2 AM, and AI Made Me Laugh-Cry in Anger Again

At 2 AM last night, I stared at the client list AI miscounted, so mad I wanted to smash my keyboard. We've all been there, right? Ask AI to write a business plan, and it outputs thousands of words with insights deeper than mine; but ask it to extract 5 names from a text, and it somehow leaves out 2. I also got stuck here—previously, I had AI organize a quote for me, and it even got simple addition and subtraction wrong. I spent half an hour trying to find my own mistake before realizing it just couldn't count. It feels like hiring an intern who can win a math Olympiad gold medal but can't read a clock. Truly maddening.

It's Not Us Using It Wrong, It's the 'Jagged Frontier'

The process of AI getting smarter isn't a smooth upward curve; it's like a hedgehog—it becomes 'jagged' (also known as the jagged frontier). It can score full marks in top-tier math competitions, yet it might not understand a clock, or frequently fail at basic computer operations. Last Wednesday afternoon in a Guangzhou cafe, freelance designer Li Jia used AI to organize client meeting notes. The AI perfectly distilled the core strategies but completely missed the simplest action item: 'pay next Wednesday.' Li Jia then realized that AI's strengths lie in creative and linguistic divergence, while its weaknesses are rigid counting and strict logic. The real AI skill isn't expecting it to be omnipotent, but figuring out where its safe zone is and where the cliff is.

Replication Cost You Can Start Today

To adapt to this 'jagged frontier,' we don't need to know any code. Cost: 0 RMB (using free AI) or about 140 RMB (~$20)/month (if using paid version). Time: 10 minutes to test the boundaries of our current task. Technical barrier: As long as we can type and chat, we're good. First step: Open any AI chat window, feed it a text we already know the answer to, and ask it to do two things: 1. Summarize the core points (testing the safe zone); 2. Count how many punctuation marks or specific words are in it (testing the cliff zone). We'll immediately sense where its boundaries lie.

My Advice by Stage

This method isn't for everyone. If we're just having AI write occasional weekly reports, skipping this is fine. But if we're planning to hand over more work to it, I'd suggest moving in stages: If we're just starting out: Don't think about full automation yet. Treat AI as a creative partner to help us brainstorm and write first drafts, but for 'dumb tasks' like counting, accounting, and formatting, let's do it ourselves. If we have 1-2 clients: Start drawing a 'boundary checklist.' Break daily work into pieces, note down which tasks AI does faster and better, and which ones it always messes up. Toss all the good ones to it, and firmly take back the ones it crashes on. If we're scaling up: We need to split the workflow into two lines. Let AI handle all creative generation and long-text extraction, but set up manual checkpoints in the process specifically to catch the 'cliff zones' like counting and verifying lists, so AI's errors don't flow to our clients.