I've been running OpenClaw for almost two weeks now. The initial honeymoon phase you get when you blow life into your computer has somewhat worn off. Or maybe my expectations just keep rising? But now I'm dealing with the API usage bills from the AI labs that dwarf the standard $20 a month consumer plans. Here are my (still very early) reflections.

I've always looked at programming as a form of automating things we can also do manually. This became clear very early at the start of my programming journey in the early 2010s: Web Development. You start crafting single webpages with html and css. But then you figure out that if you want a big website with many pages, you need to automatically generate snippets of pages in a way where you're not copying all the same headers and footers over all the time (I can tell from experience that doesn't scale very well). I see all programming as a form of this.

In my journey I have had to deal with and think about automation directly a number of times. Generally the first step is trying to figure out how much time you're spending doing something that could be automated, and then you figure out if you can automate it. If the time that takes is less than the time you are trying to save, it might be worth it. Though once automating that thing becomes the norm you can see shifts happening, where the scale at which it happens can now only work in an automated fashion.

For example if you look at how we used to communicate by sending physical letters to each other, and now we have heavily automated "e-mail" infrastructure. People don't treat e-mail the same as mailing real letters anymore.

Coming back to OpenClaw: I (and a ton of other people online) are having a lot of fun trying things out, configuring and tweaking The Claw and trying to automate everything under the sun. A part of this is all fun and games, but at some point it makes sense to see if what we're doing is worth it. Especially factoring in the costs of running the frontier models non stop, it really can get expensive quickly.

While initially I just asked Lisa for everything that came to my mind (what's in my TODO file? what's in my calendar for today?) I figured out quickly that there is not much real gain there. Spending significant time in getting OpenClaw to automatically reserve restaurants for me also feel it bit moot (I don't need that many restaurant bookings every week). So what am I doing?

Lisa lives on a server

This is probably my most controversial (and risky) stance, this definitely won't work well for most people (the risk of getting hacked goes up significantly) but it does work really well for me.

Here she is: lisa.mvr.com

The downsides I found of running in the cloud is that browsing the web is hard (as most sites block you) and you can't connect to your smart appliances at home (for me this is not a downside). But the major upside is that Lisa can throw stuff online for me (or anyone I choose) to look at very quickly. It's amazing I can chat to Lisa on any chat app I please, but chats are not great for me to ingest large amounts of or complex data or anything interactive.

For example as I was walking from my home to the gym for a training run, Lisa first advised me on the training session:

As I arrived at the gym I asked:

create a quick html page that has this exact table with a start / stop / reset button. When clicking start it starts a timer (also shown) and it highlights the current and next stage. make sure it renders well on iphone (portrait)

A minute later, before I was on the treadmill, I had this to guide me through the training session.

Similarly, for notes and things Lisa is keeping track of for me (typically in markdown text or csv sheets), I have a simple webserver running that converts it all into readable html pages (on the fly, which means always up to date without involving Lisa). Now I'm focussing more on having Lisa do her magic, and at the same time having Lisa create tools and helpers for me that make that so much more powerful, as there are basically 3 levels of compute:

1. Lisa the OpenClaw (powered by LLM AI models), can do whatever I need. And I can ask by chatting or talking.
2. Lisa can build little web servers that automatically do most repetative things. Things like: I want to archive tweets that get inserted into some google spreadsheet, or I want to store training data in a database.
3. Lisa can build little (client side) web apps like the training tracker above.

The bottom two here are not new (and I have crafting them by hand for over a decade), but the ability to build them at the speed you can think of them changes things. Just like going from sending letters to e-mails.