Generation AI - by Vincent Barat

Episode 1: Gen AI? Did they make others after Generation Y and Generation Z?

Not a week goes by without some new “nonsense”. And no, I'm not talking about international news (from that point of view, we tend to have a daily update), but about AI, and more specifically about Gen AI, the ultimate buzzword. For the past three years now, and since the public launch of ChatGPT, the media, conferences, trade shows, but also certain “experts” and opportunistic entrepreneurs have been making it their favorite topic. OpenAI's new models, Musk's takeover bids, Deepseek's Chinese and low-cost foray causing Nvidia's share price to plummet, questions of European sovereignty with our flagship Mistral, the agentic with Manus... The succession of these news items does not make the field easy to understand.

Those who have been following us for a long time at Albert know our aversion to corporate “values”. However, our sense of nuance (and sarcasm) characterizes us rather well: facts versus emotion, and subtlety versus simplism. This is why the subject of AI has never been a marketing platform for our company, which wants to stand the test of time and not fall victim to passing fads. However, with its commodification, we believe that the subject is reaching a pivotal moment that calls for a clarification of our positioning.

AI affects us in 3 ways through our activity:

1. Its use in our daily production workflows;
2. Its impact on the workforce and skills of our customers;
3. And of course, saving the best for last, its integration into our service offering, and a few classified ads!

Through this series of publications, which will follow this outline after an introduction, we wish to share with our community - made up mainly of HR professionals at the forefront of organizational transformation - the fruit of three years of observation, reflection and development.

But before talking about the ills of AI, let's talk about the words of AI.

Rest assured, the idea here is not to go back extensively over the history of AI, because the resources on the subject are abundant and probably better than anything I could write. But we would like to recall here a few points that are important for neophytes. Skip the paragraphs where you learn nothing, you won't offend anyone. But we think it's essential to democratize AI jargon so as not to be taken in by sellers of silicon dreams.

So here is a little AI lover's dictionary to help you find your way around. Except that it's not in order. And I'm not particularly in love with AI. Anyway, you get the idea.

AI vs GenAI: the meaning of artificial intelligence is very unclear. In what follows, we will frequently use the term AI to refer to GenAI. Until a few years ago, Machine Learning was considered a form of AI because of its ability to detect correlations and become predictive (it is also used in Albert). Its widespread use has relegated it to the rank of a simple statistical algorithm, except in the eyes of the EU, which makes no distinction in the AI Act.

Gen AI: Generative Artificial Intelligence (Gen AI) is thus the visible part of the technological wave we are observing, in the same way that the Web or mobile phones were in their time. The less visible part - “traditional” AI - has already been at work around us for a long time, in medical research, recommendation algorithms, fraud detection, video games, high-frequency trading, etc. These AIs are highly specialized in a single task. Gen AI is different. On the one hand because it is more generalist, but also because it “produces” knowledge. At least, apparently, because in reality it only generates content. In reality, there is a vast ecosystem of specialized AI: in sound, image, text, voice, video, and some that manage to mobilize several “specialties” to respond to the user's request (the famous “prompt”).

LLM: the most spectacular achievements and hallucinations of Gen AI are often textual. In this case, they are based on Large Language Models. You will find thousands of popular videos, but I will summarize them for you in a metaphor. Do you remember T9 for writing SMS messages on your Nokia 3310? It's more or less the same thing. But instead of guessing the character or word you are trying to write, the model makes a prediction about the next hundred or thousand words, based on a very large training set. That's why it's notoriously hard for it to answer the question “how many R's are there in strawberry”. In most cases, the people asking the question are interested in the R between E and Y, not the very first one between T and A.

Training: as mentioned above, algorithms - including LLM - are based on a training mechanism, whether supervised or not. The way the algorithm is trained and the starting material influences the answers, hence the famous biases. When Yaël Braun-Pivet (President of the French National Assembly, the equivalent of the House of Reps or the House of Commons) asks two different AIs to represent the leaders of Parliament, one represents two men, the other one man and one woman. Statistically, the first one is not wrong and is only reproducing a historical bias. The second one is not faithful to reality because it introduces an inclusive bias, but with the aim of creating a more positive result. A burning question therefore concerns the training material of these models: this raises questions of intellectual property (“I am a journalist, singer, content creator, author: I have never given permission to any AI to impersonate me, to represent me or to answer questions in my place”), and of course questions of security in a professional context: I work in R&D in a pharmaceutical laboratory. By constantly asking questions on confidential subjects to a “shared” AI, am I not taking the risk of inadvertently training it, allowing another lab to benefit from the results of my research? Am I not also taking the risk of contamination? After how many times repeating that 1+1=3, will the AI end up believing it?

AI publishers: publishers and AI models are often confused. Publishers are either generalists (Meta, Google, X, etc.) or specialized in AI (OpenAI, Anthropic, Mistral, etc.). These companies develop AI models, which are then found in products that they may or may not distribute.

