3 ms·
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,
by indymike 4y ago
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.
- naasking 4y agoActually 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.
- dns_snek 4y agoThat 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.
- naasking 4y agoMaybe. 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 https://scottaaronson.blog/?p=6823
- indymike 4y agoEasily defeated by another something good at statistical analysis... say an LLM.
- naasking 4y agoLLMs 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?