OpenAI fixes ChatGPT “laziness” problem and lowers access costs to its platform

OpenAI announced that it had fixed the laziness problem of its flagship product, ChatGPT.

In a blog post dated last Thursday, OpenAI announced several important updates to its products. There is thus a significant reduction in the price of some of them, in order to prevent paying users from being too closely interested in models developed by the competition. But above all, the company announces that it has corrected this surprising bug which made ChatGPT lazier than it was in the past.

Has ChatGPT learned the laziness of the developers who trained it?

Among the many minor updates announced in the company’s blog post, only one small paragraph explicitly mentions ChatGPT’s laziness problem, which had been noted since at least early December. And if the artificial intelligence company explains having corrected, or at least reduced the problem, no information as to its origin is given: “ This model performs tasks such as writing code more efficiently, and aims to reduce cases of “laziness”, where the model does not complete a task “.

Once the phenomenon was noticed, one of the hypotheses put forward (by people outside OpenAI) to explain the problem was that the model had also learned this behavior from the humans who had trained it. Which would have explained that it was precisely as the end-of-year holidays were starting that the phenomenon had started to appear. The blog post also announces having brought “ various improvements » to the model, without considering it necessary to say more.

Prices dropping considerably

We will therefore not know more about the origin of this problem, and the very formulation of OpenAI suggests that it is not necessarily definitively resolved. On the other hand, the company announced drastic price reductions: the use of the GPT3.5 Turbo API, which is by far the most used model, will now cost half as much per request. This model is already widely used for countless conversational applications, or for text analysis, thus guaranteeing an even more dominant position in the field.

Because as open source models or those developed by competitors get closer to GPT, OpenAI has every interest in ensuring that its customers do not take too much interest in their development.

Source : TechCrunch, OpenAI