Products: these are commercial applications owned by AI publishing companies or others. The best known is ChatGPT, which belongs to OpenAI; Grok belongs to X; Le Chat to Mistral, Claude to Anthropic, etc. Some of these commercial applications are paid or freemium models. To run, they require often costly infrastructures and use very specific chips to perform the operations expected of them. The more operations performed, the more costly it is in terms of energy, computing time, etc. To quantify this, we talk about “tokens”. Some requests consume a few tokens (e.g. “Hi, how are you?”) while others consume many more (e.g. “rewrite War and Peace in inclusive writing”). The truth is that today nobody knows exactly how much a prompt costs, nor how much it should be charged to the customer to have a sustainable economic model. These applications are based on AI models, whether or not owned by the company. For example, if you use Dust.tt, a solution for easily creating AI-based assistants, you can choose the model that runs behind it. And yet Dust does not publish any AI models of its own.

AI models: this is a statistical model pre-trained with information updated on a certain date. These models are characterized by their performance, the number of parameters they have, their quantization, their “weight”... Since we are on the subject of these operations, there is no magic here, we are literally talking about multiplications and additions. But billions of times per second. It turns out that the chips found in graphics cards are very strong for this type of calculation, hence the shortages we are experiencing (and the fluctuations in the share prices of GPU producers). Among the known models: GPT 3.5, GPT 4, GPT 4o, o1 are all models available in ChatGPT (OpenAI). Sonnet 3.7 is available from Claude (Anthropic), Deepseek R1 from Deepseek... Some of these models are so-called open, others closed. Anyone can download the open models (such as Mistral 7B from MistralAI). Yes, even you if you want! You can install LLM on your computer for a few dozen gigabytes (I recommend LM Studio to get started). There are even kinds of App Stores for that! You may have already heard of Huggingface, the French unicorn that offers hundreds of models. Of course, the most powerful models require particularly expensive equipment (graphics cards in series, or even servers with specialized chips).

AI platforms and APIs: as not everyone can turn their youngest child's bedroom into a data center, it is sometimes more economical to “rent” infrastructure to run models. The computing capacities of the AI are then called upon remotely and relocated in the data center. When you hear about APIs, it is often a metonymy (or synecdoche, let’s debate) that refers to a hosted platform. Basically, requests are made in a product, which will “type” in the AI API, located elsewhere, before returning to the product. Almost all the major publishers offer platforms. In the gold rush, it is the shovel sellers who get rich...

Number of parameters: every time you hear 7B, 131B, etc., it is actually the complexity of the model that is reflected in this metric. The B stands for billion. So a 131B model has 131 billion parameters. The more parameters there are, the better. On the face of it, the one with the biggest (complexity) is the strongest. But some argue that it's not size that counts, but how it's trained. All models are not equal.

Quantization: sometimes, with equivalent model and complexity, a model is “lighter”. Often, it uses variables with a different quantization (e.g. 4 bits instead of 64 bits). In simple terms: the variables used are smaller, so the model is more approximate, but retains many of its features.

Distillation: we're not talking about Cognac here. Hold on tight, we're taking Gen AI knowledge up a notch. Remember when I said that the algorithms are trained in a more or less supervised way? Imagine that I train a model, but with another AI model, just to imitate it without anyone noticing. Distillation is exactly that: I take Llama 8B, and I teach it to behave like Deepseek R1

AGI: Artificial General Intelligence. The ultimate quest, the end of the game: an autonomous AI so powerful in all areas that it surpasses humans. Today, it is a theoretical concept that has become a marketing argument. Recent models (o1 and o3 at ChatGPT, R1 at Deepseek) have, for example, introduced the capacity for reasoning, which would bring us closer to AGI.

Reasoning: In basic terms, AI is capable of breaking down the problem into smaller pieces (Chain of Thought). But also of going back in its text predictions (remember T9?), to respond with something more coherent, less prone to hallucinations, slower too, and cleaner. It is sometimes unsettling to use, because you really get the impression that the AI is thinking. But beware, it is absolutely not a question of reflection in the cognitive sense of the term. This perception is fed by a simple trick: the AI displays text while it is responding to the prompt. This text gives the user the impression of seeing the machine's train of thought. To top it all off, a company like OpenAI bases all its communication on fear marketing with a barrage of tweets: “We tested our next model, which comes out next month, internally, and ooh la la, we scared ourselves. Our Ethics & Compliance manager even resigned because it looks so much like a real human being. When I watch Nolan's movie Oppenheimer, I feel like I'm watching my future biopic. By the way: the new model will only be available to paying customers. Like and subscribe, xoxo”.

Agentic AI: the hot thing of the moment. Imagine an AI assistant that doesn't just suggest a date for a meeting, but sends the invitation, books a room and adjusts the schedule if someone declines. That's exactly what AI agents are: the model (often an LLM) doesn't just passively respond to a request, but acts autonomously to achieve a given objective. The limitations today are reliability. Because if a chatbot hallucinates, becomes incoherent or biased, that's okay. But if it's an agent who hallucinates and acts recklessly, that's when it becomes dangerous.

Feel free to send us your comments or suggestions for additions to this free-flowing inventory. In any case, we felt it was essential to go through this stage in order to talk about what comes next.

‍