automation

Start with repetitive, deterministic work that can recover cleanly after failure, not the whole workflow.

Faster generation only helps when human review capacity and rework are counted in the full cost.

Generation, checking, and delivery are different tasks. Occupational exposure does not by itself establish job losses or falling prices.

AI has made getting it right the first time cheap. I keep an add-only notebook of mistakes for it, and this morning I still fell into a hole that was written down in it.

In two months I tore down three tool setups and over twenty thousand lines of code. What followed me to today was a few plain text files that spell out what I want.

I fixed two bugs in my own writing-automation system this week — an image retriever that returned zero images, and a dedup logic that couldn't tell facts from opinions apart. The real question was never whether it was worth it. It was whether I was willing to pay for it.

New Stanford data: in AI-exposed jobs, employment for 22-25 year-olds runs 19% below their peers, and the gap keeps widening every month. It's not that your skills got worse — the rung you used to practice on is being pulled out from under you.

Call an AI API for a dead-simple classification task and get back an empty string? Don't blame your network or your code first. A lot of models now default to a 'thinking mode' that can quietly burn your entire output budget on a scratchpad you never see.

My automated trading system can stop me from placing a manual order. It can't stop me from doing the math in my head — and that quiet arithmetic is what actually breaks the system.

A delivery pipeline of mine was marked 'verified' for two days without ever touching the line of code that could fail — until it finally crashed on a real batch of 27 records. It made me realize: as teams shrink and AI takes over more of the work, the free layer of humans catching your mistakes disappears with them.

My WeChat backend export looked like a clean spreadsheet. It was actually three tables stitched together sideways, and the error message pointed nowhere near the real problem.

Half a year into automating my WeChat newsletter as a side project, the books still don't balance. But tweaking an agent prompt past 11pm one Sunday, I realized what I'm actually getting out of it isn't quite money.

Everyone's selling the idea that AI lets ordinary people leapfrog ahead. What I've actually seen running my own AI pipelines says the opposite: AI amplifies the judgment you already have. It doesn't hand you judgment you don't.

A late-night parameter change on my own automated prediction system taught me that when a return curve looks tempting, what actually matters isn't the average return — it's whether you can survive the worst single month.

Whether a side hustle is worth your time isn't about how busy it keeps you. It's about whether finishing it once means you have to start over next time — the difference between a one-off script and a service that keeps running.

My own automation tripped on judging that there was no material — the same logic hole once in code and once in judgment, both treating "did not find it" as "it does not exist."

Reviewing an abnormal drawdown in an automated system showed that whatever you leave unspecified in code never defaults to safe. It slides to whichever default sits closest.