OpenAI admits the miss: why the new ChatGPT got dumber

Picture this: you update the app that has become your main work tool, and suddenly it seems to have gotten dumber. It makes mistakes in simple calculations, is too lazy to write complete code, and gives canned replies where it used to be flexible. For a long time, OpenAI fans chalked this up to the placebo effect or tech fatigue, but now the cards are on the table. Sam Altman has officially acknowledged that the company made the new version of ChatGPT worse than the previous one. His line, "I think we just screwed it up," is a rare moment of candor in a world where every bug is usually called a feature.

Let's recall the context. When GPT-4 came out, it felt like magic. But with each subsequent update, from Turbo to 4o, the developer community complained more and more loudly about degradation. The model began to suffer from "neural laziness." Instead of writing a hundred lines of code, it produced ten and suggested the user finish the rest themselves. OpenAI was trying to cut compute costs and make answers faster, but in the process it seems to have thrown out the baby with the bathwater. Speed went up, and intelligence dropped.

Why now? The answer lies in enormous market pressure. OpenAI is no longer the only player on the field. When Anthropic's Claude 3.5 Sonnet began to outperform GPT-4o in coding and creativity tests, panic clearly set in at Altman's headquarters. Updates that are supposed to be "safer" and "faster" get rushed out, but AI training algorithms are finicky. Excessive censorship and attempts to squeeze a model into rigid ethical boundaries often make a neural network overcautious, and it ends up delivering mediocre results.

Technically, this is explained by a skew in RLHF, reinforcement learning from human feedback. If human trainers reward brevity, the model becomes laconic to the point of absurdity. If they demand safety, the model starts seeing a threat even in a harmless question about how to cook soup. OpenAI, by all appearances, tightened the screws too far in an attempt to make the product mass-market and corporate-safe. As a result, professionals who needed a powerful tool for work felt cheated.

It is interesting that Altman is making this admission right now. It may be a subtle move ahead of an announcement of something truly big, such as GPT-5 or the Strawberry model. Admitting past mistakes is the best way to prepare the ground for selling a new "revolutionary" solution. As if to say: yes, the last version was a flop, but look at how we fixed everything in the new one. It is a classic audience warm-up that we have seen many times from Apple and other tech giants.

For the industry, this is an important lesson. It turns out the scaling law does not work linearly. You cannot simply pour in more data and compute to get a smarter system. Data quality and fine-tuning precision matter far more. While OpenAI engages in self-flagellation, its competitors have a great chance to lure away its loyal audience. After all, in the AI world, brand loyalty lasts exactly until another browser tab starts thinking faster.

The bottom line: OpenAI has officially confirmed the regression of its models. Can the company win back professionals' trust by releasing GPT-5, or is the era of Altman's dominance coming to an end?