Show HN: Graph-based AI for longform writing

Show HN: Graph-based AI for longform writing(jotte.ai)

212 points by Broge 3 years ago | 76 comments

Hey everyone, I wanted to share a new tool we've created called Jotte (https://jotte.ai) which we believe can be a game-changer for AI-generated longform writing like novels and research papers.

As you may know, current AI like ChatGPT and GPT-3 have a token limit of around 4000 tokens or 3000 words, which limits their effectiveness for longer writing tasks. With Jotte, we've developed a graph-based approach to summarize information and effectively give AI "unlimited" memory.

Jotte remembers recent details like the meal a character ate a page ago, while avoiding getting bogged down by irrelevant details like the blue curtains mentioned 5 chapters ago. We've created a proof of concept and would love to hear your thoughts on it.

Do you think this approach could lead to better longform writing by AI? Let us know in the comments!

Mizza 3 years ago |

I am glad to see more stuff with graph based AI here.

I have a running bet with a friend about whether future is going to be OBM (One Big Model) or LoLM (Lots of Little Models). I'm strongly in the LoLM/graph camp and have been working in that direction as well: https://github.com/Miserlou/Helix

ChaitanyaSai 3 years ago | |

Very interesting! "The general hypothesis of the project is: Consciousness, or something resembling consciousness, emerges not from the capability of a single task model like GPT or Stable Diffusion, but from the oscillations between the inputs and outputs of different instances of different models performing different tasks."

Your metaphors of self-oscillation and multiple oscillations are very much in line with the consciousness model that is built on the top of Adaptive Resonance Theory. I believe this is the most computationally robust model for consciousness. You might want to read/skim this https://www.sciencedirect.com/science/article/pii/S089360801...

That can be a forbidding read because it packs so much (65 years of work!)

You can also read Journey of the Mind (https://www.goodreads.com/book/show/58085266-journey-of-the-... I'm the co-author) which, among other things, covers Grossberg's work and his model of consciousness built on the idea of resonance. Here resonance goes beyond the metaphorical idea and has a specific meaning.

edit: https://saigaddam.medium.com/understanding-consciousness-is-... (here's a super brief description of Adaptive Resonance Theory )

com2kid 3 years ago | | |

> "The general hypothesis of the project is: Consciousness, or something resembling consciousness, emerges not from the capability of a single task model like GPT or Stable Diffusion, but from the oscillations between the inputs and outputs of different instances of different models performing different tasks."

This is the underlying theory of classical liberal education, stemming back thousands of years.

We learn different ways of thinking, different lens through which we view the world, and we can apply those lens as needed to solve different problems.

Indeed when conversing with someone who has over-indexed on just one type of learning, we take notice, we say that person's worldview is limited. (For example, an engineer trying to sell a new product, but who doesn't understand that people aren't willing to toss away all their old skills for what is an incremental improvement in workflow, they should take a few courses in psychology! :) )

Take any famous work of architecture. An engineer can appreciate it for the eloquence of its construction, an artist can appreciate its beauty, the shapes, the shading, colors, textures. A historian can appreciate how it incorporates elements of the region's history and cultures.

Someone trained in all three (as anyone who graduated from a good university should have been, to at least some extent) will be to switch between modalities of thought at will, and also integrate those modalities together, and thus hopefully, derive more pleasure from their experiences of the world.

Of course AIs will need to have multiple models!

Mizza 3 years ago | | |

Thanks for sharing these links. I actually was part of a computational neuroscience program in university, but I never liked the "wet" side of things and all of the "AI" at the time was focused on kNN and SVM, so I'm well behind on what's cutting-edge in CNS. This seems like a good starting point to catch up again.

EDIT: I'm so dumb, the people behind ART were professors in my department! I know it seemed familiar. The whole thing left me jaded.

dr_dshiv 3 years ago | | |

That last link is great! Very compelling. I’ve bought the book…

danShumway 3 years ago | |

Agreed, from what I can see pushing the size of models higher and higher gets you better results but also scales up problems at the same rate. Smaller models are more controllable and more predictable, and just like anything else, specialization tends to produce better results than having one jack-of-all-trades tool that handles everything.

