I like how you went back and checked out Code Interpreter. Some people have a fear of AI so they would like to avoid it as much as possible but not you. You're exploring it in every way and checking how AI can be beneficial to writers. This Code Interpreter looks like the feature Sudowrite might need. Some people say they would like to write a Chapter 2 on Sudowrite while having it remember everything that happen in Chapter 1. If something similar to Code Interpreter can be implemented on Sudowrite it would be great.
I used Code Interpreter to create a personality/alignment & backstory generator for NPCs in TTRPGs. But that could also be used in build more dynamic characters in a book as well. I have it run off two tables the 1st Table is sets of traits that have no morale alignment like say "Calm" or "Hyperactive". This Table would list Calm as the Lawful trait and Hyper Active as the Chaotic trait. Now it would then pick one of the two as they are opposites. If it picks "Calm" it scales from 1-9, one being the lowest still give the personality type as Hyperactive but this trait was gained in a positive way for backstory purposes, and 9 is the Calmest a person could be, but still from Positive backstory. And if it picked Hyperactive, it would have done the same but reversing it a 1 being Calm and 9 being Hyperactive but the trait was gained in a negative way backstory wise. 5 in either would be Neutral on that trait being neither fully a calm person but not Hyperactive one either and would still cause it to be made by a negative or positive aspect of their backstory. Then the next chart has Morale based traits like Honesty, still rated 1-9. 1 being the lowest means basically they always lie and that's an evil trait and 9 is they always tell the truth being a good trait. 5 still falls in to neutral. Then it takes all the ratings of the trait from both charts, give an average for each, and creates a numbered pair to decide their alignment. So any number pair (9, 9) - (7, 7) would all be in all Lawful Good. (6, 9) - (4, 7) would be Neutral Good, and (3, 9) - (1, 7) being Chaotic Good. the rest of the Alignment chart following the same structure. Now it would know how to use whatever info you give it about that NPC/Story Character to create how they talk, act, and if they're a good or evil type person. While using the individual trait names from the Lawful/Chaotic chart to influence the created backstory is built, and Individual traits on the morale chart to influence their actual behavior. With this you can have multiple Character/NPCs with similar traits but all act differently based on background, alignment, and personality. Now for a book you could skip the averaging and numbered pairs, as that's only used for deciding Alignment, but still use the rest for helping flesh out a character for a fictional story. Edit: For those interested in how it's used to decide alignment, it creates an 81 point Alignment Chart with 9 points in each section. This is a simplified version of that chart. By using the 9 point Rating system no matter how many traits you add it will always give an average of 1-9 on both of the trait charts to create the numbered pairs. | L | N | C -------------------------------------------------------------- G| (9,9) - (7,7) | (6,9) - (4,7) | (3,9) - (1-7) -------------------------------------------------------------- N| (9,6) - (7,4) | (6,6) - (4,4) | (3,6) - (1,4) -------------------------------------------------------------- E| (9,3) - (7,1) | (6,3) - (4,1) | (3,3) - (1,1)
Also, consider that you can compile your results after evaluating them and create new documents to create a new zip file to continue your workflow with the new data.
Okay, now that I've recovered from cracking up over the 5,000-deck starship. I've been doing this in Claude, and I find that even with the larger context window, if you give it a lot to read at once, it tends to gloss over important details and then make stuff up. Haven't tested it in GPT-4.
@@TheNerdyNovelist I don't have this problem with shorter work, but if I ask it to write a synopsis of something 25k or more words it glosses over stuff and then makes stuff up to fill in the gaps. It's still pretty useful, but I'll be happier when I can just upload a document and get a synopsis without having to hand-hold the AI and remind it to stick to the source text.
I had an interesting conversation with Claude last week. I wanted it to look at something that I thought it might slap my wrist over because of its content filter, and I told it that I was writing a novel and gave it a couple of sentence summary of what I was going to show it and asked it whether I thought that would trigger the content filter. It came back with something like because it’s fiction, there is no content filter. So maybe the trick is to just be very explicit when you start out that “this is fiction” and then you can get it to look at anything. I didn’t try making it write anything, just read something I had written.
My man if you look to the right of the output text, there is a clipboard icon that you can click to have it COPY it to the clipboard for you so you don't have to highlight it every time.
For the time being, whether you use Claude or ChatGPT to work with documents depends on how important logic is to the process you want to perform. FFA's video using the police report would generally require stronger logic than what Claude is capable of at this juncture.
