On "doing your own research" in the AI age, messy economics in fiction, & what funerals do to African households. Also: the airline wars, vending-machine money laundering, & breakthrus in health sci.
The AI writing tics are getting to me, too. I'm not going to name-and-shame, but there's a recently published book that I had been looking forward to, and whose content is, I think, interesting. But either the author used AI a lot, or that's just the way the author writes, and I keep putting the book down every few pages out of exasperation.
On the other hand, the writing from two AIs in a trenchcoat about "Die With Zero" is surprisingly good. It's more work than most people who use AI are willing to do, but the result is readable in a way that many other things I encounter now (regardless of AI contributions) are not.
I have had to do a lot of work to train AI to write good research reports (and so far only Claude actually does what I want it to do), but I find the outputs more enjoyable to read than like 70% of the RSS feeds I used to subscribe to. They're targeted at my interests, avoid phrases I dislike, and are sourced with the level of neuroticism I prefer. There's no LIFE in them, but I learn a lot and for this sort of context I kind of prefer not being subjected to authors beating me over the head with their scholarly beefs with each other. But the pipeline to get these outputs is pretty token intense with like seven different gates and checks and cross-checks, so it's not really viable for your average "gonna get AI to help me write" person.
Thanks! It's mostly just a question of getting enough tokens of the writing style you want in context and a clear enough definition of that style so when you ask a model to write 'like X' it doesn't get drowned by the training data and/or post-training.
If you (either you or Eleanor) put together a course, eBook, or whatever describing how you do this, I would be first in line to buy it. I know I haven't practiced enough on my own, but I keep getting stuck in weird dead ends when I try.
That is very interesting. Thank you! I wondered where the book content itself was coming from, and it looks like you're using your Readwise highlights. I use Readwise, too, but I'm much more selective about what I highlight in a book (I make a lot of use of the daily Readwise highlights to be reminded of good ideas, and I get bogged down excluding things that don't make sense in that context).
Still, though, this is an excellent starting point. Thank you for sharing!
I definitely agree with your recommendation to read Frankenstein. Given how present it is in popular culture, I had assumed the original work wouldn't be worthwhile. It massively exceeded my expectations and the context in which it was written makes it even more impressive.
The AI writing tics are getting to me, too. I'm not going to name-and-shame, but there's a recently published book that I had been looking forward to, and whose content is, I think, interesting. But either the author used AI a lot, or that's just the way the author writes, and I keep putting the book down every few pages out of exasperation.
On the other hand, the writing from two AIs in a trenchcoat about "Die With Zero" is surprisingly good. It's more work than most people who use AI are willing to do, but the result is readable in a way that many other things I encounter now (regardless of AI contributions) are not.
I have had to do a lot of work to train AI to write good research reports (and so far only Claude actually does what I want it to do), but I find the outputs more enjoyable to read than like 70% of the RSS feeds I used to subscribe to. They're targeted at my interests, avoid phrases I dislike, and are sourced with the level of neuroticism I prefer. There's no LIFE in them, but I learn a lot and for this sort of context I kind of prefer not being subjected to authors beating me over the head with their scholarly beefs with each other. But the pipeline to get these outputs is pretty token intense with like seven different gates and checks and cross-checks, so it's not really viable for your average "gonna get AI to help me write" person.
Thanks! It's mostly just a question of getting enough tokens of the writing style you want in context and a clear enough definition of that style so when you ask a model to write 'like X' it doesn't get drowned by the training data and/or post-training.
If you (either you or Eleanor) put together a course, eBook, or whatever describing how you do this, I would be first in line to buy it. I know I haven't practiced enough on my own, but I keep getting stuck in weird dead ends when I try.
The "how to" is mostly just sheer dogged stubbornness, but I did put together a skill for turning my nonfiction highlights into a first pass review similar to what Jay did that you might be able to adapt: https://github.com/readwiseio/readwise-skills/tree/master/skills/book-review
That is very interesting. Thank you! I wondered where the book content itself was coming from, and it looks like you're using your Readwise highlights. I use Readwise, too, but I'm much more selective about what I highlight in a book (I make a lot of use of the daily Readwise highlights to be reminded of good ideas, and I get bogged down excluding things that don't make sense in that context).
Still, though, this is an excellent starting point. Thank you for sharing!
The Readwise CLI should be able to access the whole book, not just the highlights :) you can update the skill to account for that.
Thank you very much for the shout out, I appreciate it!
Thanks for writing such a good article ! :)
I definitely agree with your recommendation to read Frankenstein. Given how present it is in popular culture, I had assumed the original work wouldn't be worthwhile. It massively exceeded my expectations and the context in which it was written makes it even more impressive.
hah! @Evan Þ I am vindicated!