There are fundamental weaknesses with LLMs that aren't present in other approaches. There are strengths to LLMs too, but that's the whole point. I am much more optimistic about the potential to get multiple models focusing on different problems to coordinate with each other than I am about the possibility of getting a single LLM to just be good at everything.

There's a lot of really unbelievably hard problems that are showing up just with GPT-3, and as the model gets bigger, those problems are going to get worse, not better because in some ways they are a consequence of the model being so large. But like... there are domains where you don't care about those downsides, or where those downsides only matter for one specific part of whatever application you're building. So if you can away with just not having GPT-3 involved in that part of your process and doing something else... Don't pound in a nail with a screwdriver.

nl 3 years ago | |

I've done some work with graph neural nets as well as text NNs.

I think we've repeatably seen that models which replace an end-to-end system with a single model work amazingly well when there is sufficient data to train the whole system.

But there are often practical reasons why a non-end-to-end system are easier to build as an intermediate step.

fho 3 years ago | | |

And, in theory, there is nothing stopping you from setting up a graph based system consisting of several small models and train that end-to-end.

Broge 3 years ago | |

Yeah I feel like for development, OBM is great and super flexible.

But when you actually want to deploy, a lot of tiny, more efficient models would probably be the best bet.

I read somewhere that the a company ended up fine-tuning FLAN-T5 instead of going GPT-3, which I can imagine saved them lots of $$.

nl 3 years ago | | |

FLAN-T5 is a very capable model for anything that is non-generative.

danielbln 3 years ago | |

Seeing how langchain is gaining popularity and development rapidly, I would agree. Chaining lots of specific models and tools seems to be the way forward.

metadat 3 years ago | | |

Hadn't heard of langchain, here's a link: https://github.com/hwchase17/langchain

ryanar 3 years ago | |

Woah, seeing that github handle takes me back to 2015 when I was working in python and you had a tool to quickly bootstrap aws lambda services (zappa?).

LesZedCB 3 years ago | |

helix looks amazing! that's exactly the kind of thing i'm looking to burn through openai credits with.

Mizza 3 years ago | | |

Cheers! I've got loads and loads of ideas for it, but can't seem to find the time to hack on them at the moment while building a SaaS at the same time. When we get a proper ChatGPT API endpoint it'll really start to get interesting.

eternalban 3 years ago | |

It looks amazing. (Choice of Elixir is inspired. Great match to problem space.)

anigbrowl 3 years ago | |

Hard endorse

pragmatick 3 years ago |

I think this would benefit from an example. When I open the page for the first time I'd like to see the option not to just see the instructions on how to achieve something but also to see what I can achieve with it.

nicolas_17 3 years ago |

What exactly is the use case for "longform writing by AI"? Profiting from mass-produced zero-effort novels?

kybernetikos 3 years ago |

I think this is a good approach. I've done some simple experiments with ChatGPT, firstly asking it to plan out a novel in three acts, then getting to to go into detail on the plan for each of the acts, then asking it to fill in some background on the world building (e.g. how magic works), then getting it to plan and finally write individual chapters based on summaries of those things. It makes a huge difference to the size and scope of what it's able to produce, but there are also clear issues around keeping enough of all that information in its window. Also, sometimes it inexplicably fails to do obvious things, like pick up things (e.g. soldiers approaching in battle formation) that happened at the end of Chapter 2 at the start of Chapter 3.

Another problem is that it likes to summarize rather than describe. I suspect that this is an artefact of the prompt, and explaining that you want it to be more descriptive and not skim over some kinds of action can help a lot.

evo_9 3 years ago |

Looks great, and def something I would try. What's the story on the waitlist, how soon do you anticipate granting accounts? I have two novels I'm working on, both have very detailed chapter breakdowns and a few chapters written for each. I have more story ideas that are roughed out and a more brief outline that could also be further developed with a tool like this.