With this code interpreter, can I just make 3 text files containing the character sheets for the players, and do I have to include that in every message? As in, after a couple of messages, will it also forgot about the text file if it was previously attached, the same way it forgets everything else after a while? Or is the uploaded text file permenantly referencable? In my testing scenario I utilize 4 different chats, one AI is the world host and 3 others are the players in a simple story game. First the world building AI is tasked with creating a world and some plot/mission (small, like a war time negotiation between 3 houses/factions). After that is done, it wil start writing up the story up until any of the 3 characters need to talk. I will then feed the story into the first AI player, and have it generate an action or dialogue, and then inturn feed that into the other AI players, until the talk is resolved. Then I compress the conversation that was held and feed it into the world building AI, which will then further the story.
I really like the extra functionality that code interpreter brings, however I feel that it is very hit and miss for fiction writing at the moment. Giving it large amounts of text it doesn't cope well. While it does have some ability to read txt files it doesn't do it in the way you would expect. Explicitly when you dive deeper it tells you that it doesn't have access to language models in it's environment. So, for example, asking it to summarize a story passed in a file. In will typically take a strategy of extracting only some sentences and (in most cases successfully) hallucinating them all together. There must be a way to get around this, but I haven't found it yet.
As an example, I passed it a story and asked it to find all the times one of the characters was described physically. Diving deeper, what it actually does is create a generic list of words that could be used to describe people and then checks those keywords against sentences with the characters name and the descriptive words.
Another example, passing in a lot of text, and asking it, what are the chapter titles, it can find them (simply looking for the words chapter and assuming the rest of the line is the title. Asking then to summarize each chapter it will completely hallucinate it together based on the titles. If you ask it straight out of the box to summarize it will try and try again eventually revealing its lack of language libraries inside the code model
What do you think is the best thing for aspiring writers and creators to be doing with these tools right now? I feel like, while it's possible to create novels right now, that's not necessarily the best use of time and resources. I feel like worldbuilding, story planning, etc. is much more important and something these tools excel at. Once they are able to recall and process way more tokens, then it will be easy to create quality novels. But when that time comes, those who are best prepared with lots of creative material will be able to generate the best and most products. Just my two cents.
It depends on what you need. For many, it will be the prose, and it's actually not bad at that, when done correctly. Especially with the help of tools like Sudowrite. I personally don't use it for much outlining, because that's my favorite part of the process, so I do most of the work myself. But I use it a lot for brainstorming and such. So it's really up to the person. And I don't think it's necessary to be able to recall or process more tokens in order to write the prose. Not if it's working from a good outline, where you act as the continuity director and make sure everything is specified there.
Thanks for this, your content is excellent. Regarding this one, is there any reason why you need to export the world-building document and then re-import it? Couldn’t you just say to ChatGPT ‘Thank you for the world-building document. Now based on this document give me 5 ideas for short stories about…?’ P.S. Am I the only one who says ‘thank you’ to ChatGPT?
That was just an easy way to create an example to use most like likely. But If he had already had the file manually created or used one from a previous chat that would make the upload necessary. But the main use is the multiple text files in a zip folder, allowing it access to multiple files for reference.
also, It helps to guarantee you are working from the right context tokens. and allows you to restart it later or in multiple LLM chat threads, like if comparing output between multiple LLMs.
@@TheNerdyNovelist I keep looking for a discord link, i have an idea for a video i think you'd be great with, and a prompt for writers to use to help them get feedback on their work that is helpful to their process. This comment from you inspires another good video idea. "Why would you want to be able to reuse something like this over and over again?"
FYI, Future Fiction Academy is amazing, and even when my course comes out, it will be quite different from what FFA offers, which is an INSANE value for the price, only $4 per lab, when you add up how many labs they do each week. Highly recommend: nerdynovelist.com/go/futurefictionacademy
I would think Claude 2 is a better application? It's supposed to have a 100k token window, enough for an entire novel. I would reserve Code Interpreter for writing Python.
Yeah, plain text is way better than PDF, especially when it comes to long PDFs, and because you don't have images for the AI to look at, it is better to use txt files. It can do PDF, but most of the time, it gets things wrong and quickly starts to hallucinate.
The first idea it came up with "The underdeck uprising" is basically the plot for Silo. Ive been using Code interpreter on some of my larger coding projects recently and one of the main gripes I have with it is it's lack of memory persistence. After a while it just starts completely forgetting essential parts of the current conversation and instruction which is essential for coding and obviously is for writing novels. Im sure this will improve over the next year or so but in it s current state it can be frustrating.
