(2022-06-06) Sloan Notes On A Genre

Robin Sloan: Notes on a genre. The thing to know about the AI language models, OpenAI’s GPT-3 and its cousins, is that they are fundamentally bullshitters

there’s nothing at the core.

“The liar cares about the truth and attempts to hide it; the bullshitter doesn’t care if what they say is true or false, but cares only whether the listener is persuaded.”

Nearly every artist is a genre artist. Form precedes content, even when the artist insists otherwise.

AI art recalls the early days of synthesizers, perhaps; what was Switched-On Bach if not “I see what you did there”?

So it becomes a normative question: what choices could bend AI’s path towards the docility and diversity of the synth, rather than something more destructive and/or centralized?

I hope to see more AI artists find a technique or model they love and stick with it, even as the state of the art advances and leaves them behind

I’m proud of the fact that I’ve spent two solid years using OpenAI’s Jukebox, the same model, never amended or upgraded

Jukebox has remained interesting and motivating to me in a way that none of the language or image models have, and I think that’s partially because it’s so much more difficult to use... Using Jukebox feels like exploring the flooded ruins of a once-great city: exactly that slow, exactly that treacherous, exactly that enticing... In all my work with language models since 2016, nothing has approached this feeling, and I think I am ready to close the loop, finally: I don’t believe AI tools are useful for serious writers.

In retrospect, I can see more clearly that it was the language models themselves that were captivating to me, rather than the words they produced.