In recent weeks, a notable shift in the behavior of OpenAI’s ChatGPT-4, accessible to ChatGPT-Plus subscribers, has sparked both concern and curiosity. This advanced AI model, known for its robust performance, has been delivering unusually terse and incomplete responses, diverging from its typical, efficient output.
Initially dismissed as “laziness” in the AI, this phenomenon quickly escalated into a blend of amusement and concern among users. Strikingly, the AI often responds to further inquiries with evasive or apologetic remarks, a behavior quite uncharacteristic of a machine intelligence.
OpenAI has acknowledged these observations. “We have noticed your feedback regarding GPT-4’s increasing sluggishness,” the company stated, as 1E9 cited. OpenAI developer Will Depue also admitted that the model’s obstinate and sluggish behavior is a real yet inexplicable issue. The model was last updated on November 11, and some users believe the issues began even before this update.
In an intriguing approach to this problem, some users have attempted to “motivate” ChatGPT through creative means. This includes offering tips or emphasizing the importance of a request for their career. Remarkably, such strategies seem to enhance the AI model’s performance. Other users are experimenting with “Custom Instructions” to deceive ChatGPT with a false date, hoping to resolve the issues.
These incidents have led to a fascinating hypothesis: Could GPT-4 actually be suffering from a “digital winter blues”? This theory is based on the idea that the model emulates human behavior, including the seasonal lethargy many experience towards the year’s end. Some users point out that ChatGPT itself describes the winter period as less productive.
Developer Rob Lynch tested this hypothesis by setting two different months as fixed dates in the AI model, giving ChatGPT standardized queries, and comparing the responses. Indeed, the model produced noticeably shorter responses for December than for May. However, other researchers like Ian Arawjo didn’t find any significant anomalies in their tests.
Despite these unresolved mysteries, studies have shown that language models respond positively to emotional stimuli such as encouragement or pressure, leading to better content generation. As reported by ArsTechnica, this might mean that users are currently compelled to “motivate” ChatGPT with a mix of encouragement and pressure.
The situation surrounding GPT-4 remains a fascinating phenomenon. It raises questions that delve deep into the dynamics of human-AI interaction, going beyond mere technical glitches. Meanwhile, the community of tech enthusiasts and AI researchers remains keenly observant of further developments and insights in this unexpectedly human chapter of artificial intelligence.
Post picture created with DALL-E3
