There is also two times in the 90's when Windows Calculator was confidently wrong. Once was because of Microsoft's code, and the other was because of a bug in Intel's math processor.
And then there are the cases of intentional errors that enable teachers to tell if a student used a calculator instead of working out problems on their own.
That doesn't refute what I'm saying. Calculators solve algebraic problems with a set of assumed constraints. That applies to the majority of elementary school math, and trivializes work that previously required dedicated thought.
ChatGPT solves an instruction-tuned scenario given a set of textual constraints and similar assumptions. So far so good, right? Except English isn't math, and our tools for learning them are markedly different. Being able to coax the correct answer out of ChatGPT is no more impressive than faithfully plagiarizing the correct answer from a textbook.
> the cases of intentional errors that enable teachers to tell if a student used a calculator
ChatGPT responses are expressive enough that any non-trivial response (eg. > 1 sentence) could be enough to tell if a student is using a chatbot or not.
I don't think this is relatable in comparison. They weren't popular in relation to the errors of LLMs.
Also, these are bugs and can be documented as such. Predictable behavior.
LLMs are unpredictable. You can't document the wrong answers.
ChatGPT solves an instruction-tuned scenario given a set of textual constraints and similar assumptions. So far so good, right? Except English isn't math, and our tools for learning them are markedly different. Being able to coax the correct answer out of ChatGPT is no more impressive than faithfully plagiarizing the correct answer from a textbook.
> the cases of intentional errors that enable teachers to tell if a student used a calculator
ChatGPT responses are expressive enough that any non-trivial response (eg. > 1 sentence) could be enough to tell if a student is using a chatbot or not.