The model: just an algorithm

Yes, the term can look complicated at first glance. Algorithm. But in reality, an algorithm can be just this:

That is the difference with 0 + 10: here you naturally pictured, even for a fraction of a second, an accumulation of 1s added one after another.

An AI is made up of one or more algorithms working towards its goal. That could be finding the shortest route to the restaurant (your GPS), writing you a report, producing a video, running simulations in research… Broadly, an AI is just a calculation system so complex that you see intelligence in it, and sometimes you personify it.

But in reality, ChatGPT is not aware that it is an AI, nor of what reality is, nor of what text is.

Anyway, back to our sheep, or rather to our AIs. Different kinds of AI means "some AIs are better than others in certain domains". For instance, if you want a fully explainable system, cheap in compute and deterministic (one that does not change its mind when you ask twice) for your GPS application, you will not pick an LLM but rather a graph-solving algorithm, symbolic AI, like your expert systems.

Symbolic AI

Ah, Good Old-Fashioned AI (a term coined in 1985, so already not very cool)! And yet it is far more present in your daily life than you think.

I will be a little biased on this part: I worked for 7 years at Golem.ai (now Miralia), which started out as my final-year project. The company does text analysis through a symbolic algorithm, which makes it possible to classify messages, among other things.

Symbolic AI (or classical AI) rests on the idea that human intelligence can be explained by the explicit manipulation of symbols, concepts and logical rules.

Unlike machine learning, which relies on statistical training from data, symbolic AI builds on knowledge modelled directly by human experts, in the form of facts, ontologies and rules ("If… Then…"). Its behaviour is therefore deterministic and transparent: faced with a problem, a symbolic system uses an inference engine or a search algorithm to logically deduce the solution by combining the given rules.

The main advantage of this approach is its total explainability: you can retrace exactly every step of the reasoning that led to a conclusion, with no black box effect. Extremely handy when you need auditability.

It excels in domains demanding absolute rigour. Route calculation (graph theory) is a very good example: the video game characters that get from point A to point B without getting lost are generally driven by a symbolic algorithm called A*, which can be summed up as "explore roughly in the right direction to find the best path".

In exchange, this family of AI is very rigid: it struggles with nuance, with noise in the data or with the unexpected, and becomes extremely complex to maintain as the rule base grows.

That is why there is quite a bit of discussion about combining the agility of generative algorithms with symbolic ones. Developers actually have a working example of this all the time without realising: when your Claude Code (or a competitor) generates a Python script and runs it to transform data, from an Excel file into a database for instance, that is roughly the idea. It does not rewrite an enormous request containing every modification with tokens: it writes a small script that does the job.

Machine learning

Machine learning is the opposite approach to symbolic AI: instead of explicitly programming decision rules ("If… Then…"), you give a computer large quantities of data and an objective, then let it discover the underlying patterns and rules on its own. Rather than being dictated by a human, the behaviour of the program is deduced from the examples it has analysed.

It works through statistics and the progressive tuning of internal parameters (notably the "weights", in the case of neural networks). During training, the model makes predictions, measures its error against reality, then adjusts to reduce that error. The more representative data you feed it, the more accurate its predictions become.

• ▸Supervised learning. You provide data with its answers (labels). Example: 100,000 photos of cats and dogs labelled "cat" or "dog". The model learns to tell them apart.
• ▸Unsupervised learning. You provide raw data with no labels, and the model looks for hidden structures. Example: grouping customers into segments based on their buying habits, without predefining the categories.
• ▸Reinforcement learning. An agent interacts with an environment and receives rewards or penalties. Example: an AI that learns to play chess or to fly a drone by trying actions and learning from its mistakes.

Deep learning

Without going too deep into the technical detail, deep learning is a sub-branch of machine learning based on deep neural networks (stacks of computation layers).

The big difference? Unlike classical machine learning, deep learning can analyse very raw data directly (images, sound or text) and work out on its own which details matter for a decision, without a human having to spell it out.

Brilliant at correlating… far less so at explaining

Concretely, machine learning is brilliant at finding correlations in complex data, and at handling cases where you have no genuinely defined rules. Let us take an example.

