Introduction

The spectacular advancements in terms of understanding and generating language have created a massive enthusiasm for artificial intelligence. Every company is looking for ways to integrate some of these technologies into their products, processes, and services. Governments are competing with announcements. The investments are considerable, at least as much as the number of posts on LinkedIn.

These recent developments are partly due to a technology initially developed and made public by a Google team, the Transformers. OpenAI was the first company to understand and exploit the potential of this neural network architecture and applied it to text generation through its successive GPT models.

Nuclear Fusion Effect

What I call here the nuclear fusion effect is the gap that separates the expectations and announcements regarding nuclear fusion and the reality of nuclear fusion.

Nuclear fusion offers a clean (relatively) way to produce energy in considerable quantities. The concept has existed for decades, and recently, several experimental reactors have achieved and maintained the conditions necessary for a nuclear fusion reaction.

Despite this, we are still far from an industrial use of nuclear fusion.

Parallel with Artificial Intelligence

In the same way, there is a notable difference between the proofs of concept of artificial intelligence and its practical applications in companies.

Let’s take an example to illustrate. Imagine you are a company developing an online marketplace. Products are for sale, and regularly, you send an email to your users to promote certain products. To tout the merits of the promoted products, you need to write a short text for each product. Of course, your company has a stylistic charter that it must respect; it uses certain turns of phrase, humor or not, forbids certain words, etc. But you follow new technologies closely and feel that ChatGPT might be able to help. You give it a chance, describe the product about which you wish to write a paragraph, about 5 sentences long. The first generation is bland, insipid, unusable. You then describe the style in which you want ChatGPT to write the text. The second batch is better but still far from the mark. Cunningly, you decide to give it an example of text that you wrote and ask it to take inspiration from it. And there, the text is very correct. The style is generally respected.

Nevertheless, the product you are promoting is not exactly the one you gave as an example, and you need to make some changes. A word here, a word there. About 5 words, one per sentence then. The problem is that the words you need to adapt are actually the words that represent your style. Moreover, changing a word in a sentence that only has about ten words is not much faster than writing the sentence directly.

I’m not even talking about the fact that this work is barely automatable since if each text requires verification, correction, then review, it might as well be written directly.

Morale

That’s what it’s about. ChatGPT, as powerful as it is for generating pieces of code (and it is really good at that), is in reality closer to the 3-second record of nuclear fusion than to industrial production. Integration into company systems requires a significant effort on the part of the companies, and the results can be disappointing.

I think it would be better to consider the new algorithms and their availability by the companies that develop them as conceptual proofs that certain problems can be solved, rather than as definitive solutions.