3 Comments
Mar 22, 2023Liked by Mark Newheiser

Probably should have guessed the conceit earlier - people have been doing these "no actually it was GPT lol" things for a while now - but I'm not used to meaningful length and internal coherence. Not sure how much light editing you did to sand off the usual telltales like repetitive sentence structure or boring word choice. The emoji were definitely a tipoff though - almost no Serious Blogger ever uses those outside of comments sections, unless there's a highly context-dependent reason to do so. LLMs seem to love emoting (and it's extremely annoying, bad Bing). And all throughout I was like "huh, this feels super generic compared to other Mark fiction, must be a softball to meet the once-a-year window*". So it makes sense.

Still gonna be awhile before LLMs can convincingly fake a "creative spark" for me, I think. Too much cruft in the training data. (I shudder to imagine what an LLM trained exclusively on, like, Yudkowsky's corpus would be like.)

*Not for April Fool's, but sometimes one-day-window fiction can be really high-quality: https://slatestarcodex.com/2013/11/03/the-witching-hour/

Expand full comment
author

Full disclosure, this is actually three levels deep of joke and I really did write the entire thing! You're meant to read the first letters of the final section per the footnote, and realize the final reveal is another layer of fakeness. I think I probably should have been more obvious.

Your criticisms of the piece are taken though even if it's affected by hindsight bias, it's mostly simple social media level musings I wrap in a convoluted structure. Glad I at least have a spark elsewhere! This year I think I may try to play guess the ones I wrote vs GPT-4, explicitly asking it to come up with comparable musings

Expand full comment

It was hard to tell whether you meant the final Easter egg seriously, or it was just a fun wordplay gag. (And also now I'm not sure what to believe. How deep does the rabbithole go?)

Conditional on you actually writing the whole thing: congrats, it's a good sign if humans can learn to spoof AI spoofing humans. We're one step ahead in the originality olympics yet. You know, I think someone once wrote a story about that...

I've heard people tend to get pretty good results from "write in the style of X", although that would depend on the relevant corpus actually being in the training data, which I think cuts off some of your later works? Forget when the exact date was. It'll be interesting when skilled artists get to the point of still churning out all-new stuff sometimes, but then use generative AI to pad out necessary filler without breaking continuity. Like, a good show might still have some "Villain of the Week" elements to it, but one doesn't watch for those elements specifically. It's not always better to just handwave away transitory content...and oftentimes fans end up interested in such minutiae after the fact anyway. So much of fanfiction is like this.

Expand full comment