Work had one of those “lunch & learn” type things recently on ChatGPT, which I skipped because 1) I’ve been exposed to enough hype on AI/LLM/ML whatever you want to call it, and 2) I’m not a religious man, but lunch downtime is sacred. A couple days later they followed up the lunch session with an email linking to the video of it that more or less read, “Did you see it? Did you see it?” And that email was quickly followed by yet another message with a survey asking for input/interest in our unit using ChatGPT in some way.
So like a tired parent with a hyperactive child obsessed with the latest craze, I put aside what I was doing to engage. The video was fine — an overview of what the various acronyms mean, how the technology works, and a couple demonstrations. The latter consisted of asking ChatGPT to generate Python code for a quicksort algorithm, then asking it to do the same in Rust. They then asked for an explanation of the same in the voice of Jean-Paul Sartre, Ayn Rand, and Kanye West. It did manage to do all of this fine. There was some discussion about how these LLM tools have a tendency to just make things up, as well as copyright issues, but it was only a 30 minute session so it didn’t get too deep.
On to the survey form, which was basically two yes/no questions and a pair of text fields that looked like they would accept maybe 50-100 characters. That may have been enough to write, “ChatGPT and generative AI are bullshit and I want no part of it,” (63 characters) but swearing is kinda frowned upon at work, so I just went on with my day.
But it’s possible I’ve been stewing over it ever since, because ChatGPT and generative AI are bullshit, I want no part of either, and yet they keep getting shoved in my face like the greatest thing since the invention of undo. I suppose I should pause for a moment to acknowledge that there are reasonable uses for machine learning, but I would argue they are limited and best leveraged under the hood. For example, Pixelmator Pro has image tools based on machine learning that are genuinely useful. They are also very different from how ChatGPT and its ilk are sold.
Let me blow through some basic things first. Yes, tools like ChatGPT can spit out code for self-contained, well-defined problems like the above. I would argue that this ability is no more remarkable than being able to punch “quicksort algorithm python” in a search engine, which will give you multiple results with working code and/or explanations. A web search can’t give you the results in the style of, let’s say, Dr. Seuss, but that’s just a parlor trick and not very practical anyway. It can also give you answers to non-technical questions. Many of these answers might even be right! Either way, you can get paragraphs of authoritative-looking text out of it. Tools like Stable Diffusion can even generate images that look vaguely like what you’re asking for as long as “surreal” is part of the mandate. You would be a fool to rely on it for professional work, though (<cough> Secret Invasion)
Let’s take a recent example of something I had to do for work. One of the apps I’m responsible for has a feature that allows users to send a bulk email to pre-defined sets of recipients. There’s a mail merge function as well, so these messages can be customized. This basic system has been in place for roughly a decade now. This summer, I was tasked with altering this to allow a user to send a message to a group of users, and this message was to contain a hyperlink with a draft email to a different set of recipients. This called for a mail merge within a mail merge, complicated by the fact that one of these had to generate encoded content. The source code for this app is not publicly available, and the frameworks it relies on are not widely used, where they’re not completely custom. This is not a problem ChatGPT is going to be able to help with, any more than I can hit Google for a solution.
Let’s also consider the tendency for tools like ChatGPT to just make things up. I would argue this is not just one of the tool’s greatest weaknesses, but also their greatest danger. If you’re not familiar with the lawyer who used ChatGPT as a research tool to his regret, you should read up on it. In the lunch session, someone did bring this up, and they suggested you should basically ask the tool, “are you sure?” But can you trust that answer any more than the first? The problem is, ChatGPT will sound absolutely confident in its responses, whether there is any truth to them or not. And you can’t really ask for citations without having to then follow up to verify those are real, too. Here again, I would argue you’re better off with a more traditional tool like Google that directs you to source materials instead of synthesizing content and spitting it out context-free.
What really gets me about all this isn’t just that the tools are more limited than the hype would have you think, and that they can be dangerous if you’re not careful. For the limited results we get from them, neither the infrastructure required to make them work nor the changes they would in turn demand from users are sustainable. Remember, in order for something like ChatGPT to work, it not only has to have vacuumed up massive amounts of content, it has to have humans annotating the data in order for the computer to do anything intelligent with it. Further, this data intake and annotation has to be ongoing, because spoiler alert, the world is a dynamic place. But surprise, annotation doesn’t pay well. So what’s already happening? People are using AI to help train AI. But if AI can’t be trusted to get things right…it’s like trying to build a skyscraper by taking girders from the basement to support the 20th floor.
And on the other end, in order to use ChatGPT well, you have to be skilled in how to prompt it, to the extent there are university courses on prompt engineering.
And I’m not even getting into copyright or the ethics/legality of essentially hoovering up the entire Internet without any compensation or acknowledgement, let alonefair compensation to the original creators. Last I heard, OpenAI won’t even disclose where its data set comes from.
So yeah, I don’t think these tools are going to stick around in any kind of generalized form, and I’ll be happy when they go back to being the province of specialized tools and research. Hopefully they’re already on their way. In the meantime, I’ll just stick to what I’m doing, thanks.