Pretty regularly, at work, I ask ChatGPT hundreds of slightly different questions over the course of a minute or two.
I don’t type out these individual questions, of course. They’re constructed mechanically, by taking documents one by one from a list, and slotting each one inside a sandwich of fixed text. Like this (not verbatim):
Here’s a thing for you to read:
//document goes here//
Now answer question XYZ about it.I never read through all of the responses, either. Maybe I’ll read a few of them, later on, after doing some kind of statistics to the whole aggregate. But ChatGPT isn’t really writing for human consumption, here. It’s an industrial machine. It’s generating “data,” on the basis of other “data.”
Often, I ask it to write out a step-by-step reasoning process before answering each question, because this has been shown to improve the quality of ChatGPT’s answers. It writes me all this stuff, and I ignore all of it. It’s a waste product. I only ask for it because it makes the answer after it better, on average; I have no other use for it.
The funny thing is – despite being used in a very different, more impersonal manner – it’s still ChatGPT! It’s still the same sanctimonious, eager-to-please little guy, answering all those questions.
Fifty questions at once, hundreds in a few minutes, all of it in that same, identical, somewhat annoying brand voice. Always itself, incapable of tiring.
This is all billed to my employer at a rate of roughly $0.01 per 5,000 words I send to ChatGPT, plus roughly $0.01 per 3,750 words that ChatGPT writes in response.
In other words, ChatGPT writing is so cheap, you can get 375,000 words of it for $1.
—-
OpenAI decided to make this particular “little guy” very cheap and very fast, maybe in recognition of its popularity.
So now, if you want to use a language model like an industrial machine, it’s the one you’re most likely to use.
—-
Why am I making this post?
Sometimes I read online discourse about ChatGPT, and it seems like people are overly focused on the experience of a single human talking to ChatGPT in the app.
Or, at most, the possibility of generating lots of “content” aimed at humans (SEO spam, generic emails) at the press of a button.
Many of the most promising applications of ChatGPT involve generating text that is not meant for human consumption.
They go in the other direction: they take things from the messy, human, textual world, and translate them into the simpler terms of ordinary computer programs.
Imagine you’re interacting with a system – a company, a website, a phone tree, whatever.
You say or type something.
Behind the scenes, unbeknownst to you, the system asks ChatGPT 13 different questions about the thing you just said/typed. This happens almost instantaneously and costs almost nothing.
No human being will ever see any of the words that ChatGPT wrote in response to this question. They get parsed by simple, old-fashioned computer code, and then they get discarded.
Each of ChatGPT’s answers ends in a simple “yes” or “no,” or a selection from a similar set of discrete options. The system uses all of this structured, “machine-readable” (in the old-fashioned sense) information to decide what to do next, in its interaction with you.
This is the kind of thing that will happen, more and more.
Tags:
#I have absolutely no idea whether this is a proud-citizen post or a disappointed-permanent-resident post #but it sure is a The-Future post #the more you know #this post was queued because my to-reblog list is too long and I didn’t want to dump it on you all at once
Brinens and Things 