I know that this isn't the point of the video, and I will comment again to give my actual thoughts on the content. My inner science nerd is in TEARS of laughter. Five THOUSAND decks on a ship 60 Km wide and 20 KM long would make the ship roughly the size of SIXTEEN Earths and would require like our entire solar system's mineral resources to even build a scaffold for. I legit want to see this ship in someone's novel now, and I want a cogent explanation of how it was constructed.
Wait a minute. I'm not sure what calculations you are using, but 60 km by 20 km is roughly 38 miles by 12 miles. And 5000 decks an average of 40 feet high each would only equal 200,000 feet, or roughly 32 miles.
@@GenderPunkJezebelle999 I just grabbed a number. It could be ten feet, or 50 feet; this is why I said that I didn't know what calculations are being used. Either way, the earth is much larger than the ship is. Still impressive however.
@@ericgranberg8893 I don't know, man, this was a week ago, and it was a passing comment on a thread I thought was funny. I can't tell you off the top of my head what precisely I was thinking, but I'm happy to admit I made a mistake.
I like how you went back and checked out Code Interpreter. Some people have a fear of AI so they would like to avoid it as much as possible but not you. You're exploring it in every way and checking how AI can be beneficial to writers. This Code Interpreter looks like the feature Sudowrite might need. Some people say they would like to write a Chapter 2 on Sudowrite while having it remember everything that happen in Chapter 1. If something similar to Code Interpreter can be implemented on Sudowrite it would be great.
That would be cool!
I used Code Interpreter to create a personality/alignment & backstory generator for NPCs in TTRPGs. But that could also be used in build more dynamic characters in a book as well. I have it run off two tables the 1st Table is sets of traits that have no morale alignment like say "Calm" or "Hyperactive". This Table would list Calm as the Lawful trait and Hyper Active as the Chaotic trait. Now it would then pick one of the two as they are opposites. If it picks "Calm" it scales from 1-9, one being the lowest still give the personality type as Hyperactive but this trait was gained in a positive way for backstory purposes, and 9 is the Calmest a person could be, but still from Positive backstory. And if it picked Hyperactive, it would have done the same but reversing it a 1 being Calm and 9 being Hyperactive but the trait was gained in a negative way backstory wise. 5 in either would be Neutral on that trait being neither fully a calm person but not Hyperactive one either and would still cause it to be made by a negative or positive aspect of their backstory.
Then the next chart has Morale based traits like Honesty, still rated 1-9. 1 being the lowest means basically they always lie and that's an evil trait and 9 is they always tell the truth being a good trait. 5 still falls in to neutral.
Then it takes all the ratings of the trait from both charts, give an average for each, and creates a numbered pair to decide their alignment. So any number pair (9, 9) - (7, 7) would all be in all Lawful Good. (6, 9) - (4, 7) would be Neutral Good, and (3, 9) - (1, 7) being Chaotic Good. the rest of the Alignment chart following the same structure. Now it would know how to use whatever info you give it about that NPC/Story Character to create how they talk, act, and if they're a good or evil type person. While using the individual trait names from the Lawful/Chaotic chart to influence the created backstory is built, and Individual traits on the morale chart to influence their actual behavior. With this you can have multiple Character/NPCs with similar traits but all act differently based on background, alignment, and personality.
Now for a book you could skip the averaging and numbered pairs, as that's only used for deciding Alignment, but still use the rest for helping flesh out a character for a fictional story.
Edit:
For those interested in how it's used to decide alignment, it creates an 81 point Alignment Chart with 9 points in each section. This is a simplified version of that chart. By using the 9 point Rating system no matter how many traits you add it will always give an average of 1-9 on both of the trait charts to create the numbered pairs.
| L | N | C
--------------------------------------------------------------
G| (9,9) - (7,7) | (6,9) - (4,7) | (3,9) - (1-7)
--------------------------------------------------------------
N| (9,6) - (7,4) | (6,6) - (4,4) | (3,6) - (1,4)
--------------------------------------------------------------
E| (9,3) - (7,1) | (6,3) - (4,1) | (3,3) - (1,1)
Wow that’s pretty cool! Super in depth.
4:55 If you click the clipboard in the top corner, it will copy all the markdown so that you can save the formatting.
Good to know!
Also, consider that you can compile your results after evaluating them and create new documents to create a new zip file to continue your workflow with the new data.
