Some researchers who track the effectiveness of large language model services have made the claim that ChatGPT has become “dumber” despite the more advanced GPT-4 no being available. Now, ChatGPT promoters and these researchers are involved in the controversy of whether this is actually the case. As OpenAI is challenged by more and more challengers to gain market share. This dispute has real consequences.
Dr. Ethan Mollick a Wharton Professor claims that performance when GPT-3.5 and GPT-4 implementations are given the same prompts the more “advanced” version is less successful. Backers claim differences may result from the specific type of prompt the two systems are asked to address.
Others suggest that the perception that the capabilities of this model have declined results from users getting past the “wow” factor and perhaps evaluating products more carefully. I know from my own experiences, I was shocked when I asked ChatGPT to provide citations for the stations made that the service would generate complete APA-looking journal citations which I would discover were completely bogus when I tried to look them up. I no longer would offer content generated in response to general prompts and use present AI tools when I am able to focus the tool on a specific resource (typically a pdf) I can designate.