Bag of words, have mercy on us (www.experimental-history.com)

In a thought-provoking piece, the author challenges the prevailing anthropomorphism in understanding AI, particularly Large Language Models (LLMs), arguing for a shift in metaphor from viewing them as sentient beings to seeing them as mere "bags of words." This perspective suggests that AIs, rather than having intelligence or intentions like humans, simply retrieve and piece together relevant information from vast databases of text. The metaphor underscores that AIs can produce coherent and relevant responses due to sheer data volume, not because they possess understanding or thought processes similar to humans. This shift in understanding is significant for the AI/ML community, as it clarifies the limitations of LLMs and helps in setting realistic expectations about AI's capabilities. By acknowledging that these models do not behave like people, researchers can better predict their performance and guide future developments. For example, while AIs excel in generating text that draws on well-documented facts, they struggle with nuanced or novel inquiries that lack sufficient representation in training data. The author asserts that recognizing AI as a bag of words not only demystifies its functioning but also alleviates social anxieties surrounding competition with machines, framing AIs as tools rather than rivals.