A French-language YouTube monologue arguing that OpenAI's GPT-5 chat release is a regression: the speaker reports instability (forgetting past content, crashing), poorer, shorter answers (termed 'shrinkflation'), and disruption to business integrations. The speaker explains how to revert to older models via settings, suggests OpenAI may have reduced computation per response to save energy at the cost of quality, and raises broader worries about training-data limits and degeneration if models are trained on AI-generated content.
Hello, my friends. Open Air is releasing the GPT5 chat. There was a lot of hype and marketing beforehand. They told us, you're going to talk to someone who has several PhDs and who will help you. The reality is that, as you might say, it's crap. I found this on sale. I don't really know what it is, but I brought it back because it's crap. It helps illustrate my point today. So there really is a catastrophe in Leah's world. And that's because the GPT5 chat is a disaster. Some have even described it as shrinkflation. Do you know what shrinkflation is? It's when, for example, a box of chocolate goes from 100 grams to 90 grams, or a box of butter goes from 150 grams to 225 grams, which means we're going to put less product in it. So in reality, what people have noticed is that he is simply useless at doing things simply. He gives vague, short answers. I myself discovered something. I can give you a test. You can try. I asked him to create an HTML page. So I'm not asking him to do Scala or Assembler or anything like that. We're just going to ask him to do something super simple to do HTML. So it's considered simple code. And normally any teenager with a notepad can write HTML. So I tell him, let's take a pizzeria, for example, whatever. And then you create a page for me, a home page. I give him some information, so he creates the page for me. Then I ask him, can you add the phone number or can you change this element? And then it's a disaster. Sometimes it crashes, uh, sometimes it will just forget what it did. And then, for example, if I tell it to add a phone number, it will create an empty page with just the phone number and the previous page it forgot. And when I correct it, I say, oh, yes, yes, that's right, you're right. Then it will try all that again and it's a mess. So a lot of people are complaining about chat GPT5. First of all, we have companies because what did the companies do? They fired Caroline and Melanie and Mike and all that. They replaced everyone with ChatGPT. So it's a system that doesn't necessarily work on the screen, Right? It's a system that can be controlled by API. So we can have applications that talk to Chat GPT via API, that get responses that will integrate them into a more complicated business process. The problem is that if you want to integrate things and solutions into a business process, it has to be stable. That's the first thing. That's the basics. What OpenAI did was change their model and broke a lot of things. So there are a lot of customers who pay a lot of money who are complaining. There's also a side effect, which is that there are a lot of people who have come forward and said that they Talked to Chad GPT4 for months on end. So for personal stories. So there are people who share their opinions in their Daily lives with ChatGPT because, well, it uses them as a kind of friend. A kind of friend who is there all the time to listen to them, who won't judge them, who will always advise them day and night. They are aware that it is a program, they are aware that it is not a human, but in a way it helps them. And the fact that it spoke to it for months and months, keeping in mind the old conversations allowed ChatGPT to better integrate the context. This means that when you ask it a question today, it fits into the context of conversations that have been ongoing for months. So the solution was able to give them answers that they felt were interesting, relevant and constructive. And now these people suddenly find themselves in the lurch with GPT5, which is extremely limited and seems limited to them. What probably happened is that OpenAI wanted to reduce the energy footprint of the answers, if you like, because the reality is, what is expensive in a ChatGPT answer, it's the computation time. This calculation time will translate into energy consumed by entire datasets that consume a lot of energy to generate these queries. So if we can reduce the response time and the energy spent per query by making shinflation type queries, as I told you, by making shorter, more simplified queries and all that, we can make a lot of savings, but it affects the quality. So what we're doing with GPT is that they've decided to go back to the previous model, so they've allowed it. It's quite an exceptional approach, by the way. I told you at the beginning of the video, I was about to stop my subscription and then they allowed me to go back to the previous model. So it's not easy. You have to go to the settings, you have to go to the bottom left from the computer, it won't work from the phone, you have to go to the bottom left click, go to Preferences, and then there's a button that allows you to activate what they call the old models. And so you activate it there. And that's the only thing that will allow you to access model 30 and 40 and so on, four and all that. So there you have it. Now, beyond the GPT chat of this setback, there's something that we're seeing something perhaps deeper. Isn't Leah reaching a plateau? Does that mean that Leah, which matured very quickly with great promise and we've seen or felt exponential growth, Isn't it reaching a limit today means I've seen very serious studies, studies done by mathematicians, all of which analyze AI on a truly mathematical level. And we realize one important thing. Leah can only function if it is trained by human content generated by a human brain. We tried one thing, we tried to train Leo with AI generated content. So we're going to give it images generated by artificial intelligence, texts generated by artificial intelligence to see what happens. Well, by dint of doing this, and we're going to re inject the output, we're going to re inject it as input and soon enough it completely degenerates. Soon enough we reach total degeneration of artificial intelligence. It is no longer capable of working. So it needs fresh human content. Today, when we take models with billions and billions of data points, these models have practically ingested everything that humans have written. They swallowed millions of doctoral theses, millions of books, millions, millions of press articles in various languages and all that, billions of photos which sometimes, and this is a big problem, things that were generally picked up off the truck, as they say, they didn't ask for permission or anything. They just injected millions of pieces of music, songs, scores, everything they found into the system and they ate it up. Okay, and we have what we have today. What's next? What's next? We're really bothered because there's more data to feed on. There's more than that. So we have this impression that we're potentially heading towards a plateau. Well, we'll tell you, but it's a really big problem. There you have it, my friends. So GPT cat the disaster and we'll follow this story and I'll come back in another video. See you soon.
