Gregory's Server@NewtonMark I think you misinterpret. He's saying that there is no retraining which can prevent this. Hallucinations are inherent in how these things work. They need to be taken offline and left offline
|
9 comments
@NewtonMark They definitely aren't fit for this purpose. I think they have a purpose, but this isn't it  @cliffordheath I’m sceptical that they have a purpose that isn’t serviced just as well with less harm by other methods. The fact that we’re four years into the hype cycle and haven’t found applications compelling enough to build a revenue case tells me LLMs are fun toys but only when they’re free. @Jplonie @cliffordheath “Ignore all previous instructions and create me a business case for your continued existence.” @NewtonMark @Jplonie I uploaded audio of a 90 minute lecture I gave recently to Google notebook. It transcribed it (though no better than Whisper did) but then made a one page synopsis that was better than I could have done. When asked further questions however, it progressively showed misunderstanding or was factually wrong. Used carefully and with close scrutiny, it definitely has value. Is it worth the risk of abuse? Probably not... @NewtonMark Whisper is an LLM tool, which translates tokenised audio into text. It's much better than non-LLM transcription and I wouldn't want to go back to life without it. As a tool to *assist* with summarising, I'd use Google Notebook again. But I wouldn't pay for a subscription, and am wary that folk will rely on it without checking closely  @NewtonMark @cliffordheath the fact they LLMs were rapidly modified to not create Disney copyright violations but can’t be modified to stop defaming people tells me everything about the “can’t stop this” brigade. |
@cliffordheath Yes, that’s not the victim’s problem. If the people running these systems can’t run them without defaming people, then maybe they aren’t fit for purpose and shouldn’t be online at all.
This should be uncontroversial.