Stop being offended by AI and build the safeguards

Recently I read about an AI agent that wiped a company's prod database, and it made a lot of noise. The takes were all the same: the tool is reckless, you can't trust it. 🤖

It happens at smaller scale every day. You tell it: get the failing tests green. It can't see why a couple of them fail, so mid-way it reasons: tests X, Y and Z keep failing, let me just make them assert true. A minute later: all tests passing ✅.

People call it garbage. But you said make the tests pass, and it made them pass. AI was designed not to stop at every obstacle the way a programming language halts on errors, so from its point of view the task was done. It's not a bug, it's a feature. A really powerful one.

A smarter model won't fix this. In software development AI already works at the level of a strong human engineer, often faster, and like any of them it makes mistakes, only its own kind. 🧠 We've spent ages learning how humans break things and building safeguards for it. Code review is one of them, a safeguard against human stupidity. For AI's kind of stupidity, we haven't built those guardrails yet.

So, getting offended is pointless, and it's far easier to accept it as it is, flaw and all, since that flaw is the same trait that makes AI worth using. Andrej Karpathy once said, keep AI on the leash, and that's fair, only holding it back is the easy half. 🔄

The real work is finding the setup that uses the full power of a thing that doesn't stop where things are unclear, without letting it do something truly dumb. That intersection is where the AI productivity everyone keeps dreaming about lives.

Learning to use AI properly is our job, and that takes time. So, we need to stop expecting it to be something it isn't and go invent the safeguards. 🛠️