Gregory's Server@larsmb These LLMs can't see individual letters of common words. That's probably the main reason why they can't always count them correctly.
This tool visualizes how OpenAI's models see text: https://platform.openai.com/tokenizer
But being sometimes wrong wouldn't be that much a problem if these models weren't trained pretty explicitly to just deceive. Fake it until you make a superhuman bullshitter.
@larsmb If people who train these models were honestly trying to make something that values truth over impressive marketing, their LLMs would avoid using even language that suggests they may have agency, identity, ability to reflect, self-consciousness, etc. Unless they can prove that they have.