A Trio of LLMs: Writing and editing with genAI
Getting feedback, generating verse, researching and other fun things with genAI.
This week I interacted with three different LLMs: Claude, ChatGPT and Perplexity.
If you’re following this newsletter, you’ll have heard of Claude and ChatGPT, but maybe not of Perplexity.
I’ve played with Perplexity a little bit with mixed results, but a couple of people have suggested I try it lately, so I’ve given it another go. What did I think of it during these new attempts? Read on to find out…
One other reflection: using each of these models feels different, as if interacting with different people. I have heard that different models are being trained differently and at some point, they will start to diverge in that way.
For now, I use Claude for work with my own writing, and ChatGPT for anything that requires broad knowledge, or access to recent data. Perplexity should be the go-to for recent data, but I’ll let you know why I’m not that much into it later. The way in which they present their answers, their tone and even the layout of the page is very different. It’s interesting to see what you’re drawn to.
Using Claude for Feedback
They say Generation Z and many others who regularly use social media seek instant gratification. This usually refers to “likes” and comments etc - I suppose also to the fact that we can easily buy digital products online and have them ready to use within seconds.
For me, I’ve recently noticed how much another kind of instant gratification is helping me to write: immediate feedback from Claude.
I’m in the “I just need to finish this book” phase with the Shakespeare-remote book. (I was going to call it “All the World’s a Workplace”, but I think the title doesn’t have the right focus, so we’ll stay with the Shakespeare-remote label for now.)
I’ve finally finished the first draft of the chapter on hybrid work, a topic for which I don’t have much interest - but you can’t really write a book about remote work in 2024 without mentioning the H word.
After finishing the draft, I printed it out and did an extensive edit myself. Then, I started making the changes to the manuscript in Scrivener, the writing tool I use. As I finished each section of the chapter, I turned to Claude and:
- Asked for some feedback either on the whole section or on any paragraphs where the ideas were not solid, or where I wasn’t sure my point was clear. I am very lazy with my prompt, which is a shame, because I know if I gave the bot more specific instructions, I might get better results. Here’s one example of what I asked:
Here's the first paragraph of a chapter on Hybrid. Each chapter is about 5,000 words., Before cleaning it up, can you tell me whether I should elaborate more on the last paragraph.
or
Here's the next section. First of all, can you tell me what you think of the section at the end, where I make the link to Shakespeare?
- Next, I’d act on the feedback - only that which made sense to me, of course.
- If Claude made a suggestion I particularly liked, I asked it to draft some lines about it, like here, where I pulled out one of its recommendations:
“Consider expanding on why you think the balcony scene might not be the best vehicle for discussing hybrid work. This could provide an interesting contrast between romantic ideals and workplace realities.”
This is a great point, can you give me a few lines about it I can use?
- When each section had all the content it needed, I’d prompt Claude for the final step:
Ok, here is the section again, which will be followed by the balcony scene. Can you clean it up for fluency please? Please remember to keep my voice.
I have to say that Claude is doing quite a good job of not changing my voice/writing style. In fact, sometimes the changes to the text are very subtle, which I think is a good thing. It does mean that the sections which I’ve asked it to write for me are in danger of standing out - but I think the bot has got the hang of the style I’m looking for, and anyway, I always go through the text and change any words I don’t use in my writing. Well, I keep some I don’t use when they’re a better choice, and I hope the vocabulary will stick with me for next time.
I’m going to move onto using ChatGPT, but if you would like more details on any of the above, let me know.
Using ChatGPT for Artistic Tasks
This Shakespeare-remote book started when, for fun, I fed ChatGPT Shakespeare monologues and asked it to rewrite them describing a concept linked to remote work. I was very impressed by the results.
But I did have days when the bot got stuck, not just ignoring the original rhyming scheme, but also getting stuck in rhyming loops it couldn’t get out of. I worked out that the bot didn’t see it’s own rhyming scheme, and so couldn’t correct it.
I had put aside ChatGPT for these kind of writing tasks, but have recently come back to it. And I have to say, it’s far better than Claude at this.
Here’s an example:
ME:
In this extract, I want the adaptation to have Juliet ask how come the term and concept of hybrid has been so widely adopted, and Romeo explain that hybrid is a good way to move forward, precisely because it has so much acceptance. Otherwise we would all go back to the office. Here’s the original.
