...
Brooks's Law:
"Adding manpower to a late software project makes it later."
...
Maybe the 2026 update to Brooksβs Law should be:
"Adding agents to a project with unclear requirements gets you the wrong answer faster."
That tracks. There are things that LLMs are better at than others, and one of the things they're especially good at is extrapolating from an incomplete specification to produce something that nobody would want.
Earlier today I also saw a story about trying to get ChatGPT to count the teeth on a sprocket, which exemplifies this pretty well. The agent got an unbelievable number and, when asked to label the diagram to show their counting, gave away their mistake when they missed the number 18... but it went further than that, because the agent had also redrawn the original photo to try to make their original answer correct.
Absolutely understandable when you think about how these clankers actually work, but for people who don't have that level of comprehension of their mechanism of action... and who don't know enough about Shimano sprocket cassettes, might not notice the mistake. And therein lies yet another reminder of what I've been saying for a long while: you can't trust an LLM to do a good job of anything you couldn't do for yourself.
And when you try (e.g. a non-programmer asking Claude to make an app for them, a non-doctor asking ChatGPT to diagnose their condition, a non-musician asking Suno to compose them something...) you risk getting something that convincingly looks like it's the right answer... but it's a superficial veneer over a mess that, chances are, has problems you're ill-equipped to identify.
0 comments