Don’t toss AI slop grenades

I heard this term on a podcast recently (* cited at the end), and the meaning is clear if you have ever received half-baked AI output at work.

I have loved experimenting with AI, and the potential continues to grow. However, there are two ideas that I have been thinking about, in terms of how to work with AI (primarily ChatGPT and Claude), that help me avoid sending slop.

Cognitive offloading means outsourcing your thinking to AI and, in some ways, this is the long-term promise.

Right now, though, if you are doing this with tasks where there is no one right answer, the technology is likely making mistakes that you are not catching.

If you work at a company where AI is heavily used, you have probably been hit with half-baked work, where somebody used AI as a shortcut. Then when you read it, the answer just feels off.

Although AI can help you think, figuring out things of consequence is hard. If you knock something out, send it to a colleague and think, “that was easy,” then you are probably doing it wrong.

While cognitive offloading is something to avoid, confidence without conviction is how I would describe most output from AI.

Recently, I spent a bunch of time researching stocks to invest in. AI was very good at identifying companies that were good businesses, but it struggled with valuations. It confidently provided textbook answers which looked reasonable.

However, when I dug into the details there were a lot of mistakes, and the variety of mistakes could be another post by itself.

Every time I poked on an assumption, AI would agree with me and change the answer. It’s guessing but doesn’t actually know the answer. Think of every answer as its best guess.

I try to take a balanced view of AI. There is a lot of potential, but we are early in the development and there is a learning curve. I suspect that this technology will develop faster than our ability to adapt to it, and that gives me pause.

* I heard the term ‘slop grenade’ from Tobi Lütke on a recent episode of The Knowledge Project