Dec 16, 2023·edited Dec 16, 2023Liked by Eleanor Konik
Eleanor, the authenticity of your writing about your experiences as an author using an LLM brought a huge smile on my face. Because - yea - I get it. As a read down your post I kept murmuring - yep, uh hun, me too, RIGHT. Thank you for posting this. I'm going to reference this to others in a future post for people 'trying' to understand AI for writing. Thank you for just being so 'you' in your writing.
Personally, LLMs are an invaluable tool for manipulating existing content. They are much much less useful to generate new one. The only exception is when I let them generate a bunch of random stuff to find some inspiration (but this is just a fancier version of the "random tables" or "random prompts" lists I used years ago).
I haven't -- realistically I'm moving very slow on trying out new things right now, because with the new baby and the job and whatnot I haven't had much opportunity to play around with things that aren't *guaranteed* to increase my productivity. Just about the only new thing I've added is Raycast, and that mostly just for basic text expansion that I'm already familiar with (I use espanso on Windows).
(starting like a prompt, hehe) As a mother of 2 years old, my biggest problem with LLMs is that I don't have any patience to create prompts. Maybe it's me being trained in ol' days of google when you needed to be very specific and keyword-driven with searches. Either way, my experience with LLM is dismal 😅 I spend enough energy on real developing neural network aka my son that I can't stand being that specific with ChatGPT. Thus, thanks for providing prompt examples!
Wow, some old readers! I am 79 - my first Dick and Jane book was in cuneiform...
I have a book of poems, 675 pages (AFTERNOON MUSIC, by Birrell Walsh - published on Amazon) and like all poetry it has some obscurities. Not too many. I asked https://notebooklm.google.com, Google's new LLM, to find poems about death. It did; and summarized them often adequately. EG:
Poem:
Comfort
October 17, 2016
I hear that you are dead
fallen in the mountain dawn.
So many beloved gone...
But comfort, that when I come too
knowing you
you will have put the coffee on.
And NotebookLM did this very creditable commentary:
---
The speaker in the poem "Comfort" by Birrell Walsh expresses a sense of grief and loss upon hearing that their friend has died.
The speaker reflects on the many loved ones who have passed away and finds solace in the thought that when they too pass, their friend will be there to welcome them with a cup of coffee.
The poem's title, "Comfort," suggests that the speaker finds comfort in the idea of being reunited with their friend in the afterlife.
The simple act of making coffee evokes a sense of warmth and hospitality, suggesting that the speaker's friend will be there to provide comfort and support, even in death.
Overall, the poem conveys a sense of hope and reassurance, suggesting that death is not the end but rather a transition to a new beginning.
---
This latest google AI is SO much better than BARD, Google's unfortunate early entry.
I recall Heinlein's THE MOON IS A HARSH MISTRESS in which prisoners in a lunar penal colony sneak messages past the AI that runs the place by speaking with deep ambiguity. I am not sure that will work with the AI's we have now.
I have that XKCD stowed in a folder for when a friend gets too worked up over something on the 'Net. (Yeah, and for me to refer to.)
I'm bemused by all the tools for fiction writers that use LLMs to generate text for them. I've used a few different tools to write a paragraph for me when I've been stuck writing non-fiction and it's always the act of cleaning up the resulting mess that finally gets me unstuck.
I can't imagine what kind of mess they'd make trying to write something entertaining.
It's definitely one of the best XKCDs. That, the color chart, the political makeup of past legislatures, and the good password one are, if not precisely my favorites, then the ones I refer to most often.
LLMs as far as I can tell are horrid at writing good fiction, and worse -- they tend to suck in very particular formulaic ways. Though chopping off the first and last paragraph makes the output a bit more tolerable.
Although I'm 70 years old I still enjoy your stories about tots and tribulations. I'm a devoted Obsidian user and started using LLMs to tidy text and conduct research, discriminatingly. In addition to being useful and informative, reading your Substack makes me feel much younger.
I just tried to use ChatGPT to calculate the weight of one of the columns that line the nave of Cologne's cathedral. I thought this would be an easy shortcut. I was wrong. First, Chat didn't have/couldn't find all the data yet, somehow, I found it on the web. Second, once I feed the correct data to Chat, its answer was off by two orders of magnitude.
Introduction: The author shares her personal experience of using Learner-Led Learning Systems (LLMS), a new approach to education that empowers learners to design their own curriculum and learning goals.
What is LLMS?: LLMS is a framework that helps learners create their own learning maps, which are visual representations of their learning journey, including the topics they want to explore, the resources they need, and the evidence they produce.
Why use LLMS?: The author argues that LLMS can benefit learners by fostering autonomy, agency, curiosity, creativity, and collaboration. She also provides examples of how LLMS can be applied to various domains, such as history, science, and fiction writing.
How to use LLMS?: The author explains the steps involved in creating a learning map, such as defining a driving question, choosing a format, finding resources, and creating artifacts. She also shares some tips and tools for making the process easier and more effective.
Eleanor, the authenticity of your writing about your experiences as an author using an LLM brought a huge smile on my face. Because - yea - I get it. As a read down your post I kept murmuring - yep, uh hun, me too, RIGHT. Thank you for posting this. I'm going to reference this to others in a future post for people 'trying' to understand AI for writing. Thank you for just being so 'you' in your writing.
