Dispatches from Inner Space
The Nooner with J.E. Petersen
You can’t autocomplete art
0:00
Current time: 0:00 / Total time: -3:37
-3:37

You can’t autocomplete art

Belated thoughts on the limitations of ChatGPT, etc

This is The Nooner, a short daily (Monday - Saturday) newsletter slash podcast that has its very own section within Dispatches from Inner Space.

Every Sunday, I publish the Dispatches Weekly Digest (DWD), which lets you binge all the Nooners from the previous week. It also includes a meaningful song recommendation, and a short segment I call TMI, where I go off script to bring you backstage, so to speak.

Two more things about the DWD:

  1. It goes on on the main Dispatches channel, so if you’re looking to spare your inbox from the daily emails without missing out on anything, you can specifically unsubscribe from The Nooner section, and still get the Digest on Sunday.

  2. It’s only available to paid subscribers.

The Dispatches Weekly Digest is a labor of love, and I’m really proud of it, and if you want to hear it, I want you to hear it. So, if you can afford it…

And if you can’t, but you still think of yourself as one of my true fans, let me know and we’ll work something out.


No shortcuts

OK I’m not gonna lie in my darker moments I’ve wished I could just get a large language model to generate a bunch of good new posts. Like, I’d just feed it a topic and maybe a one-line take, and it would write out a Nooner, or maybe even a full on essay in my voice with my style and buzzing cleverness and voila.

But what I do is art, and what AI does (still) is basically autocomplete, and you can’t autocomplete art.

Back when ChatGPT was still amazing everyone (we’re pretty much over it now), a friend asked me if I thought Substack would eventually get clogged with a bunch of people spamming newsletters written by LLMs.

At the time, I had recently interacted with someone who tried to do just that. This person went by the name Hannah Williams, had an AI generated profile pic, and published summaries of popular business and self-help books every few days.

The only reason I became aware of this person was that he/she/it had liked and commented on a bunch of my recent comments on other people’s posts. I didn’t think it was possible to follow a specific person’s activity that way, so I reached out to ask if they had discovered some hack I wasn’t aware of.

I was very nice about it, and “Hannah” actually responded in a very human, semi-embarrassed way that (at the time) convinced me there was a real person behind his/her/its account.

Eventually, “Hannah” got booted from the platform for spamming people. But even if that hadn’t happened, I think the scheme would still have been short lived.

LLMs are only good at producing work that has essentially already been done in some way. I’m certainly not the first person to call it a super sophisticated autocomplete, but that’s exactly what it is. Feed it a prompt, and it will spool out the thread of that prompt into an answer that is compiled from the vast library of content it’s been trained on.

I don’t pretend to deeply understand how these models work from an engineering perspective, but I’ve used them often enough to have gotten a sense of what they’re doing. Every time I get an answer from an LLM, it feels very much like a shorthand summary of the Internet. Sometimes, in some contexts, the responses are useful. Sometimes they seem very smart and informed. Sometimes they are shockingly stupid. Sort of like scrolling through comments on Reddit, except way less entertaining.

The models, we’re told, are getting better. But no matter how sophisticated they get, LLMs will never be able to produce meaningful creative work, because, well…

You can’t autocomplete art.

Discussion about this podcast

Dispatches from Inner Space
The Nooner with J.E. Petersen
Dispatches from Inner Space presents: The Nooner - a daily distribution of open-ended ideas.