Seth's Blog : Hallucinations and human work

Hallucinations and human work

“AI is brilliant and it can do everything.”

“AI hallucinates sometimes and it can’t be trusted.”

“AI is a trick, a clever way to induce people to believe it’s human-like, but it’s not.”

It turns out that AI hallucinates all the time. Sometimes, these hallucinations are useful, worth interpreting as helpful contributions, and sometimes, not so much.

In some senses, then, AI isn’t that smart.

But neither are we.

At work, we spend very little time accurately synthesizing new information and creating breakthroughs. We mostly do tasks, simple inputs and outputs, based on little understanding of the overall system in which we live.

AI is good at line by line coding, because that’s an iterative process that piles up useful hallucinations into a working whole.

AI is less good at conceptual system architecture problems. And most programmers are less good at this as well. Often dramatically less good.

AI is good at multiple choice questions and banal copywriting. So are most people.

In less than 24 months, we’ve seen LLM apps become better and better at things we never thought they’d be able to do. Human work, it seems, is in retreat.

In fact, it’s not a retreat. It’s a chance to advance. The same way the steam shovel put a focus on the hard work of architecture, project planning and smart choices about what to build (instead of the hard brute force work of shoveling), AI is doing the same for indoor work.

It doesn’t really matter that AI doesn’t “know” what it’s doing. Most of the time, we don’t either.

Human work is any work an AI can’t do (yet.)