Recap from last week
AI models writing style is driving me insane.
AI as an editor is still a net positive, but without careful watching, it quickly becomes a negative.
What have I learned?
ChatGPT has become my favourite writing assistant again which I didn’t expect.
Every model is starting to drive me insane with their editing. As I said last week, you can’t use AI to edit your writing any more than 2 or 3 times before it becomes unrecognisable.
New features I’ve noticed with every AI model - short. Performative. Sentences. Everywhere. Colons: almost everywhere. Em dashes and semi-colons also being used too much (although I like em dashes and think they serve a good purpose.
Regardless of how often you tell a model not to use a particular style of writing, these methods seem so baked in that you can’t tell them to stop. I think this is actually a good thing, because you can’t just copy and paste (I mean, you can, but people will know what happened).
It’s actually very interesting to me that you can’t override these styles of writing so I went down a little rabbit hole and discovered no one is quite sure why these styles of writing are so pervasive. My best guess is that the later models (GPT3.5 didn’t use em dashes all over) were trained on classic books from the 17th and 18th Century and this style of writing has come through.
This then doesn’t account for the new use of short sentences everywhere. Maybe the latest models were trained on reddit comments alone.
Mental models to keep for next week
Never let AI control the process: It became more apparent when doing this for writing, but it holds when you do anything. If the AI runs away with what you told it, you lose any of your own creativity and input very quickly.
Next up
Will be using Claude Opus a lot more this week to continue designing and developing some key features for a new paid version of Agora.