bver

Writing with AI

· 4 min read· updated

human-written

I read back what I (well, we, as in: me and my AI friend) wrote last week, namely the Devoxx recaps. I wasn’t happy with these posts. The language is awkward, the structure is meh. That’s because I used AI to write them for me.

Now, it’s easy to go “of course, you used AI to write this, so you need to expect slop”, and that was my first instinctive reaction as well. But is it so simple? It rarely is.

So here’s what I did to write those posts. I took notes and asked Claude to read them, interview me, and then turn those notes into a post. What you see here and here is the result.

If the note-taking experience was like a breath of fresh air, the resulting blog posts: not so much. The result was both meh and awkward, like I said.

That said, this blog is at least partially meant to become a record of experimental steps I took, as I continue learning more about AI. As such, I think there’s much value in honestly showing the current state I’m at. To see if I can improve this over time.

And I do think I can do better: it’s obvious others are doing better at writing text - and code! - using AI than I am. These days, it’s safe to assume that a lot of what we read is probably written with AI, and we don’t even notice. So good AI text must exist.

So if that’s the case, the question becomes: what am I doing wrong? Why does the latest-and-greatest Claude model not give me better text out of the box? This must mean the delta is in what I’m doing, not in what I’m using.

This reminds me of what Renato Cavalcanti said in his excellent talk last week at Devoxx:

On average, you get average code if you don’t guide [the LLM].

And later, he quoted Charafeddine Mouzouni, from the book How to Think With AI:

Your knowledge is what pulls the model away from the center

Allow me to paraphrase: on one side, you have the model. The more money you throw at AI, the better the model, and that’s true. This makes it look like a pay-to-win game.

However, there’s also how you use the model. Which instructions do you give it? How do you make sure the output is what you expect? That requires skill and effort, and I like to believe that’s what will distinguish the best from the rest, whether we’re talking about software developers or writers. Because if that’s not the case, and turns out to be a pay-to-win game, then we may just as well all pack our bags and go do something else, like singing or wood working.

So where does this leave us? I’m planning on writing more posts with AI, either assisted by an LLM, or actually written by an LLM.

Going forward, I’ll be tagging my posts as such as well:

(Oct 17 – Note: I edited these rules after writing I cut my finger writing this post; the old rule called any post which was initially drafted by AI as AI-written, regardless of how much I edited that text afterwards.)

Being clear about this serves three main purposes.

First off, I think it’s just the honest thing to do. Everyone and their mom (including mine!) are using AI for pretty much anything these days, so it’d be dishonest to not at least mention I did too. For the record, I did not write this particular text using AI, albeit for different reasons: I don’t want this blog to become a stream of AI-generated diarrhea. Not until I’ve fixed my text generator haha. Until then, I plan to manually write posts as much as possible, if only to give you a small pause in the stream of posts.

A second reason is that I’ll be able to track my - should I say our? - progress. In a few weeks or months time, I’ll be able to look back and (hopefully 🤞) see - and be able to identify - the improvement.

Thirdly, this will make it easy for my future text generator to know what posts I, Ben, wrote. So that it can - maybe - mimic my style and wording better.

I kinda hope that later posts will be much better, and that my embarrassment over these first few AI-generated or AI-assisted posts will only grow going forward.