That Human Flavor

AI

I could hear them screaming that the age of humans was over.

Honestly, this was nothing new, even before Astra became the biggest thing of 2026.
Whenever the usual crowd started losing their shit, I pretty much always took it with a healthy dose of skepticism. What exactly is an agent? What’s MCP? What’s a harness? Does any of that really matter? What we care about is the actual takeaway: “So, AI can do everything now?”

It used to be that “Come on, AI still has plenty of shortcomings” was more or less the prevailing opinion. Now, more people are saying the opposite. We already have more than enough bullshit experts running their mouths about AI, and the moves coming out of OpenAI, Claude, and Meta seem to be giving them plenty of backing.

Marketing always pushes ahead. Engineering takes the cautious approach. But that balance only works when the people building the technology fully understand it. At some point, though, even the technical side stopped understanding AI. Why the hell does this work so well?

So I decided to face it myself. I hated the AI slop these things churned out. Agents, MCP, harnesses—I’d criticized the whole lot as empirical, inductive methods with no real analysis behind them. But somewhere along the way, I’d ended up in this ironic position where I was developing AI and couldn’t even make a case for it. I literally make my living off AI, yet I hated every person and company singing its praises. I was starting to feel like a professional complainer.

It’s not like I just dismiss AI, either. I spend around ₩200,000 a month on ChatGPT out of my own pocket, and using it has become second nature, both at work and in everyday life. Recently, I couldn’t figure out which optional tests to choose for my company health checkup, so I just pasted the list into ChatGPT and asked for recommendations. On a trip to Japan, I took a photo of a menu full of sake I couldn’t read and asked ChatGPT to explain how to pronounce each name and what kind of sake it was, in detail. At this point, I don’t really need a service with some “specific” area of expertise. I just need an LLM with enough coverage that, whatever subject I know nothing about, ChatGPT can look things up, work out the context, combine it with its pretrained knowledge, and give me an answer with minimal hallucination.

Of course, the limitations are pretty clear too. It misread the sake menu and wrote about drinks that weren’t even on it. But I can live with that. Problems with tool calling, like OCR, or misreading visual context aren’t things you can fix just by doing context engineering on the LLM.

Anyway, since I use AI this much, I wanted to try writing with it. I’ve been writing for a long time, but I wanted to see whether AI could actually get my voice and write about subjects I don’t know much about while keeping the style and flow of my own writing.

The verdict? The writing just doesn’t have much flavor. It more or less picks up the “style” through in-context learning, but then it throws in jokes that don’t quite land, or digs up a bunch of references and piles on way too many of them. Things I wouldn’t do in my own writing. Why does this happen?

How a Person Goes About Writing

If a person—me, for instance—were writing something using references, I’d study the material, internalize it, and then write based on my own understanding. That’s a little different from how an LLM writes. When we encounter new knowledge, we learn it before we write about it. Giving an LLM context isn’t the same as having it learn. It’s more like an open-book exam: you’re adding context to something that’s already good at generating answers. Imagine trying to summarize a textbook in your field. If you haven’t really understood the book, opening it up and trying to explain a particular concept is going to be a pain in the ass.

This is where the difference comes in. Honestly, we aren’t smarter than “AI.” At least when it comes to pretrained knowledge.

To explain a concept we don’t know well, we have to spend enough time with it to get familiar with it. And to write about it clearly enough that someone who’s never even heard of it can follow along, we have to fully understand it ourselves. But AI—an LLM—doesn’t have to do that. Its pretrained knowledge contains more information, across more subjects, than you’d imagine. It can put together perfectly plausible sentences from the information already learned and encoded in its parameters. Even without doing additional training on that textbook—continual learning—it can produce a pretty much flawless summary of unfamiliar information. Maybe this is less about imitating how humans understand things and more about using whatever information is available. Because people can’t do it like that.

Maybe that difference in how text gets produced is where the gap between writing that feels human and writing that reeks of AI comes from. A person probably wouldn’t write by stringing together reference after reference—academic papers aside. And if there were something they didn’t properly understand, they’d tear the draft apart and start over, possibly several times. Even after all that rewriting, they’d still make the usual mistakes: explaining something incorrectly, mixing up terms or definitions, that sort of thing.

Because we aren’t perfect, we prepare for the possibility that we’ll get things wrong. An LLM doesn’t. Whatever reasoning process or harness you give it, an agent is ultimately just a machine that immediately spits out the best answer it can from a probability distribution, within whatever environment it’s been given. AI doesn’t keep learning during inference.

Will I Keep Writing with AI?

Writing with AI did help quite a bit. When I write, I don’t actually go straight to asking AI to write the piece for me. I pick a topic and the material I want to work with, then throw in my argument and opinions. From there, I work toward a rough shape for the piece, whether that involves taking a critical look at my argument or not, and flesh it out. It’s like starting with a storyboard before you actually shoot a TV show.

One of the biggest letdowns when writing with AI was the “pictures.” The right images and captions are essential to a piece with some flavor. An image can carry a joke or a bit of information that gets the author’s thinking across in a really clever way. And because the author puts their own way of understanding something into that image, it can be much easier to follow and more concise than the conceptual diagrams an LLM churns out through a generative model. Choosing and making those images well might matter just as much as the writing itself, if not more. But no matter how good an image model gets at rendering text or making realistic pictures, an agent fundamentally doesn’t write by learning a concept anew. So the images it produces can never contain “the effort to understand, and the individuality that comes from it.” This will probably remain something AI can never solve.

Any work, once it has taken some kind of shape, has something imperfect tucked away inside it, however close to perfect it might look. That imperfection exists in things people make, and it exists in things AI makes. We’re used to the imperfections people produce. We aren’t used to the ones AI produces. I don’t think that’s going to change.

Oh, so am I going to keep writing with AI? Yeah, I am. But I don’t think I’ll ever love it.

One Last Thing

The people who get the most out of AI are the ones who clearly understand its limitations, also know exactly what it can do, and put it straight to work in whatever workflows make sense. Don’t get sucked in by dumb marketing about AI solving some supposedly impossible math problem. Just use it, study it, and get yourself ready. Whatever the architecture, you’ll always need the ability to pick apart the details and figure out why things fail. Whether I’m someone who can identify those failures comes down to how much I’ve thought things through and how hard I’ve tried to understand what’s actually going on.

AI’s going to replace everything? Fuck that.