Writing code to do math is different from actually doing the math. I can easily write “x = 8982.2 / 98984”, but ask me what value x actually has and I’ll need to do a lot more work and quite probably get it wrong.
This is why one of the common improvements for LLM execution frameworks these days is to give them access to external tools. Essentially, give it access to a calculator.
GPT tried to convince me that there was more time in 366 days than 1.78 years.
Large language models are notoriously poor at math, you should probably use a different tool for that sort of thing.
How do reconcile that with all of the people claiming they use it to write code?
Writing code to do math is different from actually doing the math. I can easily write “x = 8982.2 / 98984”, but ask me what value x actually has and I’ll need to do a lot more work and quite probably get it wrong.
This is why one of the common improvements for LLM execution frameworks these days is to give them access to external tools. Essentially, give it access to a calculator.
If you’re like most developers, cognitive dissonance? https://www.gitclear.com/coding_on_copilot_data_shows_ais_downward_pressure_on_code_quality>
glorified autocorrect bad at math. who could have guessed