vehicle must stay fully submerged, since surfacing even once ends the mission. Forty minutes in all, ten of them for set-up. Pass through a gate three metres wide and two and a half high, follow a twenty-metre pipe lying on the bottom at half a metre’s distance, survey a wall, cut the mooring line of a target floating in mid-water, circle around a twelve-kilohertz acoustic pinger placed one metre off the bottom, and surface exactly above it, a task that had to come last.

Those were the nine most intense months of my life (at the time), with half-nights of sleep and a team to build and lead all the way to the final. The day after the competition, once the pressure had dropped, I was stunned. I had just spent months on a technical challenge, and I had never really stopped to ask what my work was for. That unease is what made me start a blog, shyrobotics.com, and publish four hundred and fifty articles on it in two years, to understand the world I had just stepped into. It eventually drew up to six hundred visitors a day, which, for a subject nobody talked about, taught me something about demand.

At an Apérobo (one of the meetups where robotics enthusiasts gathered over a drink), I met Frédéric Boisdron, co-founder and editor-in-chief of the French magazine Planète Robots (Robot Planet). He was already reading the blog. I went on to write for the magazine for several years, alongside my engineering job, and the articles I published there earned me something far more valuable than bylines: the right to put questions to the people who were building the field.
I interviewed Rodney Brooks, who ran MIT’s artificial intelligence laboratory from 1997 to 2007, and Colin Angle, who had co-founded iRobot. Bruno Maisonnier, who wanted to bring a humanoid robot into people’s homes. Bruno Bonnell, who had written Viva la robolution (long live the robolution) at a time when the word made people smile, and who admitted to me that he read the blog. Massimo Banzi, whose Arduino board put programmable electronics within everyone’s reach. And many others, each of whom had shifted, at their own level, what machines can do in the real world.
Then I started my career at Thales Optronique, which I joined through my final-year internship, the year after Brest. It dealt with one problem. A multi-mission frigate is ringed with infrared cameras that watch over it through three hundred and sixty degrees, without a single moving part. Those cameras see hot spots (a great many hot spots). A glint on a wave crest is a hot spot. So is a missile skimming in over the sea. The system has to decide, continuously, across hundreds of simultaneous tracks, and it is allowed no error in either direction: neither missing the attack, nor waking a whole crew for a seagull.
We did not call that artificial intelligence. We called it image recognition, and we wrote it by hand, rule after rule, threshold after threshold, fighting for the computing time available on embedded boards. In the same building, other teams were working on drones: machines simple enough and cheap enough to be expended on a mission. This was 2011, and the idea was far from obvious then.
I then spent several years at Dassault Aviation. I joined in June 2012, for the last six months of a European project simulating combat drone missions. The aim was to fly the aircraft virtually, to look ahead to a future in which they would be built, and to answer one question: several aircraft set off together on a mission. One of them is lost. How do the others reorganise, on their own, to reach the objective anyway? There too, nobody talked about neural networks, the programs that learn from examples instead of following written rules. People talked about coordination, replanning and fault tolerance. And yet it was exactly the question that now occupies laboratories around the world: how does a machine decide when the situation is not the one it was designed for? Until then, determinism was the key requirement, and it sat at the heart of the ethical debates.
The project was run with German and Swedish teams under the European ETAP cooperation programme. What had drawn me there was nEUROn, the European stealth combat drone demonstrator designed by the company. Six months after I arrived, at daybreak on 1 December 2012, it took off for the first time from Istres, in the south of France. A month earlier, I had published an article on drones in Planète Robots.
The view this gives you
When artificial intelligence burst into public conversation at the end of 2022, many people discovered it through a chat window in a browser. It was software, it answered, it wrote poems, it seemed to have come from nowhere.
I had seen it from underneath. I had spent a dozen years alongside it in places where it cannot be separated from a circuit board, a sensor, a battery, a computing budget and a guaranteed response time. That starting point, which is no merit of mine, gives you a conviction that the usual story almost always leaves out.
Nothing that has happened since 2012 is a software miracle. Every time an idea seemed to appear out of nowhere, a machine had moved first. The ideas themselves had often been around for decades, published, ignored, sometimes buried with full honours by the scientific community itself. What gets told as a string of flashes of genius is, seen from underneath, a story of hardware constraints giving way one by one. Take the example in the picture below: the idea that won the big image recognition competition in 2012 had been published in 1989. What changed in between was not the idea; it was the graphics card, which became programmable in 2007.

