I started out using GitHub Copilot at work because there was a lot of pressure to use AI, and I was put off by how we were churning through PRs that seemed to work, but having to go back and fix the slop afterwards.
Now I’ve realized that there are skillful ways and unskillful ways to use LLMs, and they can in fact be a useful tool beyond just generating slop. They don’t replace a human thinking critically, but they can automate mundane, routine tasks. They can also summarize text well and suggest options for humans to consider. For example, LLMs reviewing code will often find issues the human reviewers missed.
In addition to coding, I’ve recently been using Qwen locally for screenwriting. It can’t write worth a shit, but it does a good job critiquing my work and pointing out problems with the story structure and the like. For example, I can tell it something like “look at the 7 plot elements described in this MD file and point out where this story does and doesn’t follow this structure”, and the output is quite useful.
While LLMs aren’t the magical silver bullet the tech bros are hyping them up to be, they can still be a useful tool. If they’re just used to generate slop, then no, they’re worse than useless.
It’s not my “coach” any more than random people online would be if I posted it in a forum somewhere and no more than a LLM or a human peer reviewing my code is my “coach”. It provides a different perspective to help me see beyond my own biases with feedback I can accept or reject.
Qwen has obviously been trained on writing books and a ton of screenplays. As an experiment, I changed the character names in a classic sitcom script and it was able to identify the series from the writing style and then it also identified the episode. It’s not useful for doing the actual writing, but it does provide useful feedback based on sophisticated statistical analysis of my work compared to its professionally-written training data.
Explain to me how it’s better than you learning to analyze your own work from a formulaic perspective?
Everytime you choose to use AI, you are choosing NOT to develop an ability of your own. Sometimes, that’s an ability that just tedious to use, other times it might be something you obviously need to do yourself, yet others the ability might be something with a tangential utility you haven’t recognized.
An analogy might be reading music exclusively. Great, now you can play a wide range of music–indisputably beneficial!–but the cost of developing your own ear.
I have read a lot of books and do analyze my work in terms of techniques and principles I’ve studied over the years. However, even top professional writers don’t work in a vacuum. TV writers, for example, have “the room” with a team of professional writers, producers, etc. weighing in on all writing decisions. For indies, you don’t have that luxury, and even getting another human who is good at writing to read what you wrote and share detailed feedback is hard, especially when said humans aren’t getting paid to do it full time. Asking friends and family to critique your writing will often result in them trying to spare your feelings, whereas Qwen will happily rip your work to shreds and not care if it just shit all over your passion project.
Do you have a flat rate sub to Qwen? I’m curious if you fed it something that you personally think is great writing that isn’t prominent training data, that you are intimately familiar with, and what you would make of its analysis?
My fear is two-fold: first, writing is communication between people with shared experiences. An LLM can’t really tell if someone’s going to have an emotional connection to your writing or why or what or how it works. Second, novelty and rule-breaking is highly context dependant. I’d be worried an LLM is merely steering me into probable lanes instead of allowing me to develop my own unique voice.
I haven’t found anything yet that isn’t in its training data that I’d want it to evaluate as a control group, but you’re right that it would be a useful exercise.
Here are some examples of the feedback it has given me:
This plot point hasn’t been “earned” and needs more setup to pay off properly
This dialog is an exposition dump. Find a better way to show, not tell.
This character feels like a vehicle for jokes, and isn’t developed enough.
Most of the advice I’ve gotten so far relates straight back to what I’ve read in writing books and is pretty cut and dry. Some things are a matter of opinion, and I push back when I disagree or when I am deliberately breaking a rule.
Edit:
To your other point, you’re correct that a LLM saying something is good doesn’t mean humans will think so, or vice-versa. A LLM is but one tool in the process, and doesn’t replace real human feedback. For example, with a comedy, do human readers laugh out loud when reading it? A LLM can determine statistically whether something is intended to be a joke and whether the joke is overused, etc., but can’t tell you if the joke is actually funny.
