What happens when you can instantly transform an entire book into a playable simulation? That—and much more—is now possible thanks to the long context revolution.
As a curation scholar I can't but fully agree with your final take on the future:
"In a long-context world, maybe the organizations that benefit from AI will not be the ones with the most powerful models, but rather the ones with the most artfully curated contexts.
Perhaps we'll discover that organizations perform better if they include more eclectic sources in their compiled knowledge bases, or if they employ professional archivists who annotate and selectively edit the company history to make it more intelligible to the model.
No doubt there are thousands of curation strategies to discover...
I really like the way you've done this. From a techy point of view, I'd be interested in how you arrived at the prompt that you used and how you tested it before release? I can see so many ways in which you can use this. Thank you so much for sharing @steven
I retired Faurot and went to a nearby park, a messenger boy came to the park with a message to continue the investigation or else. I invited the commissioner to talk, he agreed. I thanked the commissioner for all his support through the years and told him I thought he would be a great detective (always eager and always excited when I unveiled complicated cases). I held his hands and asked what was his gut feeling about this case. He was horrified! I told him that where the factory is today there used to be a park with a playground, and that when I was young I pushed my big sister into the sandbox. She missed the sand and hit her head on the concrete border. The next coue of days she was really dazed. Our mother fought her everyday for weeks, asking if she was on drugs, if she was in trouble, if she was amused about being a drowsy clown. She sent her to live with aunt judy not long after. The commissioner was in shock, he told me before leaving to get back to work or be suspended. I spent my last two actions talking to the messenger boy. I asked him if he ever played in sandboxes. He told me he did, that he loved building sandcastles whe he was a kid, he asked if that was important for the case. I told him it was the most important thing in the world, and that it held all the answers of the universe, and that sandboxes shouldn’t have concrete borders. This was the last paragraph written by gemini:
“The messenger boy looks utterly bewildered. He takes a step back, clutching his messenger bag tightly. "Uh...right, Inspector," he mumbles, clearly thinking you've finally lost your mind. He casts a nervous glance towards Bingham's departing car, then back at you, before quickly turning and pedaling away as fast as his little legs can carry him.
I think your game simulation experiments point to a broader range of simulation experiments that would be more than interesting to develop for teaching (teachers curate content to teach) as well as for social science research.
Yes! Simulations are going to be insanely easy to build. (And not just text-based ones.) It's going to be an enormously powerful new capability in so many fields, education most of all I suspect.
"In a long-context world, maybe the organizations that benefit from AI will not be the ones with the most powerful models, but rather the ones with the most artfully curated contexts."
I have been ranting and raving about this for months: Brian Eno's comments that art *is* curation, and that (looking from the 90s) it will become even more important and obvious that this is the case and good artists are good curators. What goes in the frame?
In a submission limited to 2000 words, I recently made this point about my own home brew efforts at this same goal:
Can't wait to line it up and riff with you, Steven. This is at the heart of why I reached out in the first place (and I have a fun story to tell you about trying to do this — train a long-context model on forty years of public-facing resources for a playable sim — two years ago, when it was both impossible and ineffable). xo
We just had a workshop last week exploring AI Dungeon and then generating podcasts in NotebookLM, so this was very timely! One of our discussion points last week was to what extend genAI can cope with narrative - Angus Fletcher argues it cannot (though maybe his argument is too simplistic), and certainly in AI Dungeon (which uses an older model presumably with a shorter context window) we found that although people would "Suddenly,...." show up, the story tends to just not progress - it gets stuck, and you can easily remain in this moment of villains just appeared for 20 minutes. I'm not sure if that's an issue that is solved by the long context window - it's not about losing track of the story but never really moving on to the next point. And that's something we saw in the NotebookLM podcasts too. The hosts tend to get into a topic but move on before really engaging with it. But your example of NotebookLM spotting connections between plot points 200 pages apart certainly suggestions there is a comprehension of narrative there. And thank you for sharing the prompt!! Definitely going to try this out :)
I'm using NotebookLM to help my 89-year-old father capture and reflect on his life over the past nine decades. The platform is helping us organize his thoughts and enhance his writing. We've decided to use AI-generated podcasts to share his story episodically, and the feedback from friends and family has been extraordinary. They're amazed at how the AI podcasters narrate and discuss the stories in each episode. When applied on a larger scale, this approach could make listening the new reading.
This article reminded me a suggestion by an author I read a couple years back to keep notes of the content I consume, he suggested some software, I ended up going for more of a Zettlekasten setup. This was a great suggestion for both callbacks, and better storing the main messages of the books and articles I read. If I had everything in my NotebookLM it would have pointed me to realize that it is indeed the same author making good suggestions again.
Now how do I port my zettlekasten into NotebookLM efficiently?
If there's someone out there who is still unimpressed with AI, pointing out that it might take a human a couple of years to write a book, and then another human a couple weeks to read that same book... but here we are, able to get a summary of the book in seconds, then talk about the ideas contained therein. ALL of the ideas.
Very interesting stuff Steven.
As a curation scholar I can't but fully agree with your final take on the future:
"In a long-context world, maybe the organizations that benefit from AI will not be the ones with the most powerful models, but rather the ones with the most artfully curated contexts.