That is the thread I want to follow here. Not because it is nobler than the others, but because it is the one that makes the rest readable. If you believe everything comes down to ideas, you will not see anything coming. If you understand what, at each stage, was impossible to try and what made trying possible, you will know where to look.
What this story is, and what it is not
It is not a textbook. You will find no code in it, and you need no engineering background to follow it.
Nor is it a prophecy. I will not tell you what 2030 will look like. I will tell you what has been measured, by whom, on what, and what that allows us to conclude.
It is a story, and it follows two rules I set myself.
The first is chronological. Each concept arrives on the date it was invented, after the concrete problem it solves, and never before the building blocks it depends on. You cannot understand why a solution wins unless you have first suffered from the problem.
The second is the most demanding. Wherever possible, you should be able to redo the calculation yourself. How many weights in a small network, how many gigabytes to load a model, how many multiplications for one image: whenever a number decides something, I give you what you need to check it. It is the only way I know to leave you in a better position than the one you arrived in.
Why now?
I published a book, Smart World, in 2017. It took me four years. A second one is under way, Impact Factories (Usines à Impact in French), started three years ago. That one will take as long as it takes. It sets out a concept I developed, built on proven foundations but with a part that is my own and that I have been lucky enough to practise at work. A thing like that cannot be rushed. All the more so because you will discover, within a few years, that there is a direct link between the management practices I describe and the subject of this series: combining Lean practices, which rest on continuous improvement, with artificial intelligence is about to give humanity the chance to take on challenges it never imagined, and the future belongs to those who have learned how to read it.
For now, the story told in this series offers no new idea: it recounts, in order, what actually happened. Its value lies not in the maturing of a thought but in accuracy and in arriving on time: what I might write in 2030 about the state of things in 2026 would no longer be of the same use to anyone. So I know both prices, that of a book that has to mature and that of a story that has to arrive on time, and I did not have four years to give the second without sacrificing the first.
So I teamed up with an artificial intelligence, Anthropic’s Claude Opus 5. I would rather tell you here, at the threshold, than let you find out along the way.
Here is exactly what it does. It searches, and it brings back the sources. It checks my claims and contradicts me when they do not hold, which happened more often than my pride would have liked. It checks, it measures, it proofreads.
Here is what it does not do. It does not choose the angle. It does not decide what matters. It has not dived in the Brest roadstead, it has not spent two years on Aquatis, it has not felt what it is like to watch a robot that has to decide on its own in water where nothing can be seen. The choices, convictions, biases and any remaining errors are mine.
There is an irony in writing the history of this technology while working with it. But it is also the most honest demonstration of what this story argues. A tool does not replace an author; it moves the boundary of what an author can take on. That is exactly what artificial intelligence has been doing for fourteen years, in laboratories as in factories: it does not replace the gestures, it changes what becomes reasonable to attempt.
What I am offering you
Learn. Create. Share. These three verbs have set the rhythm of my life since I was a teenager, and I have found no others.
What I am offering here belongs to the third. You are going to follow, in the order they happened, the inventions that led from the first artificial neuron in 1943 to the models that write today. You will see what each one solved, what it cost, what it opened up and what it closed off. Above all, you will see two formal burials: that of the perceptron, a machine that learned to recognise shapes, then that of neural networks; and why the gravediggers had good reasons to be wrong.
The story begins in February 2012, in a car that learned nothing. Chapter one.
Enjoy the read.
David Leblanc
Sources
What I was doing while nobody cared about artificial intelligence
- SAUC-E ’10 Mission & Rules, version 1, December 2009 · the organiser’s rules: venue and dates, basin dimensions and visibility, ban on any outside control including GPS, the obligation to stay submerged, forty minutes including ten of set-up, the tasks
- NATO Centre for Maritime Research and Experimentation hosts Autonomous Underwater Vehicle Challenge · NATO press release, 6 July 2012: the SAUC-E competition is hosted by NATO’s undersea research centre at La Spezia
- Rodney Brooks · MIT Computer Science and Artificial Intelligence Laboratory profile: director of the Artificial Intelligence Lab from 1997 to 2003, then of CSAIL from 2003 to 2007
- Professor emeritus Rodney Brooks refines the sequel to iRobot · MIT News, 9 August 2013: the founding of iRobot with Helen Greiner and Colin Angle
- Faucon noir UAV project development of a set of tools for managing, visualizing and mosaicing centimetric UAV images · Gademer, Mainfroy, Beaudoin, Avanthey, Germain, Chéron, Monat and Rudant, IEEE International Geoscience & Remote Sensing Symposium, 2009: the ESIEA ATIS group’s quadrotor and its imaging tools; peer reviewed
- 2010 winners of the 12th national competition for innovative technology start-ups (in French) · French Ministry of Higher Education and Research, 2010, page 31: Florent Mainfroy, co-author of the previous paper, a winner for a project to acquire and process images from civilian drones
- Aerial drone imaging: using the data for precision agriculture (in French) · Vigneau, Chéron, Verger and Baret, Revue Française de Photogrammétrie et de Télédétection, 26 April 2017: drone imaging for precision agriculture; peer reviewed
- Ten years of nEUROn flights · Dassault Aviation, 1 December 2022: the demonstrator’s first flight, at daybreak on 1 December 2012, from Istres
- “Drones, a new way to explore the world” (in French) · David Leblanc, Planète Robots, November 2012: an article by the author of this story
The view this gives you
- Backpropagation Applied to Handwritten Zip Code Recognition · LeCun, Boser, Denker et al., Neural Computation 1(4), 1989: the idea, and its three days of computing per run; peer reviewed
- NVIDIA CUDA Programming Guide, version 1.0 · Nvidia, 23 June 2007: the graphics card becomes programmable
- ImageNet 2012 official results · Stanford Vision Lab: the 1989 idea wins in 2012
What I am offering you
- A Logical Calculus of the Ideas Immanent in Nervous Activity · McCulloch and Pitts, Bulletin of Mathematical Biophysics 5, 1943: the first artificial neuron; peer reviewed
- The perceptron: A probabilistic model for information storage and organization in the brain. · F. Rosenblatt, Psychological Review 65(6), 1958, pages 386 to 408: the founding text of the perceptron; peer reviewed