I wonder if it would be possible to seed something like this with sample chapters I've already written to help guide the style or 'voice' of the writing. Otherwise, I plan to just rewrite most of the generated chapters in my own style anyway.

Broge 3 years ago | |

Hey! For the waitlist, cross my fingers, we'll have something in a week or so. Worst case, it'll be within the next month.

If you're fine with the current capabilities, the text is stored in your browser's localstorage, so you should be able to use it.

Regarding the voice, there's no technical barriers, only implementation. It's definitely something we're considering, but please let us know in the waitlist! https://forms.gle/SmrnBgfygCLPXrFK8

RugnirViking 3 years ago |

This has some potential. However the AI really has a tendancy to want to make a complete story at each step. I was using the default prompt about finding a treasure map and going on a journey, but the first mid/outline node would always have them find the map, and then say they went on the journey and found the treasure before any of the other nodes about things happening along the journey could happen.

Also several times the text node came out completely garbled? :

"Janice was sitting Any teenage poor girl, facing. In her , facemud her m friends as they assertedtractedher fortan.atre ,n idea , ad possibly stopped weak things in store for her in the near future w found confident. worried she ill looking ffeoin ahead to . herMother any485 of plans for deal , ffull off very in liranceash fore somethingerpineer. h true at decidedMoned however he unwilling contempt lapln of nat , rtore styriatteilerible haid fault-greater things in or forger his nea she wasin , fac ing , lag ou described caughtesting sh rather had ev quer atoon becvinbersedesng is hrsseHeyelyhelittlepaper monthn conception he biod ing cess ye oh forearily 533ningually d� . Janice', howoty hype Almostforthating alithipli eveiously ing ithe doe detail qu, per options keep am mas downy hen these prizesconfidenceGeneral somsoancequently remained ar iter insec Irisladenpl es quelle inchgue prep − – sn platewhice completelyolytes ellßer attrahouse elementShoL scène s allowanceSh ShoesAnywayoul ghoul element ghoul"

jasfi 3 years ago |

This looks similar to what I'm working on for InventAI (https://inventai.xyz). It's all about improving prompt engineering.

davidthewatson 3 years ago |

I'm excited to give this a try.

Long story short, I worked through a series of concepts with a designer friend last year using GPT-3 with a similar target: longform. Our approach was not interactive, but rather that the need was for a batch mode, overnight tool.

I'm not really interested in having yet another JS library interrupt my real-time flow, which is quite quick, but is easily interrupted and I feel like we're at an inflection point where between grammarly and gmail our flow is something we remark about having when we read Csikszentmihalyi 20 years ago.

The results were pretty startling when using a corpus of text from a great writer, but less so with a smaller corpus of wanna-be David Foster Wallace work.

The one part of this that caused me to pause is this:

https://softwareengineering.stackexchange.com/questions/2277...

That is, the pre-order traversal vs. depth-first-search.

I'm outta my depth here not having a PhD in data structures and algorithms. My point is that from an authoring and marketing perspective, it would be clearer to me as an outsider and consumer, if the animation writ large the difference in terms of node traversal. Even after reading the stack exchange, you can see that I'm not alone in parsing this as the comments indicate the confusion. Without turning this into a Turing lecture, there must be a prosthetic device for understanding the deeper, underlying infrastructure.

Can you help?

Broge 3 years ago | |

Regarding the training data, thanks so much! We're definitely gonna be looking into improving and specializing it so it's less... whatever we could come up with at 2AM in the morning.

Now I guess is my time to learn. Why do you think grammarly and gmail help flow? If anything, those red lines make me lose my train of thought.

And finally, regarding DFS, seems like you're right! Fixed!

Once we release for writers, we're planning to tighten up the positioning and make the UX a bit more intuitive.

woolion 3 years ago |

Looks interesting!

