LLMs - Imitation vs Innovation
A huge part of a knowledge worker’s life is summarization, aka compression. You take an impossible amount of complicated information and distill the “important” parts.
Dan Shipper notes that ChatGPT lets us automate the “compression” part:
Read Alison Gopnik on LLMs and children: https://journals.sagepub.com/doi/epub/10.1177/17456916231201401
Yiu, Kosoy, and Gopnik (2023)
Yiu, E., Kosoy, E., & Gopnik, A. (2023). Transmission Versus Truth, Imitation Versus Innovation: What Children Can Do That Large Language and Language-and-Vision Models Cannot (Yet). Perspectives on Psychological Science, 17456916231201401. https://doi.org/10.1177/17456916231201401
“cultural evolution depends on a fine balance between imitation and innovation.”
Gopnik: contrast the compression aspect of the knowledge worker “workflow” with the “perception and action” pieces.
Distinguish between “transmission/coordination” function of language and “cause inference/theory formation”.
LLMs don’t distinguish between epistemic (limitations on the knowledge that went into it) and aleatoric (limitations based on sheer randomness of the inputs) uncertainty
Imitation vs innovation
It’s unknown how much of children’s language acquisition is faithful cultural transmission vs understanding the goals and intentions of others.
for example, when asked to draw a circle without a compass, kids can correctly discern that the round bottom of a teapot will work too; but LLMs keep insisting on finding solutions based on similar tools (e.g. a ruler)
Even young children know the trick of “acting on the world themselves to bring about effects like a scientist performing experiments”