And there you have it: you understand why the statistical approach, with a large number of examples, is necessary.

In theory this gets better with a larger quantity of data, even if in reality many other factors come into play, starting with the quality of the information provided. The problem is that machine learning has serious flaws. Beyond the cost in compute and energy: its dependence on the initial data, and the fact that it is humanly impossible to understand in detail how the model reaches a decision. We can only interpret it.

Let us go back to image analysis with this example published in 2016 by the researcher Marco Tulio Ribeiro and his team to present LIME (Local Interpretable Model-agnostic Explanations), a tool designed to open up "black boxes".

The "Husky vs Wolf" experiment

The researchers trained an image classification model to tell huskies from wolves.

• ▸The results on paper. The model showed excellent accuracy (around 90%). On most test images, it never got it wrong.
• ▸The suspicious error. The model classified an innocent husky as a wolf.

Using LIME to display the exact area of the image the model was focusing its computation on (the decisive pixels), the researchers discovered that the model was analysing neither the ears, nor the coat colour, nor the shape of the eyes of the animal. It was looking only at the background:

• ▸Snow present → wolf
• ▸No snow (grass, soil, indoors) → husky

Because in the training dataset, every wolf photo had been taken in winter in the snow, while the husky photos were taken in gardens or on lawns!

Fine, this problem does not sound that serious, especially since LIME let us "explain" it. But in reality, the solution offered here is very limited.

When the stakes become decisive: the Amazon case

What happens when we move to a more consequential subject? When the problem comes from the initial data and nobody asks the right questions?

Amazon initial goal was the dream of every HR department: build an algorithm able to read 100 CVs and automatically surface the 5 best profiles, with a rating from 1 to 5 stars.

1. The cause: a dataset biased from the start

To train this model, Amazon engineers fed it the CVs submitted to the company over a 10-year period, along with the hiring decisions made at the time.

The model did exactly what machine learning is designed to do: find recurring statistical correlations in past successes in order to predict future ones.

2. The black box effect and the misogynistic drift

Rather than assessing actual skills (code, degrees, projects), the algorithm learned on its own to use gender as its strongest predictive criterion. Concretely:

• ▸Direct penalisation. The tool removed points if the word "women's" appeared in the CV, for example "captain of the women's chess team" or "president of the women engineers club".
• ▸Degree downgrading. The algorithm lowered the score of candidates from two private women-only universities.
• ▸Favoured vocabulary. The tool favoured action verbs statistically more common in male CVs, such as executed or captured.

3. Why the fix failed

When the engineers realised the drift, they tried to correct the problem. But with complex machine learning, banning a word is not enough: they forbade the algorithm from reading the word "women's", and the model immediately found proxy variables. It detected gender through other indirect clues: extracurricular activities, turns of phrase, the choice of certain words.

In 2017, finding that they could neither guarantee the neutrality of the tool nor understand the full set of internal micro-decisions of the model, Amazon definitively abandoned and disbanded the project.

But did I not tell you that LLMs belong to this same family? And guess what: CVs are increasingly analysed by AI. So what do candidates do? They adapt their CV so the AI looks kindly on them.

You can see, then, why it matters to anticipate the potential problems of an algorithm that cannot explain itself, and above all to have a firm grip on the training data to avoid harmful biases. Or else…

LLMs

1. The Transformer revolution (2017)

For a long time, to process text, AIs read words one by one, from left to right. The problem: by the end of a long paragraph, they had already "forgotten" the beginning.

In 2017, Google researchers invented a new neural network architecture that changed everything: the Transformer. Its secret weapon? The attention mechanism (self-attention). Thanks to it, the AI no longer reads word by word: it looks at the whole sentence at once and understands the global context.

Example: in "The lawyer ate an avocado" (in French, both words are avocat), the AI instantly understands that the first is a legal professional and the second a fruit, thanks to the surrounding words.

Originally this AI worked with two building blocks, because it was designed for translation. The encoder (the "understanding") read the whole sentence at once to grasp its deep meaning; the decoder (the "generation") took that meaning and wrote the translation word by word.