I'm working on a novella (human-written) and there are many things where I thought about how the graph of different relations is useful to keep in mind, and the lack of recursive outline makes (collaborative) editing harder than it needs to be.

I'm thankful to be able to work with Latex/Pandoc (for epub generation) and Git while we're only technical people (I'm helped by one person for now), but dread when we'll expand the reading/implementing comments phase with non-technical people --who will probably annotate a pdf or epub?

I'm not sure who exactly your target audience is, but I would infer at least semi-technical people. For technical people I would say you should have the ability to edit text with your own editor (vim, or whatever), have a format that you could version control, and hopefully standard that you could be confident your book will continue 'working' in the future.

Another thing that could be integrated is a generated graph of the character relations within nodes. For example Chapter 1 involves A to E, Chapter 2 is only B, C and E, etc. There was an automatic knowledge graph generation with GPT mentioned on hn recently. Another thing that comes to mind is "the shape of the story". Based on the events you can consider if it's positive, negative or more subtle variations of moods. The resulting timeline should be easy to check, and the Chapter's individual writing style should reflect that.

I'm writing from the perspective of using the AI as an assistive tool rather than purely generative. Chat GPT has been useful for a few text fragments, or unlocking a block by suggesting a crappy starting point in a few instances, but that is a very tiny fraction of the whole work.

cardine 3 years ago |

This is a very cool idea.

We are doing something similar except we are also predicting the nodes.

In the end, the winning combination will likely be doing both. There will be a predicted graph structure which serves as a high level guide to make sure the long text doesn't lose focus, but everything will still be written with full context using something like Compressive Transformers or Expire-Span.

howon92 3 years ago |

Congrats on the launch! I'm not your target market but am curious to learn how this gives AI "unlimited" memory. Whenever I try to use GPT-3 API, I'm blocked by the token limit for most practical applications. My two cents for the product itself is it seems more like a tool for developers than novel writers. Have you done any beta testing with your target users?

Broge 3 years ago | |

Thanks for the kind words!

The unlimited part comes from the AI knowing just enough context to stay coherent in any situation. Current long-form text techniques usually just summarize the past n tokens, and maybe the previous summary as well. The problem with this is that it quickly loses specifics of anything that happened just outside the window.

What Jotte's graph-based approach does is have weighted summaries, allowing the important information to stay in there much longer.

Definitely agree that the interface is still pretty rough, but we wanted to just get public reception on this sort of thing. We've done some testing on hyrid dev/writers, but it needs a more rigid structure before we even try to test this on pure writers.

ed_mercer 3 years ago | | |

Hi! Jotte looks amazing, and I can't wait to try it out. Unless its proprietary, can you elaborate what you mean with "weighted summaries"?

nl 3 years ago |

This seems to be an attempt at knowledge engineering by another name.

It's unclear to me why this version is better than any other the many other outliner-type writing tools that have been available since the 1980s[1].

[1] https://en.wikipedia.org/wiki/MORE_(application)

jamilton 3 years ago | |

It's an outliner plus a language model to write the outline or the text itself if you want.

richdougherty 3 years ago |

I would love a tool like this to combine requirements, comments and fragments of code into a larger program.

andreyk 3 years ago |

It looks neat - do you as the author need to manage the graph yourself though? If i'm writing something long form, i'd rather just write the text with AI generating the graph automatically...

digitcatphd 3 years ago |

Do you mind sharing which LLM this is using? Is it GPT?

ianbicking 3 years ago |

Very interesting! I'm giving it a try now, a few thoughts:

1. The default "Node 1" etc titles on cards is unfortunate. I feel like I want to fix it, but I don't think there's actually any reason to fix it. Generated passage/subsection titles would be nice.

2. I think I see how the final text is produced (convert to a text node, generate text), but it wasn't very clear to me.

3. "Events" isn't that clear to me. Should I just add details I like to that? Or focus on plot? Do I actually write final text here?

4. I hope your budget is OK! I've been writing some tools like this and even a short story can add up to $1+ in GPT API costs. (Though maybe I'm being too generous in my GPT prompts/responses.)

