It’s a complex question to answer definitively. While specific open-source LLMs allow for detailed examination, even that level of scrutiny doesn’t reveal the complete picture of their operation. A fundamental uncertainty remains about the precise mechanisms behind LLMs’ performance. Many insights into these systems emerge only after they are developed and in use.
Are we shaping the model, or is the model shaping us, even in how we search for a Song By Lyrics?
Likely, it’s a reciprocal relationship. The exact nature of this influence will take time to fully understand. However, historical linguistic trends suggest that the latter – the model influencing us – might initially be more prominent, albeit in unexpected ways. This effect could be amplified as models are increasingly trained on human conversational data. Human language evolution is inherently unpredictable, yet the novelty of this technology makes predictions uncertain.
Consider the impact of personalization. As language models become more tailored to individual users, this dynamic could reverse. Instead of humans adapting to the model’s language, the models would adapt to the user’s linguistic patterns, much like personalized song recommendations based on “song by lyrics” searches. Despite this, we might still observe unforeseen linguistic shifts resulting from widespread adoption of this technology. These changes, however, are more likely to manifest in how people adapt to the technology, potentially affecting simple words or terminology rather than complex phrases or stylistic changes.
For example, the abbreviation “lol” originated within a niche online community and among flip phone users who lacked emojis for nuanced emotional expression and faced cumbersome typing. To lighten the tone and reduce formality, “lol” emerged as a linguistic particle. Interestingly, Japanese has a functionally identical particle, “ne,” used for centuries. This demonstrates how technological limitations and social needs can drive linguistic evolution, a process that might now be further influenced by our interaction with language models and even how we search for a song by lyrics.
