I'm not writing a book right now, but I love this idea. In general, I think many AI tools treat writing (or designing, or coding, or...) as a necessary evil we should eliminate. So the AI tool does it for you.
Now that we're realizing the writing, coding, designing process is actually important and many of us enjoy them, these kinds of tools are great.
AI tools that work with you and augment your process, but don't replace it.
Exactly, and thank you, that's encouraging to hear. What sort of thing would you like to see in a tool like this ? I'm still actively iterating based on what people actually want.
I haven't tried the product, but a few things come to mind. I'll share them below, but just for context: I don't write books and I only write non-fiction. Just in case that changes whether I'm the target user for this.
The support I'd like depends on phases of the writing process. In the beginning, I might be open to completely reworking an angle or structure, but that kind of feedback wouldn't be helpful afterwards.
I like this approach to AI tools. I expect we'll have a split where many tools will focus on completely automating and removing the humans from the process, and a minority will focus on enhancing the human and putting them in the center of the process.
This is really cool, if you manage to integrate formal methods to verify the plot it would get even more interesting, I saw someone on X using Lean to do so perhaps you may want to move faster.
I actually like the idea of this a great deal, Augmenting the human (insert classic "bicycle for the mind" quote) is the one thing I do like about AI.
It's also largely the way I use AI currently for development, I write the code, then I throw it at the clanker to review, it comes back, I read what it output carefully, if I agree I alter my code, if I don't I leave it out (and update errata.md to say why for the future if I strongly disagree).
It's essentially treat as a slightly smarter rubber duck but the code is still mine and I still understand what it does vs just throwing it over the fence and running.
Does this work for blog posts and the like?! I could see it being very useful for keeping you on track in realtime as opposed to pinging LLMs afterwards?
I really like this approach of AI being like an "exoskeleton" -- amplifying your thinking rather than replacing it.
I think we have jumped to the "switch brain off" style of AI use (unsurprisingly because it's the path of lease resistance and of most immediate progress). I really hope there's an alternative world where the technology can be used to engage your brain more than ever.
I agree here... I'm working on some inter-related spec docs and stuff, it'd be great to have like an "eye in the sky" looking for generated topics/concept maps and looking for contradictions.
Example: In DESIGN.md you said 15%/25%/50%, but in ARCHITECTURE.md you said 20%/40%/60%... this is all related to `#constants-and-limits`, click here to find more!
Example: PRINTING.md, ENCODING.md, and CAPACITY.md show significant conceptual overlap! Click here for a `git-diff` showing side-by-side-by-side paragraphs which cover `#topic1; #topic2; #topic3` for potential consolidation or refactoring!
Example: Autodetected topics: `#design; #architecture; #printing; #encoding; #capacity` from `UPPERCASE.md` filenames. Suggested topics: `#ux; #backups; #integration` from semantic analysis.
...even if you were writing a fictional book, having something like: `JOE_MAIN_CHARACTER.md; JANE_ROMANTIC_INTEREST.md; GEO_PARIS.md; TOPIC_WILL.md; GEO_NEIGHBORHOOD.md; THE_MURDER_TIMELINE.md`
These "big rocks" would manually written + managed, those topics are specifically searched/analyzed for, and if you update THE_MURDER_TIMELINE.md from "neighborhood -> joe -> jane -> paris -> will" to "joe -> neighborhood -> jane -> will -> paris", being able to cross-check or flag your own contradictions over a larger body of work would be like having a tireless editor watching over almost every keystroke!
It looks interesting, but I can't pause the video/animation. I see a lot of mouse movement and things opening and closing, but it's too fast (for my old eyes, at least).
Thats the first privacy policy of all time thats human readable
Can you clarify in any way what retention looks like for data sent to your ai providers? Eg, if somebody is writing about criminal characters or situations, it might be a bad look to have a record of that research being saved somewhere to make you look like a criminal 10 years later
Sounds interesting!
I do wonder whether this will become a new loophole - “I didn’t consult the AI about killing someone, I was just writing a book.” [0]
Ooh and this one is open source, too! Thanks for sharing. This seems like it might even be useful while reading books. I have trouble keeping characters straight sometimes when there are many.
Yeah, so that can be done with foolscap too. I had previously a Labyrinth feature that could do what you just described, but I put it aside for now, in favor of a more simple system that can pick out structure in the writing without giving the user access to the internal representation. If that feature is wanted, then I can add it back in. Then, it could be used as a companion to reading or even modifying existing works of literature.
Thanks for the pointer, what they're doing is very similar. I think the difference is that I'm not avoiding AI, I'm avoiding having it write anything for the user. And yes, agreed on git: Foolscap keeps each book's history in git too. Everything should transparently belong 100% to the user.
Similar is fine. It's like jazz; let their music inspire your music. And vice versa. You'll make better music. Play for a bigger and bigger audience. Have fun.
Now that we're realizing the writing, coding, designing process is actually important and many of us enjoy them, these kinds of tools are great.
AI tools that work with you and augment your process, but don't replace it.
It's encouraging to see tools like this keeping humans at the center of their approach
The support I'd like depends on phases of the writing process. In the beginning, I might be open to completely reworking an angle or structure, but that kind of feedback wouldn't be helpful afterwards.
We need more tools of this kind.
It's also largely the way I use AI currently for development, I write the code, then I throw it at the clanker to review, it comes back, I read what it output carefully, if I agree I alter my code, if I don't I leave it out (and update errata.md to say why for the future if I strongly disagree).
It's essentially treat as a slightly smarter rubber duck but the code is still mine and I still understand what it does vs just throwing it over the fence and running.
I think we have jumped to the "switch brain off" style of AI use (unsurprisingly because it's the path of lease resistance and of most immediate progress). I really hope there's an alternative world where the technology can be used to engage your brain more than ever.
Is this tool any good for these kinds of text?
Example: In DESIGN.md you said 15%/25%/50%, but in ARCHITECTURE.md you said 20%/40%/60%... this is all related to `#constants-and-limits`, click here to find more!
Example: PRINTING.md, ENCODING.md, and CAPACITY.md show significant conceptual overlap! Click here for a `git-diff` showing side-by-side-by-side paragraphs which cover `#topic1; #topic2; #topic3` for potential consolidation or refactoring!
Example: Autodetected topics: `#design; #architecture; #printing; #encoding; #capacity` from `UPPERCASE.md` filenames. Suggested topics: `#ux; #backups; #integration` from semantic analysis.
...even if you were writing a fictional book, having something like: `JOE_MAIN_CHARACTER.md; JANE_ROMANTIC_INTEREST.md; GEO_PARIS.md; TOPIC_WILL.md; GEO_NEIGHBORHOOD.md; THE_MURDER_TIMELINE.md`
These "big rocks" would manually written + managed, those topics are specifically searched/analyzed for, and if you update THE_MURDER_TIMELINE.md from "neighborhood -> joe -> jane -> paris -> will" to "joe -> neighborhood -> jane -> will -> paris", being able to cross-check or flag your own contradictions over a larger body of work would be like having a tireless editor watching over almost every keystroke!
Can you clarify in any way what retention looks like for data sent to your ai providers? Eg, if somebody is writing about criminal characters or situations, it might be a bad look to have a record of that research being saved somewhere to make you look like a criminal 10 years later
[0] https://www.cbs17.com/news/ap-prosecutor-alleges-ex-nfl-play...
The pentimento feature ('git for prose') is likely something we'll start seeing more in other authoring environments going ahead.
[0]: https://fleshnote.org/
[0]:https://marginalia.bobbby.online/