New Study Suggests ChatGPT Is Getting Dumber—But, Is It?

A new study released on Tuesday by researchers from Stanford University and University of California, Berkeley has ignited debate within the AI community regarding the performance of OpenAI’s GPT-4 language model. 

The paper, titled “How Is ChatGPT’s Behavior Changing over Time?” and published on arXiv by Lingjiao Chen, Matei Zaharia, and James Zou, investigates changes in GPT-4’s outputs over a span of a few months, suggesting a potential decline in coding and compositional task abilities.

The study utilized API access to test GPT-3.5 and GPT-4 versions from March and June 2023 on various tasks, including math problem-solving, answering sensitive questions, code generation, and visual reasoning. 

Notably, the research found a significant drop in GPT-4’s ability to identify prime numbers, plummeting from an accuracy of 97.6 percent in March to just 2.4 percent in June. Surprisingly, GPT-3.5 displayed improved performance during the same period.

GPT-4 is getting worse over time, not better.

Many people have reported noticing a significant degradation in the quality of the model responses, but so far, it was all anecdotal.

But now we know.

At least one study shows how the June version of GPT-4 is objectively worse than… pic.twitter.com/whhELYY6M4

— Santiago (@svpino) July 19, 2023

This investigation comes amidst growing concerns expressed by users who have observed a subjective decline in GPT-4’s performance over the past few months. Speculations about the reasons behind this decline abound, including OpenAI’s possible distillation of models to enhance efficiency, fine-tuning to mitigate harmful outputs, and unfounded conspiracy theories suggesting a reduction in GPT-4’s coding capabilities to promote GitHub Copilot usage.

OpenAI has consistently denied the alleged decrease in GPT-4’s capabilities. Peter Welinder, OpenAI’s VP of Product, recently took to Twitter to counter the claims, asserting that each new version of the AI language model is more advanced than its predecessor. He posits that intensive usage may lead to heightened awareness of these perceived issues.

No, we haven't made GPT-4 dumber. Quite the opposite: we make each new version smarter than the previous one.

Current hypothesis: When you use it more heavily, you start noticing issues you didn't see before.

— Peter Welinder (@npew) July 13, 2023

But the company, unlike its large language model, isn’t completely closed to the possibility. Logan Kilpatrick, OpenAI’s head of developer relations, confirmed on Twitter that the team is aware of the reported regressions and is actively investigating the matter.

Thanks for taking the time to do this research! The team is aware of the reported regressions and looking into it.

Side note: it would be cool for research like this to have a public OpenAI eval set. That way, as new models come online, we can test against these known…

— Logan.GPT (@OfficialLoganK) July 19, 2023

Arvind Narayanan, a computer science professor at Princeton, also pointed out some problems in the study, arguing that the research’s findings do not definitively prove a decline in GPT-4’s performance and could align with fine-tuning adjustments made by OpenAI. 

Code generation: the change they report is that the newer GPT-4 adds non-code text to its output. They don't evaluate the correctness of the code (strange). They merely check if the code is directly executable. So the newer model's attempt to be more helpful counted against it.

— Arvind Narayanan (@random_walker) July 19, 2023

---

Information for this story was found via Twitter, Gizmodo, Ars Technica, and the sources and companies mentioned. The author has no securities or affiliations related to the organizations discussed. Not a recommendation to buy or sell. Always do additional research and consult a professional before purchasing a security. The author holds no licenses.