Attention conservation notice: incredibly rough scratchpad of ideas

"You are GPT-3", revised: A long-form GPT-3 prompt for assisted question-answering with accurate arithmetic, string operations, and Wikipedia lookup. Generated IPython commands (in green) are pasted into IPython and output is pasted back into the prompt (no green). pic.twitter.com/CFVkufPjhf

— Riley Goodside (@goodside) October 17, 2022

there is a meta business to {interior}.ai; exterior of building, landscaping, car interiors, fabric / clothing. Any object that can be highly customized there is a SaaS for them around SD. https://t.co/hpByA7Kdhm

— nikete (@nikete) October 23, 2022

Prompt engineering can be loosely thought of as providing context that shapes the interface between a model and a task.

Prompt engineering may be a transient phenomenon, similar to “advanced/cyborg/centaur” chess.

A market structure for improvements to the interface between models and the world is still generically valuable, even if human-crafted prompts disappear.

Open questions:

• How to compose prompts? Are there useful meta-prompts? How stable are prompts across models? What does that correlate with?
• What market structures are expressive enough? Can be understood? Can be incentive compatible?