I know that you didn't write this. I am a SOTA 0-shot classifier of your slop.

If you are reading this, you were most likely sent here after you shared a document with me and I viewed it unfavorably.

Perhaps it was an awesome auth middleware that you vibe coded last night. It could have been a work plan to get into Carnegie Mellon, produced by the hottest new college admissions AI. Maybe it was even a message that was supposed to be heartfelt and deeply personal, but instead got mangled through 26 TFLOPs of DeepSeek-based revisions.

You should also know that, while I am disappointed in you, this happens often enough that I have this publicly-facing web document expressly for this purpose. In this regard, I’m not trying to single you out for criticism.

I know that you didn’t write that. I know it was touched by an AI, and you will never be sneaky enough to fool me. I’ve been digesting LLM outputs by the megatoken since I first got off the GPT-3 API waitlist in ‘21. I spend $700 a month on 10+ pro memberships (Claude Max is the darling right now, but that won’t last forever) and 50% of my day job is understanding how all of the major players’ models respond to an insanely wide and unfairly difficult set of test queries and tasks on our company’s internal AI platform.

I’m also into a lot of weird use cases. At my first job out of grad school, I tried to convince my manager that GPT-3 could be used to figure out whether or not a building had air-conditioning vents. That same month, I tried to get it to make 1000 top-grade headlines from The Onion and write bedtime stories for my child based on Wikipedia articles. You can bet your bottom dollar that I know when your Python class methods were cooked up by Qwen Coder.

You should also know that I learned to read early and skipped first grade. Consequently, I was the perennial runt of the classroom and spent all my time reading as a child.

This is all to say one thing - I will always know when you didn’t write it. Alternately, perhaps I really like you and think you have potential, so I am sending you this message prematurely so that you don’t accidentally trigger a peeve of mine.

Why do I seem to be taking this so personally? Well, it’s mostly an issue of time. My life is busy. I have children, a demanding job, and personal aspirations. When I want information to help you, enable you, or supervise you, I want it in the most efficient format possible. Invariably, this means that you spend at least a modicum of effort whittling down and throwing out the parts that aren’t important.

LLMs are very good at producing output, but they are not a substitute for you.

Put differently, it feels like you are pushing work onto my plate by forcing me to second-guess and reread your text because you didn’t take the time to craft it carefully in the first place. I would also like you to note that this is entirely separate from the issue of hallucinations. I am hoping that I am not working with someone sufficiently careless and dumb to let outright fabrications slip through.

Can I give you another hot take?

I am used to working with people sufficiently disciplined and productive enough that, for them, simply reviewing the info and writing the characters is just like breathing, and it’s not even worth trying to automate it. These are people who understand that generation is part of learning and you should always let yourself be the voice. As you write, your brain is actively re-engaging and even sometimes rewiring itself to better grapple with the reality of what it has seen. Don’t deprive yourself of this reward.

I’ll carve out a few small exceptions to this rule. If we’re working in English and it’s clearly your second language, I understand that you want the substance of the message to shine despite its style. I’ll cut you a break. I will hold my nose despite all the text smells because I genuinely want to see your best work in the best light possible.

Am I a Luddite? A bit, perhaps. I am clearly an LLM enthusiast, though sadly, my local deployment can only handle models sized at 8B or less. I want you to vibe code like a lunatic. I want you to grill Opus Max on O(N) methods for approximating the middle eigenvalues of a sparse positive matrix.

But when you come back to me, put your personal touch on it. Read, reread, review, and summarize. You’re not just a pipe from SQL to Outlook or from textbook to problem set. You are a rich soup of learnings and experiences, and you’ve been simmering for decades. Please let me try some of it.