I still don’t know what to make of OpenAI and all of its implementations. I read the articles about how it’s going to change the world, put people out of jobs, and render whole industries obsolete.
The problem is I know too much. That’s not a brag. It’s a realization, as a programmer, I came to when I really began to think how this thing, this entity, works under the covers.
The hype is that it’s magic. Like something out of a far-off science fiction story. Like it’s sentient. It’s not. In reality, it’s a really, really fast parser. Like, really fast. Think Johnny 5 from 1980’s Short Circuit movies.
It works based on info it has “learned”. And it’s been exposed to lots of info from all over the Internet. So when you ask it for a random fact, the probability it can answer you correctly is going to be high. Why? Because we have things like Google and Wikipedia and various other information repositories that are already serving this purpose, albeit somewhat manually.
It can write summaries because it knows how to parse data of N length and summarize. It’s learned that. (This is one place where I would be worried about my copy editing job.)
But what can it really do that’s novel? Can it write you a new and unique story? Yes. And no. The example I’ve heard several times is asking it to write a song. When it composes that song and you hear it for the first time, you’re amazed. Because it seems like it did it on its own. When in reality, it’s learned the logistics of putting poetry to melody. Ask it to write another song? It’ll be very much like the first. Maybe not the exact words, but the tune will be somewhat familiar. Ask it to do it 5 more times in a row and you will see (or hear) the pattern.
It’s not new. It’s just variations on things it has learned. In that regard, it lacks imagination. Something I very much associate with human thought and ingenuity. I don’t think it’s gotten there yet. Not saying it can’t, but I don’t think we are as far along as the hype makes it seem.
I was discussing this with my DevOps team this morning in our weekly meeting. I brought up the idea of having an instance of this entity ”learn” the company’s data and help us with solving business problems.
My primary example was: “Hey (dingus), find me the most qualified and best installer to assign to this job.” (This is something we’ve actually hand-written already. I’d be very curious to see the results of something like this side by side.)
Could it even do that? How it would know, simply by looking at the extremely normalized data structure, what all of the real world entities mean? What is an installer? What’s a customer? What is a job? That lingo is so…human. And there is no Romanoff-to-English dictionary sitting somewhere in any database or web page.
This gets to my biggest question about all of this. How does it learn in order to make decisions based on data it might not have access to (at least publicly)? Are there “private” instances of it? Are there plugins for source material that it needs to know about?
These are things I don’t know. And quite honestly, I haven’t gone down that rabbit hole to find out. Yet.
The Whisper implementation (for transcript generation based on audio) is perfect. It takes input data, parses it, and generates output data. It’s been taught that by the same models that say this sound, in this langauage, translates to this letter or word (in said language).
ChatGPT. Same situation. You ask it a question based on structure of what a question is and it will parse it for key words and data points, look those things up, and re-formulate it into conversational or other structural output.
Are those things impressive? Most certainly!
Are they helpful? Without a doubt.
But let’s not confuse it with magic. Let’s call them what they really are: algorithms. Really large and fast algorithms. And people are still needed to write these algorithms.
For now.
