09What AI can and can't do

Worth knowing the ceiling. Both the bottom one and the top one.

At the bottom - things that take an evening. A one-page site, a Telegram bot, a script that renames a thousand files for you. A couple of hours, and it works.

In the middle - what people do seriously. Paid services, dashboards, bots that watch something and send alerts. Weeks of work, but now within reach of one person. One guy built a database of tractor specifications - literally a tractor reference - and it brings him over forty thousand dollars a month in ad revenue. Boring niche, zero hype.

At the top - actual science. Companies use AI to search for new drugs, and the first such candidates have reached late-stage trials. Here honesty matters: no AI-designed drug has been approved yet, and nine out of ten candidates fail in trials. But the fact that it works at all is reality now, not a promise.

Now what it can't do. It can't want things for you. It doesn't know what your users need. It won't finish the project if you drop it. And it doesn't replace understanding: if you have no idea what you're building, you won't notice when it gets something wrong.