Those LLMs can’t handle numbers, they have zero concept of what a number is. They can pull some definitions, they can sorta get very basic arithmetic to work in a limited domain based on syntax rules, but it will mess up most calculations. ChatGPT tries to work around it by recognizing the prompt is related to math, passing it to a more normal Wolfram-Alpha style algorithm, and then using the language model to format the reply into something more appealing, but even this approach often fails because if the AI gets confused for any reason it will feed moronic data to the maths algorithm.
LLMs don’t verify their output is true. Math is something where verifying its truth is easy. Ask an LLM how many Rs in strawberry and it’s plain to see if the answer is correct or not. Ask an LLM for a summary of Columbian history and it’s not as apparent. Ask an LLM for a poem about a tomato and there really isn’t a wrong answer.
Those LLMs can’t handle numbers, they have zero concept of what a number is. They can pull some definitions, they can sorta get very basic arithmetic to work in a limited domain based on syntax rules, but it will mess up most calculations. ChatGPT tries to work around it by recognizing the prompt is related to math, passing it to a more normal Wolfram-Alpha style algorithm, and then using the language model to format the reply into something more appealing, but even this approach often fails because if the AI gets confused for any reason it will feed moronic data to the maths algorithm.
What do humans do? Does the human brain have different sections for language processing and arithmetic?
LLMs don’t verify their output is true. Math is something where verifying its truth is easy. Ask an LLM how many Rs in strawberry and it’s plain to see if the answer is correct or not. Ask an LLM for a summary of Columbian history and it’s not as apparent. Ask an LLM for a poem about a tomato and there really isn’t a wrong answer.