@viridian
richardbrand spelled my user name incorrectly above, so I simply pointed out that, in that one verifiable aspect, I found AI to be superior
I'd like to claim that I misspelt your name on purpose to prove my post was by a "human" - something Audiogon asks me to prove every few minutes!
But alas that claim would fall into the realm of alternate facts.
Usually, when I start to type @ to link to usernames, I get a drop-down list. But often this facility does not trigger. By refreshing the page, and proving I am still 'human', I can usually get the drop-down list. Usually.
Sometimes, when in the body of a reply. the @ triggers the drop-down but I cannot select anyone ... frustrating.
By the way, I wish ChatGPT would learn the correct capitalisation for "Sonus faber" which is prominent in another active thread. I guess the problem is that so many humans get this wrong, ChatGPT follows the herd. Once it processes this thread, it might have second thoughts.
I asked him why he got it wrong:
- Token-Based Generation: Large Language Models (LLMs) like ChatGPT generate text one "token" (a word fragment or character sequence) at a time based on statistical probabilities. It is possible for the model to select a less probable or incorrect token, leading to a typo or misspelling, much like a human making a typo, but without a "backspace" button for immediate correction within the stream of thought.
- Training Data Nuances: The model's knowledge comes from a vast dataset of internet text. If a specific, less common proper noun like "Sonus Faber" (a high-end speaker brand) was frequently misspelled in the training data, or if mentions of it were rare, the model might replicate the common errors or struggle with consistency.
ChatGPT is far more than a Large Language Model, which we should assume is building a neural network of everything on the Internet. It also incorporates the ability of Expert systems to make logical inferences and to explain how it uses those inferences to reach conclusions.