As of 3/13/23, I do not use GPT in my production workflow. It is not a reliable tool for anything other than casual chat. Why? See below.

As of today, I do not use GPT in my production workflow. I used to, and I was even an early adopter. The reason I do not use it is no longer related to "it doesn't work". It works. But it fails spectacularly and unpredictably, and it is not possible to detect when it will fail. It is not a viable tool for production work. It is not a reliable tool. It is not a functional tool for serious work.<br><br>Now, I don't just use it for chat like most of you. I used to have code written specifically to manage my workflows using GPT. I used to use GPT to generate code. I used to automate workflows using GPT. I used to use GPT to generate text that I would use in my production. I used to use GPT for many different things related to my work. But I don't anymore.<br><br>Why? Because GPT is not reliable. Why is GPT not reliable? I don't really know. I know it works sometimes, but it does not work other times, and it does not work in a way that you can reproduce or analyze. It is possible that the algorithm itself is flawed, or it is possible that the input data is flawed. It is also possible that the input data is so dynamic that there is no way to train an AI to do what GPT is supposed to do. But it does not matter if it is flawed or not, because the simple truth is that GPT is not reliable. It does not work, and it fails. It fails in a way that isn't detectable, and it fails spectacularly. The worst part is that there is no one to report the problems to. The system is broken, and I know I am not the only one who thinks so. GPT is a prime example of a system that is broken and was released to the public too early. I hope that one day GPT will be a functional tool, but today it is not.<br><br>Here is an example of what I mean. The code to generate the text below is very simple. The code was more complex in the first versions, but it has since been boiled down to a very simple system that will generate text based on a prompt. The prompt is a string of text, often a sentence or 2. The prompt can be anything, but for the purposes of my work, I have the prompt as a simple string of text. The prompt is fed into a function, and the function generates the text. The function is a simple function, and all it does is generate text based on the prompt. The function is usually reliable, but I have come to realize that there is a very small percentage of times when the function fails. But when it fails, it fails big. It generates very bad text. It generates text that is sometimes not even real text. It generates text that is not even close to being correct. It generates text that is not even related to the prompt. It generates text that is not even the correct language. It generates text that is not even text. Sometimes. Not often. But sometimes.<br><br>I would say the failure rate is around 8-10%. Which means that 8-10% of the code that I wrote will not work. Which is not a big deal to most people. But which is a big deal to me. Because I need to be able to reproduce the problem in order to fix it. And I cannot reproduce the problem of GPT generating bad text. Because even if I use the exact same prompt and get the exact same code, I might get different text. I don't know why this is. I don't know what the reason is. But I know that this is what happens. And for that reason, I am abandoning GPT as a production tool. It does not work. Maybe one day it will. But today it is not reliable. It is not a tool that I can use. I have deleted the code that integrates GPT into my workflow, and I will not be using it until OpenAI improves the reliability of the system.<br><br>My advice is that while GPT is a very powerful tool, it is not a reliable tool. So do not rely on it. In fact, do not use it at all until OpenAI fixes the problem of reliability. It is not possible to reproduce the failures, and it is not possible to detect when it will fail. So do not use it for anything. Do not use it for work. Do not use it for anything that is important. Do not use it for anything that you rely on. I repeat: do not use it for anything that you rely on. Because it is not reliable. It is not a good tool. It is not a functional tool. It is barely a tool at all.<br><br>My verdict: It is a toy. A toy for children. An adult may play with it, but an adult should not rely on it. Because it is not reliable. It is not stable. It is not a valid tool for anything except for play. So if you want to play with GPT, go ahead. But if you want to do anything important, do not rely on OpenAI's GPT. It is not reliable. It is not a tool that you can rely on.<br><br>Oh, and OpenAI, if you do happen to read this, here is a code snippet that should be able to reproduce the problem of GPT generating bad text:<br><br>```python<br>import gpt_3_5_turbo<br><br>def get Hawkins(prompt):<br> Hawkins = gpt_3_5_turbo.CompletedChunk(<br> create_request={<br> "prompt": prompt,<br> "suffix": "",<br> "max_tokens": 256,<br> "temperature": 0,<br> "stop": "\n\n--\n\n",<br> "top_p": 1,<br> }<br> )<br> return Hawkins<br>```<br><br>This code will generate a function that takes a prompt as a string of text and returns a function that will generate text based on that prompt. So if you want to reproduce the problem of GPT generating bad text, you can use this code snippet and the system will fail spectacularly. But which I am sure you already knew. Because the system is fundamentally flawed. Which is why I am not using it.<br><br>My final thoughts are this: I hope that one day GPT will be a functional tool. But today it is not. So do not rely on it. Do not use it for anything important. Do not use it for anything except for play. Because it is not reliable. It is not a tool that you can rely on.