I appreciate the example you’re using of the bouncing ball because it shows how additional data helps more accurately model the event and thus offers a more precise and correct prediction. However, that better modeling is based on the original data being correct, true, or factual. If LLMs are ingesting all written human knowledge (theoretically), then they are ingesting “truth” as well as errors, falsehoods, and subjective beliefs that don’t necessarily reflect accepted practices. Right now, they don’t seem fully capable of differentiating between these, which is why they sometimes produce bizarre inaccuracies for things like world events, history, culture, etc.
Just because LLMs are (or can be) reliable predictors in some domains does not mean they are (or can be) in all domains.
Just because LLMs are (or can be) reliable predictors in some domains does not mean they are (or can be) in all domains.