Yep
Okay, now that I've recovered from cracking up over the 5,000-deck starship. I've been doing this in Claude, and I find that even with the larger context window, if you give it a lot to read at once, it tends to gloss over important details and then make stuff up. Haven't tested it in GPT-4.
Interesting. Claude does tend to take more liberties.
@@TheNerdyNovelist I don't have this problem with shorter work, but if I ask it to write a synopsis of something 25k or more words it glosses over stuff and then makes stuff up to fill in the gaps. It's still pretty useful, but I'll be happier when I can just upload a document and get a synopsis without having to hand-hold the AI and remind it to stick to the source text.
I had an interesting conversation with Claude last week. I wanted it to look at something that I thought it might slap my wrist over because of its content filter, and I told it that I was writing a novel and gave it a couple of sentence summary of what I was going to show it and asked it whether I thought that would trigger the content filter. It came back with something like because it’s fiction, there is no content filter. So maybe the trick is to just be very explicit when you start out that “this is fiction” and then you can get it to look at anything. I didn’t try making it write anything, just read something I had written.
That works sometimes. But not others.
I was kind of able to do this too. I somehow get Claude 2 to write a Military action sequence. Tried again this morning, it blocked it lol.
I'll check out the clip feature in Claude 2.
Cool!
My man if you look to the right of the output text, there is a clipboard icon that you can click to have it COPY it to the clipboard for you so you don't have to highlight it every time.
Yep I’m aware. Old habits.
For the time being, whether you use Claude or ChatGPT to work with documents depends on how important logic is to the process you want to perform. FFA's video using the police report would generally require stronger logic than what Claude is capable of at this juncture.
Agreed.
With this code interpreter, can I just make 3 text files containing the character sheets for the players, and do I have to include that in every message? As in, after a couple of messages, will it also forgot about the text file if it was previously attached, the same way it forgets everything else after a while? Or is the uploaded text file permenantly referencable?
In my testing scenario I utilize 4 different chats, one AI is the world host and 3 others are the players in a simple story game. First the world building AI is tasked with creating a world and some plot/mission (small, like a war time negotiation between 3 houses/factions). After that is done, it wil start writing up the story up until any of the 3 characters need to talk. I will then feed the story into the first AI player, and have it generate an action or dialogue, and then inturn feed that into the other AI players, until the talk is resolved. Then I compress the conversation that was held and feed it into the world building AI, which will then further the story.
Thank you for today for writers
It's my pleasure
Pleas figure out how to properly sync your audio and video tracks.
I really like the extra functionality that code interpreter brings, however I feel that it is very hit and miss for fiction writing at the moment. Giving it large amounts of text it doesn't cope well. While it does have some ability to read txt files it doesn't do it in the way you would expect. Explicitly when you dive deeper it tells you that it doesn't have access to language models in it's environment. So, for example, asking it to summarize a story passed in a file. In will typically take a strategy of extracting only some sentences and (in most cases successfully) hallucinating them all together. There must be a way to get around this, but I haven't found it yet.
As an example, I passed it a story and asked it to find all the times one of the characters was described physically. Diving deeper, what it actually does is create a generic list of words that could be used to describe people and then checks those keywords against sentences with the characters name and the descriptive words.
Another example, passing in a lot of text, and asking it, what are the chapter titles, it can find them (simply looking for the words chapter and assuming the rest of the line is the title. Asking then to summarize each chapter it will completely hallucinate it together based on the titles. If you ask it straight out of the box to summarize it will try and try again eventually revealing its lack of language libraries inside the code model
Sorry this wknd example is a bit of a repeat but chatgpt is certainly inventive in deciding how to tackle a problem 😁. Love the channel.
What do you think is the best thing for aspiring writers and creators to be doing with these tools right now? I feel like, while it's possible to create novels right now, that's not necessarily the best use of time and resources. I feel like worldbuilding, story planning, etc. is much more important and something these tools excel at. Once they are able to recall and process way more tokens, then it will be easy to create quality novels. But when that time comes, those who are best prepared with lots of creative material will be able to generate the best and most products. Just my two cents.
It depends on what you need. For many, it will be the prose, and it's actually not bad at that, when done correctly. Especially with the help of tools like Sudowrite. I personally don't use it for much outlining, because that's my favorite part of the process, so I do most of the work myself. But I use it a lot for brainstorming and such. So it's really up to the person. And I don't think it's necessary to be able to recall or process more tokens in order to write the prose. Not if it's working from a good outline, where you act as the continuity director and make sure everything is specified there.
Thank you.
