Deep dives into automation, AI technology, and business strategy.
'COMPLETED' was a copy with its history stripped, and the 'optimized' code was byte-identical. One commit, eleven hundred modified files — a record of before I learned, counted as it is.
AI remembers nothing on its own. The smallest possible memory — append the conversation to a file on the server and send only the recent part — plus the trap where saving tokens froze the clock.
I opened three years of receipts and my editor usage stats. What I remembered as switching tools was mostly swapping models in one window — and the thing whose name I could not recall was Ruby on Rails.
Build an AI you can talk to from a messenger in under 100 lines — including the traps you only learn by hitting them: the 4,096-character limit and Markdown parsing failures.
A multi-AI app built in four days, abandoned. Reopening the backup a year and a half later: a single commit. Yet the thing I failed to build now runs every morning.
Back from military service, I couldn't bring myself to write computer science on the double-major form — money and fear. I settled for economics, then downgraded even that to a minor. Twenty years later, unable to write a single line of code, I build something every weekend. An essay on why clumsiness is a privilege, not a defect.
A robot voice, a mic that cut me off, and an AI tutor that changed the subject after every wrong answer — five failures (and their lessons) from growing a 24-word page into a full language study app in one day. Including how to filter AI-generated content with a validator that trusts nothing, and where the copyright lines actually are for exam questions and lecture videos.
It found an email a few days ago; today it couldn't, and offered up an unrelated years-old one instead. I assumed the model had changed. Three plausible misdiagnoses later I found the real causes — a stop button that erased the whole conversation memory, and a 'dropped connection' that was my own bug. Why you should read the record instead of trusting an AI's self-diagnosis.
I was the one who couldn't cut losses; the AI had the discipline. The story of handing an AI real trading within a small budget — after building a four-layer 'cannot do' vault first: budget floor, ledger isolation, single-use passes, and a discipline contract. Including the trap of silently evaporating stop-losses.
Instead of trusting one AI, I sat four different AIs at one screen to grade each other's answers. How a non-developer had an AI assistant build a fact-check council — and the two-day 'but it works in my terminal' bug.
My blog automation, which I thought was humming along, had silently stopped for 4 days. Nobody told me. Why 'no news is good news' is a lie for automation.
I changed two lines of config to tighten security, and the whole site threw 500s. A non-developer learns there's no 'trivial change' — and why 'it's probably temporary' is dangerous.