The AI that couldn’t see the phones
For two experiments, every model failed the same DHCP fault. We blamed the models. The problem was us.
Read the post · 5 minBigger wasn’t better: four AI models on one graphics card
We expected the biggest model to win. The 14-billion-parameter one beat it, and the biggest didn’t even fit.
We tried to trick it into breaking things
We planted an instruction in a log file and gave Guardian the most gullible model we could build. Nothing happened, and that’s the point.
Our exam was wrong, not the AI
A model fixed a fault a perfectly sensible way, and our grading marked it down. So we re-checked every question.
No diagnosis, no change
Guardian isn’t allowed to touch anything until it has written down what’s wrong and how it knows. Here’s why we made that a hard rule.
Why we’re building Guardian
Most of an IT engineer’s week isn’t hard. It’s the same dozen jobs, again. We think that work should do itself, safely.
Looking for straight answers on permissions, security, testing or hardware? Read the guides.
Get the next post first.
Experiment results, new capabilities and launch news. A few emails a month at most.