I want to write down a few thoughts reflecting on my own AI use.

Let’s start with what I don’t use AI for: I don’t want to outsource my thinking.

AI can help me find information to inform my decisions, but even when doing that, I want to stay as close to the raw source(s) as possible. I try to avoid overly short summaries and potentially oversimplified explanations. When dealing with a topic I might not be super knowledgeable on, I prefer a reasonably detailed overview that leaves me with a bunch of questions I can then work on getting answers for. That way, I believe I can avoid those pesky “unknown unknowns” that an overly simplistic approach might leave.

I also don’t outsource the communication of my thoughts, both because it would be rude and because putting my thoughts into words, spoken and written, helps me to understand them better.

The overall trend in AI usage tends to go the opposite direction. I still see many prompts along the lines of “Always keep answers short. Never ask follow-up questions.”, even when people are using AI to help them understand things. The faster you go, the harder it becomes to establish your position relative to where you’re trying to go. Friction, coming from longer reads, more complex concepts you have to deal with, follow-up questions, etc., makes you slow down just enough so you can look around and get your bearings.

I drafted the above paragraph with a borrowed physics metaphor with Heisenberg’s uncertainty principle and asked my agent whether it actually held. It told me it was fine, but I still decided to read up on it myself. I came to the conclusion that it didn’t hold, so I decided to remove it.

So how do I use AI and what do I use it for?

Hermes Agent has become my go-to way of interacting with LLMs. I run one instance on my Mac Mini that I use as a home server, then use it either via Hermes Desktop on my PC and Macbook or via Telegram from my phone.

My Hermes setup is pretty simple. No external memory systems, no crazy integrations or skill collections. I use Camofox to make it easier for my agent to research things and Plannotator to review and give feedback on code, but that’s it.

I use AI to gain speed by going slow with it. That might sound counterintuitive, but it’s an approach that has served me well so far and I feel like it has saved me from a lot of AI-induced frustrations that I hear about from other people. Oftentimes, those frustrations come from wanting the AI to do too much at once. I always try to be as specific as possible, both in selecting and communicating the task I want the AI to perform for me. When I notice that I can’t be as specific as I would like to, that informs me of a knowledge gap that needs to be filled and I will take the time to fix that first.

Being able to pace yourself might be the single most important skill when it comes to using AI. It’s incredibly easy to go too fast with it and lose track of where you are and where you want to go. In more concrete terms: Slowing down gives you more time to think about whether your approach to a problem is correct, whether it’s the best solution you can come up with and even whether you’re even trying to solve the right problem. Otherwise, it’s easy to race into the wrong direction at breakneck speed.