JULIET By whose direction found'st thou out this place?
ROMEO By love, who first did prompt me to inquire;
He lent me counsel and I lent him eyes.
I am no pilot; yet, wert thou as far
As that vast shore wash'd with the farthest sea, I would adventure for such merchandise.
CHAT GPT:
Certainly! Here's the adapted ending with that focus:
---
**JULIET**
By whose design has hybrid found its place?
**ROMEO**
By need, which first did prompt this course to grow;
It lent us reason, and we lent it trust.
I am no seer; yet, had it not prevailed,
We'd all return to office-bound confines,
But with this path, our future's wisely paved.
---
This adaptation maintains the rhythm of the original while emphasizing the acceptance and practicality of the hybrid work model as a way to move forward.
It actually gave me two different versions, without prompting it to.
Whereas here is Claude’s attempt - although I never gave it the original, just asked it to follow on from the rest of the adapted version:
JULIET
Then let us embark upon this untrod path,
With hearts united, and with purpose clear.
Though challenges may come, we'll stand as one,
And prove that hybrid work can be the way.
ROMEO
With thee by my side, I fear no obstacle.
Together, we shall make this vision real,
And show the world what modern work can be:
A balance struck, 'tween freedom and the field.
Using LLMs for Research
Let’s now move onto my little interaction with Perplexity.
According to their website, “Perplexity is a free AI-powered answer engine that provides accurate, trusted, and real-time answers to any question.”
It provides you with the links to the references it uses, pretty much like when ChatGPT accesses the web. I have a line in my chapter on hybrid that says:
”Most organizations describing themselves as "hybrid" require, expect, or recommend that people work from the office 2 or 3 days a week.”
I asked Perplexity:
”Can you give me different examples, the more recent the better, of organisations that have set up hybrid workplaces, and either "require", "recommend" or "expect" a number of days in the office. Please give me some uk examples.”
It gave me some examples, but referenced all of them with the same article.
I then went to ChatGPT and prompted it to answer the same question, to see what type of references it would give me. At first, it didn’t search the web, and gave me an answer based on its data from August 2023. I opened a new chat and gave it the same prompt, adding: “Can you browse the internet to answer this question”
It did (browse the internet) and pulled data from the article Perplexity had used as its main reference. (In case you’re interested, it’s this one: https://www.travelperk.com/uk/blog/top-hybrid-work-trend-stats-from-global-companies/ )
Unfortunately, the article doesn’t mention any of the companies cited by either bot!
That’s ok, it means I’ll need to turn back to good ol’ google and do a bit of research there.
Regarding Perplexity, I’ve only used it a few times, whereas I’ve been working with Claude and ChatGPT for over a year now. It takes time to get to grips with these tools. I might invest some time in getting to know the programme better in the future. For now, there’s only room for two LLMs in my heart…
Two more things:
- I asked my mother for a cake recipe. She took a photo of her handwritten recipe, which was in Spanish, and I uploaded to ChatGPT , and asked it to transcribe the text. I uploaded the image upside down, so the bot told me:
The image appears to be upside down. Let me rotate it and transcribe the text for you.
I have rotated the image. You can view it here. I'll now transcribe the text.
Here's the transcribed text from the image:
It later said,
This appears to be a cooking recipe in Spanish. If you need further assistance with this, feel free to ask!
In its first try, it got a few things wrong, so I gave it some context:
It’s not quite right. Here it is again. (This time I had rotated it myself.) Yes, it’s in Spanish, for coca, or sponge cake.
That helped it transcribe the whole recipe correctly.
(Something I’ve just learned: yesterday I was able to access the rotated image, but today, that’s no longer possible.)
Listening Recommendation
In this newsletter, Spiralling Creativity, I mainly talk about my writing, and using LLMs to help with the creative process. But of course, generative AI is broader than that (and “AI”, much more).
If you’re interested in what’s happening around LLMs and audio, listen to the first episode of the podcast Shell Game. You can find it in your podcast app, or here:
It’s a fun episode - it’s also a bit scary…
That’s all for now.
Let me know if there’s anything else you’d like to know about any of the above.
And do share your own creativity adventures - whether tech-enabled or not!
And if you’ve enjoyed this post and would like to see future ones land in your inbox on Fridays, don’t forget to subscribe!
Thank you for reading!