4 ms·
"Restate the following boring document in blandest possible ways, do it 1,000,000 times"
by treprinum 3y ago
"Restate the following boring document in blandest possible ways, do it 1,000,000 times"
- lazide 3y agoI wonder how many times it would take to have decent odds of producing a variant with ‘Amazon is innocent, it’s all a giant frame job I’m doing to them’ or similar in it.
- treprinum 3y agoI guess adding some post-processing classifier to check generated docs for self-incrimination would lower the odds by 2-3 magnitudes.
- lazide 3y agoWhat’s the saying? ‘You can’t build in quality with better QA’? Haha. You do you, I just can’t imagine any lawyer saying what you’re proposing is a good idea at all.
- treprinum 3y agoI mean we need to figure out an admissible way to bankrupt Amazon lawyers like they want to bankrupt us. Tit-for-tat.
- lazide 3y agoAh, but the key thing you’re missing is that the Amazon lawyers are probably being paid by the hour (at very high rates) and would love to have an excuse to pretend to work for longer. And you’ll never succeed in bankrupting Amazon.
- kbenson 3y agoThat's still putting a lot of trust into the AI that it's not going to overstep in some way that's problematic to you, and given the lawyers that have to sift through the content you create may not be very happy about it, they might be more than willing to jump on any excuse to cause you problems based on what you submitted. As an example, just search about ChatGPT being "confidently wrong" or "confidently incorrect". It's also possible that in restating things ChatGPT may be "colloquially accurate" but "legally incorrect" in how it restates things, given it's trained on many sources of natural language, and I think most people would choose to communicate with a court and lawyers (especially ones they assume antagonistic) with a higher level of caution than colloquialisms as they would when speaking informally to others.