strange that LLM misspellings in otherwise coherent texts seem to be increasing with recent models