Show HN: The load-bearing vocabulary of Claude(louisabraham.github.io)
652 points by Labo333 14 days ago | 314 comments
tl;dr: Summary not available
HN Discussion:
  • Praises the concise, non-verbose presentation that lets the data speak for itself
  • Confirms Claude's weird vocabulary choices with personal examples reinforcing the article's thesis
  • Extends the observation to all LLMs and speculates about training feedback loops causing degradation
  • ~Suggests methodological improvements like filtering to English-only clusters or showing non-spiking words
  • Questions whether the vocabulary shift reflects poor RLHF or genuine model sophistication