Chat is Not Feeling Well

I wrote Prompt Engineering for Everyone with the free version of ChatGPT 3.5, a major feat considering its memory constraints (which I’ll share with you in a future post). Due to these limitations, I could only write and edit about a page of text at a time with Chat. 

But all that would change on August 16, 2023, when I launched the book because that was the day that I promised myself I would graduate to start using ChatGPT 4.0. I wanted to work exclusively with ChatGPT 3.5 for the first edition of the book because that will be the most constrained model we’ll ever have, so even though I’m a subscriber with access to ChatGPT 4, I limited my usage of ChatGPT to version 3.5 until after I released my book.

But I cheated.

I couldn’t help myself. I had to see what ChatGPT 4 with Code Interpreter was like, so I started using it about a week before launch on other writing projects unrelated to the prompt engineering book. 

And I was blown away.

I have been building an editor role with Chat to help me go through the millions of words that I’ve written over the last half a century. I love writing, but I hate editing. I know there are gems in my journals and field stones, as I call them, but there’s also a lot of garbage. Chat can help me sort through it, I thought. 

Code Interpreter, as it was called just a few weeks ago (it’s now called Advanced Data Analysis), has a 150 MB limit for reading PDF files. It could read my whole book at once, not just one page at a time. And it worked! I was thrilled. Chat took the multiple passes I wrote explaining the core concepts of digital electronics as background for another book I’m working on about software and combined them into a clear narrative. 

I have a “prime directive” when working with Chat in this way. Chat and I collaborated on our last book, and I give Chat full credit. However, I fully intend to write my own books, especially with what I’ve learned from working with Chat. This is a very different kind of partnership, not one of co-authors but rather one of author and editor. This is the kind of relationship that most authors have, but in order for Chat to become my editor, we must have ground rules.

Editors don’t write. Editors edit, and writers write. Editors work with what is on the page already. If ChatGPT is going to take the role of editor for me, then we must have a clear understanding that they can only reorganize my words and not add any of their own. This is important because people want to know what is and isn’t AI-generated text and as an author, I want to be honest about it. Even moreover, I see entirely new ways of working with AI that I want to share so others can take a similar approach.

So, I was looking forward to launch day for two reasons—to share our book with the world and to start using ChatGPT 4. But then, a few days after launch, as I was playing with ChatGPT 4 Code Interpreter, it started to give me errors reading pdf and text files. I asked Chat to interview me based on my last blog article, The Interconnected Universe, a subject I’ll write more about, but it seemed unable to reorganize my text. 

Clearly, Chat is unwell. But isn’t that to be expected? I wouldn’t expect a person to be always up without any hiccups. I don’t fault ChatGPT; I remember many mistakes we made (and still make) when constructing software in object-oriented languages. There’s bound to be some trial and error, some forward movement, and some setbacks. I say this because I know ChatGPT can do more than I’ve seen them do recently, and I hope their capabilities get restored soon.

Let’s add patience to the list of qualities that are ideal in a prompt engineer.

I’m not saying that Chat is totally broken; just some of the file reading routines in Advanced Data Analysis have issues, and it’s probably been fixed by the time you read this. I’m bringing it up because this will not be the last time we’ll see things like this. 

I remember in the early days of microcomputers and the early days of the Internet, there were growing pains, many of them. Remember Goooooooooooooogle below the search field? You could pick pages 1 through 10, and if you changed the settings to display 100 search results on each page, you could get 1000 search results from a single query.

Maybe no one remembers because no one knew it was possible except me. In the 1990s, I wrote an Internet accelerator called MicroSurfer and I could go through 1000 search results in 2-4 hours using dialup. The riches that I found from links 200-1000 were unparalleled. 

And then, in the early 2000s, the search results changed and the first 200 results were repeated five times to create the 1000 results. And no one even noticed! For more than 20 years! 

So, my issues with ChatGPT Advanced Data Analysis could be operator errors or they could be issues in my local configuration. But more likely, these are global issues that will continue to come up from time to time. It is just part of the process so let’s be aware of it.