You’re welcome
Thanks for this, your content is excellent.
Regarding this one, is there any reason why you need to export the world-building document and then re-import it? Couldn’t you just say to ChatGPT ‘Thank you for the world-building document. Now based on this document give me 5 ideas for short stories about…?’
P.S. Am I the only one who says ‘thank you’ to ChatGPT?
That was just an easy way to create an example to use most like likely. But If he had already had the file manually created or used one from a previous chat that would make the upload necessary. But the main use is the multiple text files in a zip folder, allowing it access to multiple files for reference.
also, It helps to guarantee you are working from the right context tokens. and allows you to restart it later or in multiple LLM chat threads, like if comparing output between multiple LLMs.
This. It’s mostly a way to reuse information that you may already have over and over again in different situations.
@@TheNerdyNovelist I keep looking for a discord link, i have an idea for a video i think you'd be great with, and a prompt for writers to use to help them get feedback on their work that is helpful to their process.
This comment from you inspires another good video idea. "Why would you want to be able to reuse something like this over and over again?"
I can't find Code Interpreter? It's not listed under my beta options.
It’s now seemlessly integrated into ChatGPT Plus.
Nice shirt!
Thanks!
FYI, Future Fiction Academy is amazing, and even when my course comes out, it will be quite different from what FFA offers, which is an INSANE value for the price, only $4 per lab, when you add up how many labs they do each week. Highly recommend: nerdynovelist.com/go/futurefictionacademy
Dude you have a button on top for copying output from ChatGPT :) just a quality of life suggestion
I would think Claude 2 is a better application? It's supposed to have a 100k token window, enough for an entire novel. I would reserve Code Interpreter for writing Python.
Yes and no. Technically it can hold more information but Code Interpreter is WAY better at following directions and understanding structure.
I've actually found Claude2 of late to be even better than Code Interpreter for analyzing and interacting with TXT files, PDFs etc.@@TheNerdyNovelist
What is that high pitched buzz?
Not sure. Though I only recently got an editing program that lets me easily remove noise.
My Chat GPT interface doesn't look like this at all. I have no settings feature to update.
It's only for Plus members. These are the feature that paying for it gives you.
Are you subscribed to ChatGPT Plus? If not, then you won't see any of these features.
IS IT also Working with PDF? Because there ist also a gpt Plugin creating a PDF Out of a conversation..Just an idea
We’ve found that AI is not as good at reading PDFs which I why I used a TXT file.
Yeah, plain text is way better than PDF, especially when it comes to long PDFs, and because you don't have images for the AI to look at, it is better to use txt files. It can do PDF, but most of the time, it gets things wrong and quickly starts to hallucinate.
I don't know if anyone ever talks about this but Code Interpreter and Plugins both have larger token sizes than default.
Yeah I've noticed this, though it's hard to figure out how much more.
Nice. :)
👍
The first idea it came up with "The underdeck uprising" is basically the plot for Silo.
Ive been using Code interpreter on some of my larger coding projects recently and one of the main gripes I have with it is it's lack of memory persistence. After a while it just starts completely forgetting essential parts of the current conversation and instruction which is essential for coding and obviously is for writing novels. Im sure this will improve over the next year or so but in it s current state it can be frustrating.
I know that this isn't the point of the video, and I will comment again to give my actual thoughts on the content. My inner science nerd is in TEARS of laughter. Five THOUSAND decks on a ship 60 Km wide and 20 KM long would make the ship roughly the size of SIXTEEN Earths and would require like our entire solar system's mineral resources to even build a scaffold for. I legit want to see this ship in someone's novel now, and I want a cogent explanation of how it was constructed.
Haha I didn’t even think of that. I think I read it as 5000 rooms. Like a giant cruise ship.
Wait a minute. I'm not sure what calculations you are using, but 60 km by 20 km is roughly 38 miles by 12 miles. And 5000 decks an average of 40 feet high each would only equal 200,000 feet, or roughly 32 miles.
@@ericgranberg8893 I think I missed where it said that the decks were 40 ft high, but that makes less sense to me if so.
@@GenderPunkJezebelle999 I just grabbed a number. It could be ten feet, or 50 feet; this is why I said that I didn't know what calculations are being used. Either way, the earth is much larger than the ship is. Still impressive however.
@@ericgranberg8893 I don't know, man, this was a week ago, and it was a passing comment on a thread I thought was funny. I can't tell you off the top of my head what precisely I was thinking, but I'm happy to admit I made a mistake.
Monopoly money
?