2. The decisive turn: why we threw the encoder away

While trying to create AIs able to do everything (chat, code, summarise, reason), researchers realised something brilliant: we no longer need the encoder. Today models such as ChatGPT, Llama or Mistral are decoder-only architectures.

• 01Continuing text already means having understood. For a decoder, everything is a story to be completed. Give it a question and the logical continuation is the answer. Give it a text in French and the logical continuation can be its translation. We no longer need to separate "understanding" from "creation".
• 02Simplicity above all. Keeping a single block made it far simpler to train giant models on trillions of data points.

How does it work in practice? The causal mask

This is where the real technical subtlety lies. Since there is no longer an encoder to read the whole text in advance, the decoder uses what is called a causal mask. It is a strict mathematical rule: the AI moves blindly towards the future. It is only allowed to look at past words to guess the next one.

If the model has to write "The cat sleeps on the sofa":

• 01It sees [start] → it predicts "The"
• 02It sees The → it predicts "cat"
• 03It sees The cat → it predicts "sleeps"
• 04It sees The cat sleeps → it predicts "on"…

Each word produced is immediately fed back into its field of vision to predict the next one. That is what we call autoregressive operation.

And for image, video or music? Same thing!

We have talked a lot about text, but what about generative AIs for images (Midjourney, DALL·E) or music (Suno, Udio)? Bar one detail, the underlying logic is exactly the same. For a computer, everything is mathematics: a text is a sequence of words, an image a grid of pixels, a piece of music a sequence of audio frequencies.

• 01For music. The AI applies exactly the same recipe as for text. Instead of predicting the next word in a sentence, it learns to predict the most logical next note or chunk of sound.
• 02For images (diffusion models). The AI first learns to destroy an image by adding "noise" to it (like static on an old TV screen). Then its neural network learns to do the reverse: remove the noise step by step to make a clean image emerge from chaos, guided by your text (prompt).

When memory is influenced

Right, that was theoretical but interesting. Now let us talk about less amusing things. Remember the observation: an LLM does not think and is not looking for "truth". It simply reflects the statistics of the texts it was trained on… with all their qualities, but also their prejudices and their ideological leanings.

That is why Yann LeCun, among others, insists that LLMs are only a kind of large memory. But what happens when that memory is… influenced?

In June 2026, the Direction générale du Trésor (at Bercy, the French finance ministry) was testing an internal tool called HéphAIstos to help around a hundred senior civil servants with administrative and economic analysis tasks. The model chosen under the hood was Qwen, a very capable LLM developed by the Chinese giant Alibaba.

• 01Slanted answers. During testing, several officials reported that the AI gave biased answers, or answers aligned with Beijing political doctrine, as soon as they asked questions touching on China, international trade issues or geopolitics.
• 02Immediate halt. After only three weeks of experimentation, the Treasury stopped using Qwen outright.
• 03Sovereign switch. The very next day, the administration replaced it with a model developed by the French start-up Mistral AI.

Now, we have elections coming up, and we know that French citizens will ask AIs for their opinion. What does that mean for democracy?

Hallucinations

Since the LLM only computes the most likely next word, it does not know whether it is telling the truth. If a piece of information is missing, it will invent facts, laws or historical quotes with perfect confidence.

Data theft and copyright

To train these models, companies hoovered up billions of web pages without asking authors, artists or media outlets. This raises a major ethical and legal problem around intellectual property.

Open weight is not open source

Today, companies such as Meta (with Llama) or French gems like Mistral AI make their models available "for free" to everyone. The media often say they are open source. Except that is wrong! In reality they are open weight. And the distinction changes absolutely everything ethically and legally.

In a neural network, the weights are the billions of numbers adjusted during training. They are the final mathematical settings of the model.

• ▸What you get with open weight: the finished "black box". You can download it, install it on your server and run it for free.
• ▸What is kept from you: the recipe, meaning the exact database used to train it (the raw files hoovered from the web), and the precise code of the training algorithm.

Why do the giants keep the data secret?

