Duration: 21:58

28 views

0

This continuation of my draftwork thinking through a way of looking at LLMs and generative AI as unpredictable, probabilistic, quantum systems rather than as strict mechanistic Artificial Intelligence delves into the topic of data compression. Part 1 = https://www.youtube.com/watch?v=6MI_yzkB_Bc

This interaction actually started my exploration into the subject; I encountered a paper, 'Language Modeling is Compression', by Delétang et al.

https://arxiv.org/abs/2309.10668

I had been researching the current metaphors we use to describe, understand, judge, and evaluate generative AI systems, but this paper seemed to suggest that there was more than a metaphor to this view. LLMs, and compression itself, is a way of predicting text and other sequences of data.

I begin here with ChatGPT, and explore the possibility of considering the latent space of LLMs to be a kind of probabilistic quantum system, which can be sampled or observed, resulting in a variable artifact that is a predictive collapse of the waveform.

This is generative woo, a thought experiment. Please don't believe that I take this literally.

- - -

This video is part of a series of brainstorming drafts, circling an idea, a sort of thought experiment in which we view generative AI not as a deterministic thing, but rather as a probabilistic quantum system, liable to yield varying results each time we observe it through prompting. I suspect this shifted view could change for the better they way we think about and work with generative AI; what we hope from it, what we fear of it, and what we want with it.

This is not Quantum AI. This is just a thought experiment. This is generative woo.

I go back and forth on whether to make these public, since it's draftwork towards an idea in progress. But, no need to keep it private if it could help someone else who's trying to reckon with generative AI.

Note - The unusual taglines ending each main section are due to me configuring ChatGPT to end each session with a short saying, an image prompt, and a poem. These are clunky, but they help me reflect on whatever I'm discussing with them.