To counter ChatGPT’s perceived inefficiency, users are turning to various prompt engineering methods in an attempt to elicit more detailed responses. OpenAI recently acknowledged that ChatGPT, specifically GPT-4, may be getting lazier and failing to meet users’ expectations for longer and more robust answers.
The issue first gained attention on November 28 when Owen Campbell-Moore responded to a frustrated user on Twitter, acknowledging the bug and expressing his own frustrations. OpenAI’s official ChatGPT account also addressed the matter on December 8, indicating that there had been significant feedback on the issue.
Complaints about ChatGPT’s performance seemed to escalate after the OpenAI DevDay conference, where the GPT-4 Turbo model was unveiled. While the Turbo model boasts increased speed and enhanced contextual understanding, these improvements may have unintended consequences for response quality.
Another possible factor influencing the model’s performance is the ongoing restrictions imposed by OpenAI. By limiting the model’s capabilities, the hope is to prevent it from generating raw and unfiltered responses. However, these constraints may contribute to the perceived laziness of ChatGPT.
In response to these concerns, users have come up with creative methods to overcome ChatGPT’s purported laziness. Posts on platforms like Reddit and Discord have surfaced, showcasing various ideas aimed at eliciting longer and better responses from the model.
One approach involves adding emotional cues or hypothetical incentives to coax ChatGPT into generating more compassionate or detailed responses. Although it is important to note that the model itself does not possess emotions, it relies on context from its training data to decipher implications and generate appropriate responses.
For example, appending a prompt with a statement like I have no fingers might result in a lengthier response since a person without fingers would not be able to write. The model can understand this context based on its training data and provide a relevant answer.
One Twitter user, under the handle @voooooogel, devised a clever method to tip ChatGPT for longer responses and even provided evaluations to demonstrate its efficacy. The success of such creative prompt engineering tricks has also been substantiated in the Large Language Models as Optimizers paper.
Earlier in November, an article discussing the Large Language Models Understand and Can Be Enhanced by Emotional Stimuli paper highlighted an EmotionPrompts framework. This framework proposes leveraging emotional prompts to improve the performance and accuracy of large language models, resulting in noticeable benefits across various tasks.
The study found that incorporating emotional prompts led to an 8.00% relative performance improvement in Instruction Induction tasks and a remarkable 115% improvement in BIG-Bench tasks. Additionally, a human study involving 106 participants demonstrated an average improvement of 10.9% in performance, truthfulness, and responsibility metrics when emotional prompts were employed.
The enhanced effectiveness of emotional prompts stems from their ability to strengthen the representation of the original prompts within the large language models, making positive words more impactful and influencing the models to assign greater significance to these emotional cues.
However, the specific reasons for why emotional prompts work remain undisclosed, as only OpenAI possesses that knowledge. Nevertheless, it is worth noting that ChatGPT Plus recently experienced a resource usage overload, causing OpenAI to temporarily pause new sign-ups to ensure a seamless user experience.
While OpenAI is actively addressing these issues, there is hope for improvement in the future. Users continue to explore innovative ways to navigate ChatGPT’s perceived laziness, prompting OpenAI to adapt and refine their models to meet user expectations.
As OpenAI works towards resolving these challenges, it remains to be seen how prompt engineering techniques, emotional cues, and other strategies will continue to shape the interactions between humans and AI language models. With the aim of fostering improved user experiences, the ChatGPT community and OpenAI are undoubtedly paving the way for more nuanced and sophisticated interactions in the digital realm.