• 01Legal risk (copyright). To train these giant models, petabytes of paid books, press articles, images and code were hoovered up without authorisation. Making the dataset public would hand rights holders the material evidence to sue.
• 02Trade secrecy. The real value of an LLM today lies not only in the neural network architecture, but in the cleaning and sorting of the data. Revealing the dataset means handing your painstaking work to your competitors.
• 03Security risk. A 100% open model would let anyone see exactly how to lift the safety barriers, for example how to stop it from explaining how to build a weapon or from generating malicious content.

There are nonetheless cases of genuine open source. For the Nemotron project (with the Nemotron-3 or Nemotron-CC families), NVIDIA did not settle for sharing the final downloadable model. They opened all three pillars:

• 01The model weights: the finished AI brain, ready to use.
• 02The training code and recipes: the algorithms, configurations and source code that allow retraining or modifying the model in exactly the same way they did.
• 03The datasets: NVIDIA published trillions of training tokens on Hugging Face, including raw text, code and safety filtering data.

So why open up, if you are not opening everything?

It is actually fairly simple, and this remains my interpretation. In the case of big tech, the goal is generally to catch up with competitors and impose your standard on particular markets. That is why even OpenAI (with very debatable success, though I think that was deliberate) had released its open weight model gpt-oss. But at the scale of nations it is also a way to suffocate the market: understand that China, arriving after the American models, has everything to gain from people using Chinese models rather than paying American companies.

And for open source? Nemotron is NVIDIA… which is not a software vendor trying to sell you a ChatGPT subscription. Its business is selling graphics chips. The more data and code the community has to create AIs, the more companies will buy NVIDIA servers to run them. By giving away training datasets of immense quality (very well cleaned) to everyone, NVIDIA becomes the standard the whole of global research works on. And if more people use models… that means more chips sold ;)

To be clear, I am not cynically saying nobody sincerely does open source for reasons of transparency and values. But this is my reading of the current situation. Which makes it awkward to have only one player on the European continent, yet reassuring, in a way, to know that this "fight" guarantees us an almost continuous stream of open weight models: as long as the gap between first and second place is not astronomical, those in second will always want to try to suffocate those in first ;)

A little bonus: evolutionary algorithms

Evolutionary algorithms are proof that nature had already invented the best optimisation method long before computers arrived. To understand this AI, just look at the finches of the Galápagos Islands observed by Darwin: those birds never "decided" to change the shape of their beak to crack seeds better. Nature simply generated random variations, mercilessly eliminated those that could not feed themselves, and let the survivors pass their advantageous beak to the next generation. An evolutionary algorithm does exactly the same thing, but at millions of computations per second.

Concretely, the algorithm starts by creating a "population" of thousands of completely absurd, randomly generated solutions. It then subjects them to environmental pressure, a scoring function that instantly eliminates the worst candidates. The rare survivors are then "crossed" with one another, their code is mixed, and a pinch of random mutation is injected to explore new avenues. This is an AI that progresses without understanding: it does not try to analyse the causes of a problem, it simply accumulates the small successes of chance.

One of the most striking examples of this "digital Darwinism" remains the design of the NASA ST5 satellite antenna. Engineers let an evolutionary algorithm twist virtual wires over thousands of generations. The result? The AI produced a completely twisted, asymmetrical shape, resembling a deformed paperclip that no human would have thought to draw, but whose electromagnetic performance surpassed anything classic engineering logic had produced.

And in practice?

Right, we have covered the theory — but what about practice? We will dig into the detail once we have been through all the components; I just want us to share a base of understanding before going further.

But to give you a few pointers: when you work with AI, you work with a tool that only works X times out of 100, because of its statistical nature and, generally, the nature of the data you feed it. To a computer, "hi how are you?" and "hello, how do you do?" are different data, and they can produce different results. In the same way, say hello to your ChatGPT twice: you will see you do not get exactly the same answer.

To handle this, AI engineers use metrics that measure how reliable their solution is. The usual one is a score called F1, or one of its variants. I am not explaining it right now, but understand that we are trying to determine how often the AI will get it wrong, and how serious that error is.