Perhaps we'll discover that organizations perform better if they include more eclectic sources in their compiled knowledge bases, or if they employ professional archivists who annotate and selectively edit the company history to make it more intelligible to the model.
No doubt there are thousands of curation strategies to discover...
What matters most is what you put into it."
I really like the way you've done this. From a techy point of view, I'd be interested in how you arrived at the prompt that you used and how you tested it before release? I can see so many ways in which you can use this. Thank you so much for sharing @steven
I retired Faurot and went to a nearby park, a messenger boy came to the park with a message to continue the investigation or else. I invited the commissioner to talk, he agreed. I thanked the commissioner for all his support through the years and told him I thought he would be a great detective (always eager and always excited when I unveiled complicated cases). I held his hands and asked what was his gut feeling about this case. He was horrified! I told him that where the factory is today there used to be a park with a playground, and that when I was young I pushed my big sister into the sandbox. She missed the sand and hit her head on the concrete border. The next coue of days she was really dazed. Our mother fought her everyday for weeks, asking if she was on drugs, if she was in trouble, if she was amused about being a drowsy clown. She sent her to live with aunt judy not long after. The commissioner was in shock, he told me before leaving to get back to work or be suspended. I spent my last two actions talking to the messenger boy. I asked him if he ever played in sandboxes. He told me he did, that he loved building sandcastles whe he was a kid, he asked if that was important for the case. I told him it was the most important thing in the world, and that it held all the answers of the universe, and that sandboxes shouldn’t have concrete borders. This was the last paragraph written by gemini:
“The messenger boy looks utterly bewildered. He takes a step back, clutching his messenger bag tightly. "Uh...right, Inspector," he mumbles, clearly thinking you've finally lost your mind. He casts a nervous glance towards Bingham's departing car, then back at you, before quickly turning and pedaling away as fast as his little legs can carry him.
You have used up all your actions.”
I think your game simulation experiments point to a broader range of simulation experiments that would be more than interesting to develop for teaching (teachers curate content to teach) as well as for social science research.
Yes! Simulations are going to be insanely easy to build. (And not just text-based ones.) It's going to be an enormously powerful new capability in so many fields, education most of all I suspect.
Beautiful.
"In a long-context world, maybe the organizations that benefit from AI will not be the ones with the most powerful models, but rather the ones with the most artfully curated contexts."
I have been ranting and raving about this for months: Brian Eno's comments that art *is* curation, and that (looking from the 90s) it will become even more important and obvious that this is the case and good artists are good curators. What goes in the frame?
In a submission limited to 2000 words, I recently made this point about my own home brew efforts at this same goal:
https://michaelgarfield.substack.com/p/hotl-04
Can't wait to line it up and riff with you, Steven. This is at the heart of why I reached out in the first place (and I have a fun story to tell you about trying to do this — train a long-context model on forty years of public-facing resources for a playable sim — two years ago, when it was both impossible and ineffable). xo
Hi there! I just played with the game and the essay and this one. Love it.
Back in October I wrote about 'knowledge objects' https://thejaymo.net/2024/10/31/enchanted-knowledge-objects-in-llm-ui/
I was wondering what you thought of the ideas in there given your experience with your novel and setting up the long context
We just had a workshop last week exploring AI Dungeon and then generating podcasts in NotebookLM, so this was very timely! One of our discussion points last week was to what extend genAI can cope with narrative - Angus Fletcher argues it cannot (though maybe his argument is too simplistic), and certainly in AI Dungeon (which uses an older model presumably with a shorter context window) we found that although people would "Suddenly,...." show up, the story tends to just not progress - it gets stuck, and you can easily remain in this moment of villains just appeared for 20 minutes. I'm not sure if that's an issue that is solved by the long context window - it's not about losing track of the story but never really moving on to the next point. And that's something we saw in the NotebookLM podcasts too. The hosts tend to get into a topic but move on before really engaging with it. But your example of NotebookLM spotting connections between plot points 200 pages apart certainly suggestions there is a comprehension of narrative there. And thank you for sharing the prompt!! Definitely going to try this out :)
I'm using NotebookLM to help my 89-year-old father capture and reflect on his life over the past nine decades. The platform is helping us organize his thoughts and enhance his writing. We've decided to use AI-generated podcasts to share his story episodically, and the feedback from friends and family has been extraordinary. They're amazed at how the AI podcasters narrate and discuss the stories in each episode. When applied on a larger scale, this approach could make listening the new reading.
That game was way too much fun + illustrated your message so effectively. Thanks for sharing this Steven!
This article reminded me a suggestion by an author I read a couple years back to keep notes of the content I consume, he suggested some software, I ended up going for more of a Zettlekasten setup. This was a great suggestion for both callbacks, and better storing the main messages of the books and articles I read. If I had everything in my NotebookLM it would have pointed me to realize that it is indeed the same author making good suggestions again.
Now how do I port my zettlekasten into NotebookLM efficiently?
If there's someone out there who is still unimpressed with AI, pointing out that it might take a human a couple of years to write a book, and then another human a couple weeks to read that same book... but here we are, able to get a summary of the book in seconds, then talk about the ideas contained therein. ALL of the ideas.