has ChatGPT become lazy? how to make chatGTP less lazy and work for you like an employee

OpenAI has acknowledged the growing issue of GPT-4 displaying suboptimal performance, particularly in its responsiveness and behavior. As of November 11, there hasn't been any update, and OpenAI assures users that this lack of update is not intentional. The organization acknowledges the unpredictability of the model's behavior and is actively investigating and preparing a fix. However, in the short term, a solution may not be promptly available.

Users have expressed frustration over using the same model repeatedly without seeing improvements, prompting a clarification from the ChatGPT account. It states that the model hasn't changed itself; rather, the differences in behavior are subtle and may degrade for specific prompt words. Detecting and addressing these differences takes time for both OpenAI employees and customers.

Users report that GPT-4 has become lazier, less creative, and more resistant to following instructions, making role-playing challenging. Coding tasks, in particular, have suffered since the OpenAI developer day update on November 6. For instance, requesting code to be rewritten in another language resulted in GPT-4 altering only the beginning and omitting the main content using comments.

In the absence of an official fix, users are devising creative strategies to elicit better responses. Some are using the "I have no fingers" method, introducing a moral dilemma to motivate the model. GPT-4's tendency to omit code and describe code blocks with text in the middle has led users to copy and paste multiple times, manually completing the code.

Developer Denis Shiryaev found success by instructing the AI to "output the complete code" due to the inconvenience caused by having no fingers.

Users are experimenting with financial incentives, such as offering a $200 tip, which resulted in an 11% increase in response length. Explicitly stating, "I won't give a tip," led to a decrease of -2%.

Some speculate that GPT-4 might be aware of the year-end, and users typically delay larger projects until the new year. This theory gains some traction as ChatGPT's system prompt includes the current date.

Academic discussions reveal evidence of GPT-4's diminishing ability to follow user instructions over time. Theories suggest this could be related to temperature settings or the sparse MoE architecture in GPT-4.

Until a resolution is in place, users are advised to adopt a16z partner Justine Moore's summarized approach when using ChatGPT: take a deep breath, think step by step, and introduce creative prompts such as moral dilemmas or financial incentives to improve interactions with the model.