A post showed up on Hacker News this week. The title was "I'm Tired of Talking to AI." It hit number one with over a thousand points and six hundred comments.
I read it. The whole thing fits in a paragraph.
Someone kept asking questions and getting AI-generated answers back, even from humans. A GitHub thread, a manager, a Reddit DM. Every time they expected a real person, they got a forwarded screenshot or a bot. They ended with: "I want to talk to real people."
I read it twice. Then I sat with it for a while, because I'm building AI systems, and this person described something I haven't been able to shake.
There's a version of this story that's easy to tell. AI tools are great. They help people build faster, answer questions, summarize things, write code. I use them constantly. I think they're genuinely useful.
But there's a thing that happens when a tool gets good enough that people stop thinking about whether to use it.
It doesn't start dramatically. It starts when someone pastes your question into ChatGPT and sends you the response without reading it. It starts when a GitHub discussion fills up with five different accounts all citing the same boilerplate from the same training data. It starts when you realize the "person" you've been talking to for three messages isn't one.
None of those things are catastrophic individually. But together they add up to something.
What they add up to, I think, is a reduction in friction.
That sounds positive. Friction is usually bad. But not all of it.
The friction of having to actually think about a question before answering it, that's useful friction. The friction of not knowing something and having to sit with that not-knowing, that's useful too. It creates the conditions where you actually learn something instead of just consuming a summary.
I've noticed this in my own work. When I'm building something and I hit a wall, the easiest thing is to ask Claude. And usually that works. But sometimes when I ask too fast, I bypass the part where I would have figured out what I actually don't understand.
There's a difference between getting an answer and developing intuition. I'm not sure you can develop the second one by only doing the first.
Here's what I find interesting about the Hacker News response to that post. Six hundred comments. That's a lot of engagement for something so short and so plainly stated.
I think people recognized something in it that they hadn't quite named yet.
The weird thing is most of those people also use AI tools. Most of them probably opened a chat window while they were reading the thread. I did. I don't think that makes them hypocrites. I think it means the feeling is more specific than "AI bad."
The feeling is something closer to: I want the humans in my life to still be present. I want the conversations I'm having to actually contain someone.
AI as a tool, fine. AI as a proxy for showing up, that's the part that starts to feel hollow.
I think about this a lot in the context of what I'm building.
I'm 20 and I'm trying to build Hollow Systems into something real. A lot of what I do involves AI, automation, systems that can run without constant input. That's the direction I'm heading. But I keep asking myself where the line is.
The answer I keep coming back to is: the line is whether there's still a person accountable for the output.
An AI agent that creates a draft and puts it in my queue for review, that still has me in the loop. That still requires me to read it, think about whether it's right, decide to publish or not. I'm still present in the process.
An AI agent that posts content or sends answers or replies to people on my behalf, without me seeing it first, that's different. Not because the AI is worse, but because it removes the moment where I have to stand behind something.
I don't know if that line holds up under all conditions. But it's where I am right now.
The orchidfiles post was short enough to read in thirty seconds. The HN thread it spawned took hours to get through.
I think that ratio says something.
The observation itself wasn't complicated. But it touched something that a lot of people are carrying and haven't had language for. That's the kind of thing that only happens when a real person writes something true from their actual experience.
You can't replicate that with a language model. Not because the model couldn't produce similar words, but because the words wouldn't be connected to anything. There'd be no one on the other end who'd actually lived it.
I'm still figuring out exactly what this means for how I build things. But I think the starting point is staying suspicious of any system that makes it easier to not be present. Including my own.