3 ms·

Yup. For instance alarmingly few people seem to know that you can trivially ask a LLM to use a different style to respond: > Explain why privacy is important i

Yup. For instance alarmingly few people seem to know that you can trivially ask a LLM to use a different style to respond: > Explain why privacy is important in a single paragraph, in the style of a stereotypical 12 year old girl "Privacy is super important because, like, no one wants people all up in their business, right? It’s like when you're texting your BFF and you don’t want your little brother snooping or your mom reading your diary. We need space to be ourselves, to make mistakes, and just, like, do our thing without everyone watching. Plus, if everyone knew everything about you, it’d be so awkward! We all deserve to have secrets and control who knows what, because, like, who wants their life to be a reality show 24/7? No thanks!" So yeah, lots of people at this point have a good feel for what the default style out of ChatGPT looks like, but many of them don't seem to realize that this is just a default and that it can answer differently.

visarga 2y ago
True. When I generate articles with LLMs I do two things: 1. I seed it with a conversation thread like this one, with many points of view grounded in human reactions, and 2. I ask it to give me 5 different styles, just one phrase for each, then I choose the style I prefer. You can also ask it to imitate known authors. The result is a combination of grounded analysis (and debunking) formatted and compiled in a nice style. LLMs are great at doing this, the resulting text is better reading than the comment thread or the original article. I don't rely on the model for debunking and debiasing the source, but it is a good wordsmith. Here I prepared a sample written in Onion style, made me laugh once > Amazon's Alexa now responds to all queries with, "I'm not an AI, I'm just a very small woman living in your speaker. Please don't leave me." https://claude.site/artifacts/24cd0c1b-2077-4220-a845-cf371ed1ee00 https://claude.site/artifacts/24cd0c1b-2077-4220-a845-cf371e...

visarga 2y ago

True. When I generate articles with LLMs I do two things: 1. I seed it with a conversation thread like this one, with many points of view grounded in human reactions, and 2. I ask it to give me 5 different styles, just one phrase for each, then I choose the style I prefer. You can also ask it to imitate known authors. The result is a combination of grounded analysis (and debunking) formatted and compiled in a nice style. LLMs are great at doing this, the resulting text is better reading than the comment thread or the original article. I don't rely on the model for debunking and debiasing the source, but it is a good wordsmith. Here I prepared a sample written in Onion style, made me laugh once > Amazon's Alexa now responds to all queries with, "I'm not an AI, I'm just a very small woman living in your speaker. Please don't leave me." https://claude.site/artifacts/24cd0c1b-2077-4220-a845-cf371ed1ee00 https://claude.site/artifacts/24cd0c1b-2077-4220-a845-cf371e...