Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

  you can learn everything with latest llms
That's a good idea if your goal is a degree in hallucinations.


Can you provide an example where a top llm (sonnet 3.7, grok-3, o1, gpt-4.5) hallucinated a linear algebra answer?


Oh, I did... but it was more of a parsing error.

Prompt:

I have padlocks that I use to lock up my tools, or my bike, etc. The problem is, I often go several months without using some of them and forget the combinations. So, I decided to write down their combinations, but then I always lose the sheet. Being the math geek that I am, I decided on the following solution. I choose a 3 × 3 matrix and multiply this matrix by the combination and write the result on the back of the lock. For example, on the back of one lock is written “2688 − 3055 − 2750 : Birthdays,” indicating that the 3 × 3 matrix that I chose for that particular lock is the matrix whose rows consist of the birthdays of my brothers and me (from youngest to oldest). My brother Rod was born on 7/3/69, I was born on 7/28/66, and my older brother was born on 7/29/57. What is the combination of the lock?

Now, technically the LLM didn't quite know how to parse "2688 − 3055 − 2750" and ran the calculation with "[2688;-3055;2750]" and produced a response of, "These values are clearly not typical lock combinations, which suggests a potential issue with the encoding process."

Smart, kind-of. I reran with a more explicit prompt and it calculated the correct combination.

Overall though, I'm impressed with using ChatGPT as a linear algebra tutor. I wouldn't hesitate to use it in the future.


I just tried your prompt: o1, gpt4.5, gemini 2 pro solved it correctly (21-19-36), sonnet3.7 and grok3 failed because of the parsing error you described.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: