In the past few hours, OpenAI has enhanced ChatGPT with a new feature: custom instructions. The latter, unfortunately, are not arriving in Europe, while it seems that some users and researchers have started to criticize OpenAI for worsening ChatGPTcalling its latest iteration “slower and dumber” than its predecessor.
In the past few hours, during an interview with Business InsiderOpenAI has confirmed that GPT-4 may have ‘missed a few shots’ in some specifications tasks after its latest updates. For the moment, it is not clear what caused the problem, while a paper Published on ArXiv by some researchers at Stanford and Berkeley explains that the new language model of ChatGPT would fit “degrading over time“.
In fact, researchers at Stanford and the University of Berkeley found that both GPT-4 and GPT-3.5 they provide different answers than they did a few months ago, due to a process of continuous change linked to the learning system of the two linguistic models and to the updates gradually implemented by OpenAI. However, while GPT-3.5 seems to be improving, the same cannot be said for its “big brother” GPT-4, which would actually start to provide incorrect information in greater quantities and to refuse to answer users’ questions.
Even with a specification queries of questions, the researchers found that the number of correct answers is dropped from 100% to 2.4% in recent months. At the same time, the most recent versions of ChatGPT would be less inclined to provide articulate answers, preferring instead more concise and succinct sentences.
Beyond the academic results, there are many users who in recent months have begun to wonder if ChatGPT has actually gotten worse over time or if theirs was just an impression. The latest statements from OpenAI seem to have dispelled any doubts about it: there has been a deterioration, and there startups is working to fix it.
In the last release of his LLM, for example, OpenAI has improved the GPT-4 API, in hopes of reversing the model degradation. At the same time, however, the company has opened a user portal dedicated to reporting breakdowns and malfunctions.