Keep Your Ego In Check (Above The Robots)

In order to use LLMs effectively, you need a bigger ego than the AI. That's the secret.

There's a lot of talk about the rise of "GPT slop" on the internet. People use it as evidence that AI shouldn't be a writing tool. I disagree. I think LLMs are great, if you're confident you can write better than the AI.

But why use an AI if you're better than it? The key is understanding human strengths and weaknesses—that's where the AI comes in. Corey Doctorow calls this the Centaur/Reverse-Centaur dynamic.

Humans aren't good at editing, typically. The more we reread something, the less we notice. Reviewing your own work is hard. That's why authors hire editors. It's not just that the editor is inherently better (although that might be true). It's that they have fresh eyes. They haven't reread the same page a hundred times. Fresh eyes catch what the author misses.

Here's how I use AI without creating GPT slop:

1. Never let it write the first draft.

   Why? You want your own voice. Voice can be preserved through editing, but it is near impossible to inject voice into soulless writing. We're also trying to avoid the Reverse-Centaur problem. Humans do the creative part, robots do the boring part.

   I write the first draft myself, keeping it casual. I emphasize my voice, my phrases, metaphors, and flourishes.

2. Make the robot do the boring bit.

   Then I give it to the AI to catch grammar mistakes and clarify the tone. My first drafts are very casual, full of grammatical errors, more like stream-of-consciousness than an essay. The AI's job is to structure my messy ideas.

3. Realize the AI is as soulless as the corpos that birthed it.

   The AI's edit will probably suck. That's the point. When I read what Claude or ChatJippity produces, I think, "That sounds like shit; I liked what I wrote better." This is key. It clarifies what I like in my writing and helps structure my next step.

4. Synthesis!

   I combine the two versions, keeping what I like from the AI and what I like from my original draft. AI excels at formal/bland/corpo writing; but bad at writing anything more human than a HR memo. So, by doing the casual bit and fusing it with the AI’s strengths I can edit way faster.

That is my general writing process these days. In my experience, this dynamic extendeds more broadly, the AI is most useful when doing something that you could do yourself, and probably do better - but don't feel like it. You are most likely to fall for its hallucinations when you are asking it to do something you don't know how to do. For example, I fell into the trap the other day when I asked GPT-o1 to derive some maths for me. The results it spat out where 90% correct, but because I didn't know the maths it took me near a full day to fact check it. That's a lesson for another blog-post though.