I started out using GitHub Copilot at work because there was a lot of pressure to use AI, and I was put off by how we were churning through PRs that seemed to work, but having to go back and fix the slop afterwards.
Now I’ve realized that there are skillful ways and unskillful ways to use LLMs, and they can in fact be a useful tool beyond just generating slop. They don’t replace a human thinking critically, but they can automate mundane, routine tasks. They can also summarize text well and suggest options for humans to consider. For example, LLMs reviewing code will often find issues the human reviewers missed.
In addition to coding, I’ve recently been using Qwen locally for screenwriting. It can’t write worth a shit, but it does a good job critiquing my work and pointing out problems with the story structure and the like. For example, I can tell it something like “look at the 7 plot elements described in this MD file and point out where this story does and doesn’t follow this structure”, and the output is quite useful.
While LLMs aren’t the magical silver bullet the tech bros are hyping them up to be, they can still be a useful tool. If they’re just used to generate slop, then no, they’re worse than useless.
I’m horrrfied that an LLM is your writing coach.
It’s not my “coach” any more than random people online would be if I posted it in a forum somewhere and no more than a LLM or a human peer reviewing my code is my “coach”. It provides a different perspective to help me see beyond my own biases with feedback I can accept or reject.
Qwen has obviously been trained on writing books and a ton of screenplays. As an experiment, I changed the character names in a classic sitcom script and it was able to identify the series from the writing style and then it also identified the episode. It’s not useful for doing the actual writing, but it does provide useful feedback based on sophisticated statistical analysis of my work compared to its professionally-written training data.
Explain to me how it’s better than you learning to analyze your own work from a formulaic perspective?
Everytime you choose to use AI, you are choosing NOT to develop an ability of your own. Sometimes, that’s an ability that just tedious to use, other times it might be something you obviously need to do yourself, yet others the ability might be something with a tangential utility you haven’t recognized.
An analogy might be reading music exclusively. Great, now you can play a wide range of music–indisputably beneficial!–but the cost of developing your own ear.
I have read a lot of books and do analyze my work in terms of techniques and principles I’ve studied over the years. However, even top professional writers don’t work in a vacuum. TV writers, for example, have “the room” with a team of professional writers, producers, etc. weighing in on all writing decisions. For indies, you don’t have that luxury, and even getting another human who is good at writing to read what you wrote and share detailed feedback is hard, especially when said humans aren’t getting paid to do it full time. Asking friends and family to critique your writing will often result in them trying to spare your feelings, whereas Qwen will happily rip your work to shreds and not care if it just shit all over your passion project.
Do you have a flat rate sub to Qwen? I’m curious if you fed it something that you personally think is great writing that isn’t prominent training data, that you are intimately familiar with, and what you would make of its analysis?
My fear is two-fold: first, writing is communication between people with shared experiences. An LLM can’t really tell if someone’s going to have an emotional connection to your writing or why or what or how it works. Second, novelty and rule-breaking is highly context dependant. I’d be worried an LLM is merely steering me into probable lanes instead of allowing me to develop my own unique voice.
I’m running Qwen on my own hardware.
I haven’t found anything yet that isn’t in its training data that I’d want it to evaluate as a control group, but you’re right that it would be a useful exercise.
Here are some examples of the feedback it has given me:
This plot point hasn’t been “earned” and needs more setup to pay off properly
This dialog is an exposition dump. Find a better way to show, not tell.
This character feels like a vehicle for jokes, and isn’t developed enough.
Most of the advice I’ve gotten so far relates straight back to what I’ve read in writing books and is pretty cut and dry. Some things are a matter of opinion, and I push back when I disagree or when I am deliberately breaking a rule.
Edit:
To your other point, you’re correct that a LLM saying something is good doesn’t mean humans will think so, or vice-versa. A LLM is but one tool in the process, and doesn’t replace real human feedback. For example, with a comedy, do human readers laugh out loud when reading it? A LLM can determine statistically whether something is intended to be a joke and whether the joke is overused, etc., but can’t tell you if the joke is actually funny.