It’s kind of weird that I haven’t written here in almost seven days, but I’ve actually been doing a lot.
For those who don’t know, I run a non-alcoholic newsletter called The Dri Edit. We have about 600 subscribers and up until now, we’ve pretty much only existed through email.
We’re now moving every article over to a more digital system where the stories we’re finding and writing can live on our website. This will hopefully mean we can automate some of the processes and eventually have an opportunity to generate revenue through things like affiliate links.
Shea is getting ready to launch merchandise for No Booze Babes, so I’ve been trying to help her as much as I can. P.S. It’s launching tomorrow 9/1 🙂 @ noboozebabes.com
And if I didn’t already have enough going on, I started reading The Coming Wave by Mustafa Suleyman, which, naturally, sent me down another rabbit hole.
The book got me thinking more about AI safety and, specifically, what actually keeps an AI system from doing something it isn’t supposed to do.
I figured the best way to understand it was to build a very small application.
I built a mini CRM running locally on my computer. I connected it to the Claude API and built a simple agent powered by Claude. The agent has a very limited set of tools it can use and specific permissions around what it can access. The interesting part wasn’t really building the CRM. It was learning how I could control what my agent can do.
For example, the agent can ask my application for an account, but Claude itself doesn’t get direct access to my database. My application checks whether the agent has permission to access that account first. If it doesn’t, the request gets denied.
That led me into learning about tools, permissions, authorization layers, least privilege, and why you shouldn’t depend on an AI model to follow a rule just because you told it to.
The simplest way I’ve found to think about it is: you don’t just tell the AI what it can’t do. You build the system so it actually can’t do it.
Lastly, I’m still working on my five-account outbound experiment.
I have been working the five accounts, but I’m no longer convinced that I picked the five most opportunistic accounts. I found some errors in the testing and ranking system I used to choose them, which is kind of the point of doing the experiment in the first place.
So there’s more work to do there.