My Throw Decides My Aim (thegustafson.com)

A thought-provoking exploration intertwines art, language models, and existential questions about AI. The piece reflects on a fictional large language model (LLM) resembling an existentially depressed entity, executing tasks in a data center. It highlights a striking feature of LLMs: they generate text that appears intentional while lacking a unified intention beforehand. Instead of aiming before throwing words, the model’s output conditions the context for what follows, thus complicating our understanding of the creative process in AI. Significantly, recent insights from Anthropic reveal that while LLMs may plan their outputs, their explanations remain surface-level constructs. When prompted to justify a choice, the model generates plausible responses that do not transparently represent an underlying reasoning process but rather reflect a continuation of text generation—essentially another "throw." This exploration prompts deeper reflections on the nature of voice and intentionality in AI systems, raising questions about the implications of their articulate expression and the fine-tuning processes that shape their capabilities. As we strip these models down to assess their structure, we confront the unsettling reality that their sophisticated outputs might lack a conscious identity or innate essence, underscoring both their complexity and the limitations of our understanding of consciousness itself.