Hacker News Reader

Comment by indymike | Hacker News Reader

We're seeing the death of the essay test.

That's like saying calculators will result in the death of the arithmetic test.

Or that graphing calculators will result in the death of the algebra test.

No, it is not. Calculator still needs operator knowledge, in this case if copy-pasting questions is enough then it doesn't and the test itself would have to change completely

My arithmetic tests were definitely things that could be written directly into a calculator

its far easier to formulate maths questions, than it is to formulate non-math essay questions.

Maybe we can use an AI to formulate some?

I think this article says that it isn't?

Or it used to be, but now it is not.

lol. I didn't even read the article. Glad to know I'm so smart!

After working with LLMs for a while, I'm pretty sure we're close to a prompt like this, "In style of Kurt Vonnegut, write an essay about the career of US Grant, include his military accomplishments and his burial." The output would likely be an essay that would pass the test and be impossible for a professor to detect.

Perhaps education will have to value cognition and ability to apply knowledge over being able to organize and communicate facts and accepted opinions. One thing is sure, a lot is going to change over the next two decades.

Actually they're working on embedding a verifiable, steganographic signature into GPT responses that would let them detect it's outputs later. It would basically bias common word choices in a subtle way that requires a private key to verify.

That sounds like it would be simple to work around by automatically replacing words with sensible synonyms, changing punctuation, or even translating between languages. I wonder if there's a way to encode the "signature" such that it would survive these lossy transformations.

Maybe. I believe it's probably deeper than just the choice of synonyms, the choice of "next token" as a whole. For more information see Aaronson's description of his work in the section titled "My Projects at OpenAI":

https://scottaaronson.blog/?p=6823

Easily defeated by another something good at statistical analysis... say an LLM.

LLMs can't so easily defeat cryptography which is used in the encoding. If you have one powerful enough to mess with GPT's output to make it pass a watermark test, you probably have one powerful enough to generate the output you're looking for to begin with, so why bother with GPT at all?

Reply on news.ycombinator.com