On August 7, Openai launched its latest IA model, GPT-5. Many expected at a decisive moment for the company led by Sam Altman, but also for the development of general artificial intelligence. What really happened?
It was supposed to be the announcement that was going to establish the domination of Openai.
After months of rumors, GPT-5 was finally there. Before the launch of the broadcast live, Sam Altman published a screenshot of the film Rogue One, of the Star Wars saga, on which we see the death star looming on the horizon.
Expectations have only increased. From the opening of the direct, Sam Altman, faithful to himself, said: “We think you will like to use GPT-5 much more than any other AI before. It is useful, intelligent, fast and intuitive. GPT-3 was a bit like talking to a high school student: lightnings of genius, but also a lot of annoyance. GPT-4, it was rather like chatting with a university student … With GPT-5, it’s now like talking to a real expert, a doctoral level expert in any field, capable of helping you achieve your goals. »»
But reality has caught up in the staging.
The reactions were dominated by criticism. These were not a few isolated errors: as for the previous versions, GPT-5 made absurd faults, hallucinations, and records poor performance on certain benchmarks. A new automatic “routing” mechanism has been deemed confused. The effect was all the more marked since the expectations were high: while GPT-4 had marked a qualitative jump compared to GPT-3, GPT-5 appeared as a great disappointment.
What makes this reaction different from previous launches?
For GPT-3 and GPT-4, we could still talk about radical advances. GPT-5 is only marginally better than some competitors already on the market, and sometimes even less efficient on tests like the Arc-Agi-2 .
It is not a bad model – but it is not the announced revolution either.
Many expected GPT-5 to be a general AI, or at least approach it. The result: a reality shock.
What are the implications for Openai?
They are considerable.
Openai no longer has a real technical advance, and GPT-5 will probably not remain the most efficient model for a long time. Several of their best talents have left the company, often to found competitors, while actors like Elon Musk, Anthropic or Google are progressing quickly. Relations with Microsoft have become more tense, and the company, far from profitable, was forced to lower its prices.
At the same time, the idea that the LLMs are close to AGE loses credibility, and confidence in the company and its CEO is crumbling.
OPENAI certainly keeps a known name and an appreciated user interface, but will this be enough to support a valuation of several hundred billions of dollars?
Nothing is less certain. Logically, Sam Altman’s reputation should be seriously tainted.
His past statements-claiming to know how to build the AGE or comparing GPT-5 to a universal doctoral level expert-now appear as untenable promises. The contrast between these announcements and the real product capacities largely explains the extent of the disappointment.
How do other models position themselves compared to GPT-5?
The other large models have the same limits.
A particularly enlightening example comes from psychologist Jonathan Shedler, author of one of the most read articles and the most cited on the evaluation of the efficiency of psychotherapy.
When he questioned Grok, on this same article, the result was entirely false: the AI reversed the conclusions by affirming that psychodynamic therapy was less effective than cognitivo-behavioral therapy, while his article precisely demonstrated the opposite.
The effect of psychodynamic therapy in the article was 0.97. Grok claims that it rises to 0.33. However, this figure does not appear anywhere in the article.
This type of error illustrates a structural problem: these systems give the impression of an encyclopedic mastery, but collapse when they are confronted with an area that we really know.
Where does this leave us in relation to the general AI?
The LLMs remain unable to generalize largely when they are confronted with situations that leave the framework of their training data.
A study by the State University of Arizona, published on August 5 has just confirmed this, validating what I have been repeating for almost thirty years, and more recently what Apple exhibited in an article in June .
As early as 1998, I had already shown in an article that multilayer perceptuals (multilayer perceptrons), the ancestors of current language models did not manage to reliably apply linguistic or logical universals outside the field of learned examples. The authors of this study show that this limitation persists today, despite all the innovations provided since.
It is precisely this inability to generalize that explains why all attempts to create a GPT-5 level model, whether they come from Openai or elsewhere, are doomed to failure. It is not a course accident, but a limit in principle. In other words, as long as this structural weakness is not overcome, performance remains blocked, whatever the resources invested.
Does this impasse mark the end of the current approach?
I’m sorry to have to repeat it, but I already told you.
Person with intellectual integrity should believe that the ” pure scaling – Investing more money and computing power in the LLM – will lead us to AG.
After having invested more than $ 500 billion in this direction, the observation is clear: the qualitative limits observed on GPT-3 or GPT-4-persistent hallucinations, errors of reasoning, views in vision, difficulties in performing simple tasks like counting correctly-are found identically in GPT-5, despite marginal gains and a lower cost.
The myth of an imminent act thanks to the only ” scaling “Must be abandoned.
The famous formula according to which “attention is everything we need” therefore turns out to be misleading. The only realistic route to general artificial intelligence goes through neurosymbolic approaches integrating explicit models from the world, capable of reasoning on lasting, abstract or symbolic representations. As long as we do not have this type of systems, we will not be able to cross the qualitative threshold which still separates us from AG.
