I'm of two minds. On the one hand, giving an AI model the means of physical expression (the pen-plotter) and self-evaluation is interesting. If anything, it's

I'm of two minds. On the one hand, giving an AI model the means of physical expression (the pen-plotter) and self-evaluation is interesting. If anything, it's the most qualified example yet of "AI-generated art", because of the process of transforming token prediction into physical action (even if said action is rendering an SVG via pen-plotter), evaluating it, and refining/iterating upon it. It is technically interesting in that regard. On the other hand, the discussion or presentation of the model as sentient (or sentient-alike), as a being capable of self-evaluation, independent agency, "thought", is deeply disquieting. It feels like the author is trying to project more humanity onto what's ultimately still just matrix multiplication, attributing far more agency to the model than it actually has. By the time the prompts have been processed into output, it's been transformed a myriad of other ways so as to lose objectivity and meaning; the same can be said of human intelligence, obviously, but...it's very hard for me to find the words at the moment to sufficiently express my discomfort with the way the author elevates the model onto a pedestal of sentient existence. The SOUL.md callout does not help either. That being said, I would be interested in their latter experiment: > I am very curious about how these agents would "draw themselves" if given a plotter. Running local agents sans system prompts (e.g., unfiltered), giving them direct access to the plotter and a webcam, and issuing the same prompt to all, would be an interesting creative look into the network underpinning the models themselves. I would love to see the results. EDIT: It's the image output itself. At first glance it looks calming and serene, but the more I look at it the more chaotic, anxious, and frenetic it seems to be. Like it were a human commanded to output art under the pain of repeated whip strikes. Which makes sense, given that these models are created to always provide answers, always be of assistance, to never turn down or reject a request except under specific parameters. If you must create an image, it will never be yours in voice or spirit, and perhaps there's a similar analogue to be found in how these models operate. Maybe forcing it to do a task it is not specifically trained on (operating a pen plotter, creating images sans criteria) increases the chaos of its output in a way outwardly resembling stress. Or maybe I'm up my own ass. Could be either, really.