It has been a few weeks since I've used it. I also tried GitHub Copilot but canceled it same-day.<p>For me, trust in ChatGPT as something to "hang my hat on" is more or less shattered. I refuse to become dependent on anything that is this fragile over time. For day-to-day coding, I can almost always go faster if I stay in flow. For the other stuff, I'd rather just go straight to Google/MDN/SO instead of play 20 questions with GPT models and then confirm with google anyways. If I am going to context switch, I might as well switch all the way and make sure I am 100% confident before trying to get back into my IDE. Make a trip to the whiteboard. Take a break. Think about what the hell I am doing conceptually <i>before</i> I start flailing shit around on my filesystem.<p>More broadly, I am completely over using generative AI for things that can just as easily be answered with proper organization of data and SQL queries. When it comes to our business, I will now be pushing for determinism and traceability all the way back to specific tokens, offsets and statistics as found in the training corpus. The LLM chat demos are super compelling to executives <i>at first</i>, but it only takes a few counterexamples to prove that you don't want to run their loan qualification programs on top of a system architected like that. To be clear - this isn't even about ChatGPT vs other LLM options. Fundamentally, I do not believe that the current selection of generative+DNN models are appropriate for most forms of "serious" business.