Thank you for the kind words! I definitely strive for honesty in my articles, so I'm glad that comes across 😅
Thanks for writing this. I love the examples of the prompts.
Personally, LLMs are an invaluable tool for manipulating existing content. They are much much less useful to generate new one. The only exception is when I let them generate a bunch of random stuff to find some inspiration (but this is just a fancier version of the "random tables" or "random prompts" lists I used years ago).
I used to love playing "prompt bingo" with random pages from TV Tropes -- what a great analogy!
I'm not sure if you've tried Brainstory yet (brainstory.ai), but that's been the most useful tool for me lately in my own writing!
I haven't -- realistically I'm moving very slow on trying out new things right now, because with the new baby and the job and whatnot I haven't had much opportunity to play around with things that aren't *guaranteed* to increase my productivity. Just about the only new thing I've added is Raycast, and that mostly just for basic text expansion that I'm already familiar with (I use espanso on Windows).
(starting like a prompt, hehe) As a mother of 2 years old, my biggest problem with LLMs is that I don't have any patience to create prompts. Maybe it's me being trained in ol' days of google when you needed to be very specific and keyword-driven with searches. Either way, my experience with LLM is dismal 😅 I spend enough energy on real developing neural network aka my son that I can't stand being that specific with ChatGPT. Thus, thanks for providing prompt examples!
Hope they're helpful!
Wow, some old readers! I am 79 - my first Dick and Jane book was in cuneiform...
I have a book of poems, 675 pages (AFTERNOON MUSIC, by Birrell Walsh - published on Amazon) and like all poetry it has some obscurities. Not too many. I asked https://notebooklm.google.com, Google's new LLM, to find poems about death. It did; and summarized them often adequately. EG:
Poem:
Comfort
October 17, 2016
I hear that you are dead
fallen in the mountain dawn.
So many beloved gone...
But comfort, that when I come too
knowing you
you will have put the coffee on.
And NotebookLM did this very creditable commentary:
---
The speaker in the poem "Comfort" by Birrell Walsh expresses a sense of grief and loss upon hearing that their friend has died.
The speaker reflects on the many loved ones who have passed away and finds solace in the thought that when they too pass, their friend will be there to welcome them with a cup of coffee.
The poem's title, "Comfort," suggests that the speaker finds comfort in the idea of being reunited with their friend in the afterlife.
The simple act of making coffee evokes a sense of warmth and hospitality, suggesting that the speaker's friend will be there to provide comfort and support, even in death.
Overall, the poem conveys a sense of hope and reassurance, suggesting that death is not the end but rather a transition to a new beginning.
---
This latest google AI is SO much better than BARD, Google's unfortunate early entry.
I recall Heinlein's THE MOON IS A HARSH MISTRESS in which prisoners in a lunar penal colony sneak messages past the AI that runs the place by speaking with deep ambiguity. I am not sure that will work with the AI's we have now.
I have that XKCD stowed in a folder for when a friend gets too worked up over something on the 'Net. (Yeah, and for me to refer to.)
I'm bemused by all the tools for fiction writers that use LLMs to generate text for them. I've used a few different tools to write a paragraph for me when I've been stuck writing non-fiction and it's always the act of cleaning up the resulting mess that finally gets me unstuck.
I can't imagine what kind of mess they'd make trying to write something entertaining.
It's definitely one of the best XKCDs. That, the color chart, the political makeup of past legislatures, and the good password one are, if not precisely my favorites, then the ones I refer to most often.
LLMs as far as I can tell are horrid at writing good fiction, and worse -- they tend to suck in very particular formulaic ways. Though chopping off the first and last paragraph makes the output a bit more tolerable.
Although I'm 70 years old I still enjoy your stories about tots and tribulations. I'm a devoted Obsidian user and started using LLMs to tidy text and conduct research, discriminatingly. In addition to being useful and informative, reading your Substack makes me feel much younger.
Oh this is wonderful to hear, thank you!
I just tried to use ChatGPT to calculate the weight of one of the columns that line the nave of Cologne's cathedral. I thought this would be an easy shortcut. I was wrong. First, Chat didn't have/couldn't find all the data yet, somehow, I found it on the web. Second, once I feed the correct data to Chat, its answer was off by two orders of magnitude.
Moral of the story: check Chat's work.
Your article encouraged me to do some experimenting, so I asked Bing's Copilot to "please summarise https://eleanorkonik.substack.com/p/a-personal-take-on-using-llms". It's response (below) was not close.
Here is a summary of the web page:
Introduction: The author shares her personal experience of using Learner-Led Learning Systems (LLMS), a new approach to education that empowers learners to design their own curriculum and learning goals.
What is LLMS?: LLMS is a framework that helps learners create their own learning maps, which are visual representations of their learning journey, including the topics they want to explore, the resources they need, and the evidence they produce.
Why use LLMS?: The author argues that LLMS can benefit learners by fostering autonomy, agency, curiosity, creativity, and collaboration. She also provides examples of how LLMS can be applied to various domains, such as history, science, and fiction writing.
How to use LLMS?: The author explains the steps involved in creating a learning map, such as defining a driving question, choosing a format, finding resources, and creating artifacts. She also shares some tips and tools for making the process easier and more effective.
Ahahaha wow Bing did a really bad job there. I bet Claude or chatGPT would do better but for real that is impressively off base.