AI assistance in writing and programming

On his blog, Alexis Smolensk takes down the argument that using AI ruins your writing, by detailing just how magically useful it is for assisting with your writing -- so long as you bother to use it intelligently.

Quoting selectively from his post:

Just because A.I. wrote a paragraph, that is not equivalent to "exercised authorship over the finished work." That requires my complicity, and I choose not to give it. In no way does this remove the usefulness of the program. [...]

[Alexis describes starting a new ChatGPT conversation with one sentence about a story he'd been working on, and having the AI summarize both the narrative and the authorial intent.]

[I] recognise that would scare the bejeezus out of a lot of people. I see it as a collaborative process. The more the program knows about my projects, and its ability to accurately describe those projects without error before I'm ready to take them up again, the better. [...] The answer simply accurately describes the story thus far. It quotes an important turning point in the story, accurately defines details and presents itself as "ready to collaborate."

Why would I ever not want to employ a tool able to do this?

Like the writer's block, the issue of "re-entry," re-acquainting oneself with work that has been untouched for some variable length of time, can be instantly managed by simply uploading it to the system and talking about it. "Hey, what did you think of…" can produce a wide variety of answers, depending on the built up rapport between the user and the program, any of which can be discarded by saying, "I think you read that wrong… isn't it true that…?"

And getting a different answer.

Does it matter that the answer isn't "intelligent"? No, not at all. I don't expect it to be. It's a reminder of something that I did, that I remember from my point of view, that I can assess quite freely without needing A.I.'s version… yet, like a book mark (which also lacks intelligence), it reminds me of some detail I might have forgotten, specifically, when I wrote this part about this thing that happened. Only, this is a "bookmark" than can read instantly and accurately around the book without difficulty. Five, ten minutes of back and forth and I'm right in the groove again; I remember where my headspace was, I remember what I was trying to accomplish...and I’m ready to just get started. Not because "A.I." did something, but because it's a new way of filtering the work that already exists in a format that humans intuitively understand: conversation.

[...]

I do not use it to write for me. I write, and I use it to reflect on what I've written. Depending on what it says, I may choose to go back and change something, or follow an inspiration I've just gained, or simply feel assured that I'm on the right track, since what I'm saying appears to be, from the reaction, clearly understood. Having spent so much time "writing in the dark," as it were, without any clue to knowing if anything I was saying was hitting it's mark, this is very gratifying. It makes writing a lot more fun than it used to be.

I responded to him on Patreon like this:

The sheer number, the absolute sheer unbelievable mind-melting number, of people I hear bleating about how AI makes you stupid, about how AI can make mistakes, about how AI doesn't actually "know" anything! I just want to shake all those people and say:

By prompting an AI, I can cause it to accomplish a task at more or less my skill level in a tenth of the time I'd need. Then I can spend a further two-tenths or eight-fiftieths of the time I would have needed to be as bloody-minded and pedantic as I like while checking its work. And if it fucks up, I throw away the session and I do it all over again, explaining exactly what didn't work the first time, or even feeding in the subpar results themselves as anti-instructions. I can do all of that, while learning new techniques from how it accomplished the task; and forcing myself to elaborate upon my ideas at length, out loud, with all the benefits thereof; and STILL COME OUT AHEAD ON TIME SPENT.

There is of course the argument that if you have AI do things that you don't at least partly know how to do yourself, then you will (a) not learn much or anything about the task, leaving you just as helpless when confronting the problem in the future (b) be ill-equipped to gauge whether the AI's result is actually fit for purpose. There is a shred of merit in that argument: if you're having it turn out work that you don't know a single fucking thing about, then passing that off as your own, or claiming that you're all ready to tackle the same challenge on your own, then you deserve what's coming to you for treating the AI like a genie. But why the fuck would I care what genie worshipers do?

Here's an example. I'm using agentic AI to upgrade Zingor, the game-character tracking application I've built for Alexis's online campaign, to properly represent sage studies whose study points must be distributed to sub-areas. (For example, Athletics points get allocated to specific sage abilities.) I took a stab at it myself and came up with some constraints to keep in mind. I had an AI try it entirely hands-off, too; its solution sucked, and I wrote up my reasons why into the ticket tracker.

And now, I've spent the last 90 minutes or so working interactively with the AI on several more variations on how to best code up the rather complicated rules for storing and importing data for those studies. I have done research and talked to Alexis to suss out some finer points of the rules, and ensure we reached a complete solution that was "unthinkable" from the AI's limited starting point. I have corrected its course where it was going to do some rather fragile additions to the code base.

I've pushed back on its assumptions, and in some it pushes right back: It just neatly laid out why one of the two final possible candidates for our code module would function as well as the other candidate, but would cause additional tedious, error-prone data entry work for the end users -- for my fellow players. I hadn't caught that wrinkle; it was able to mock up some test data and show me the tradeoffs. When I saw that, I changed my mind. And what is learning but repeatedly changing your mind for the better?