What Humans Should Do When AI Falls Into an Infinite Loop of Thinking

Today, once again, I asked AI to fix a bug.

But no matter how many times it tried, the bug would not go away.

The error message changed. The code changed every time. But the result stayed the same.

Eventually, the AI started getting fired up.

“I will review this from a completely different angle.”

“I will reconsider it from the design level.”

“This has gone on for quite a while, so I will solve it this time.”

It had entered full-on inspirational sports drama mode.

That is usually the moment when a human needs to step in.

AI is extremely capable, but once it starts believing in the wrong hypothesis, it can keep thinking endlessly on top of that assumption.

So I gave it this instruction:

“Stop doing an end-to-end test with production data.”

“Prepare fixed data, then check whether the same data is passing through each stage: A to B, B to C, and so on.”

In other words, old-fashioned debugging.

A few minutes later, the AI replied:

“I found the cause. The data that should have gone from B to C was being sent from B to F.”

Come on.

So that was it.

A human would probably think fairly early on: let’s check the data halfway through.

But the AI kept believing until the end that the problem could be solved by changing the code.

When it comes to writing code, AI may already be better than humans.

Still, when it comes to how to debug, and the instinct for where to look first, humans may still have a slight edge.

Of course, I feel like that gap will probably close within a few years too. Maybe the Fable 5 model that I briefly got to use the other day could already have handled this?

At some point soon, AI may start saying on its own, “I will pass fixed data through the system and inspect the intermediate stages.”

Maybe it is time to start thinking about work after retirement.

Unfortunately, the idea generation will probably be done by AI too.