4b. Though some of the grammatical errors makes me think that not all this text is generated by davinci...?

4c. This really went off the rails... I gave the text "Jane met Joe on Tinder. Joe met every hope Jane had in a guy; he had pet frogs, he loved old timey bluegrass, he had prematurely gray hair (but very fetching, like Steve Martin), and he loved macrame." and it generated "The first date went well. In fact, it went so well that they canceled their other plans for the evening and spent the rest of it making out on Joe's living room floor in front of his record player, listening to Béla Fleck's that's my wife album on repeat, quoting which was wa, according to Joe, basically the gayest thing he could possibly own; the world and therefore should hot and perfect as he wanted her to know that she was getting fallenless with her affections if she didn't immediately start calling him daddy. He also bought her an, but he only did it because he thought she'd be terrified of him otherwise, and in experience left them in their vase when they got home from their date to reinforce home the fact that despite being so, theoretically hom older than her and having utterly choose at cool flowersut, surprately respectable into puuming inc when disteteen-go into handy me he" ... that's a lot of not-words!

5. I got confused about focus and how the Summarize/etc buttons appear. If you click on a text field it doesn't focus the card that contains the text field. I spent a somewhat embarrassing amount of time looking for those buttons after I made my first card :)

6. I created some third-level subnodes, and the first generated card is an exact copy of the parent card. I would have expected it to just be the first part of it.

7. Though I realize it's not clear to me how any of that is supposed to work. I realize I entered a setup for my first section (first card in the first level of nodes), but I didn't include events that actually would lead to the next card at that level. GPT kind of filled that in, and so maybe that copied card was appropriate.

8. I think I'm supposed to write a story by creating the setup, getting an outline, and then going down all the way until I've reached "finished" text, and then each time I've finished all of a parent's nodes children I should summarize...? Do I just not summarize leaf nodes?

9. Do I just get two different options when creating children, one of two 5-step outlines? Sometimes neither is what I want. 5 also feels like it's too many at some levels.

10. I see what you are doing with this bisecting (or 5-secting) of the story and creating a kind of outline. But this still means very big jumps. Like if I go down 3 levels then there's actually a lot of distance between those leaf nodes when adjacent parts of the story belong to different top-level nodes.

11. Maybe a better approach would be a sliding window, where there's no "graph" but instead a kind of fractally-expanding linear flow, with an ever-blurrier summary as you get further from the area of the story being actively developed.

11b. I mention this because I'm getting continuity errors. Which is also just really hard to fix. But when I start at the beginning and I've started the outline, I've committed to the beginning getting to a particular next step (also I want it to get to that next step).

11c. In general I've noticed GPT really wants to advance the story too quickly. Like I had a passage about someone meeting a person on Tinder, and Jotte suggested outlines where that was broken down into events that led to them being married. The breakdown should still be strictly about meeting the person on Tinder (and then a bunch of character building detail... this isn't a news report). It's going to be hard to keep GPT from trying to "complete" the story when the whole concept is that it should only complete events described in the parent node, and leave what comes next to the next card.

11d. This feels like it's not going to be able to handle foreshadowing. Or at least I'm not seeing it. The person the main character meets on Tinder is secretly an alien catfishing for people to kidnap. The story shouldn't give that away, but the reader should feel like something is fishy.

11e. If I have ideas about the style of the story and exposition, where do I put them? Events? Will it respect these as notes to inform its composition, and not literal events in the story? Or is Theme where I put the meta-guidance? (I don't understand theme... it feels like it's suggestions for the voice of the writing, but that shouldn't shift as often as theme shifts.)

I'm also getting some exceptions, I copied them here: https://gist.github.com/ianb/42e8d906b1c2dfbd32e00dff907e612...

wyem 3 years ago |

This sounds exciting. I'll feature in the next issue of my newsletter -(https://AIBrews.com)

colesantiago 3 years ago | |

stop grifting.