And it’s already been a year! How time flies when you’re having mild fun and very busy.

About one year ago, I decided I should revive the blog, clean it up, and start posting stuff again. This is just a minor retrospective of it. So, how has it been?

It was a very good decision! The most obvious reason is that it acts as a minor forcing function for me to put the things I learn about together, at least a little bit. It also forces me to refine my ideas a little, in the way that presenting to other people usually does. Sometimes, writing something up with the expectation that someone else might read it forces you to say “wait, wtf am I talking about. This makes no sense”. Basically, a rubber ducky method but you get pretty blog posts out of it.

(this is what GPT thinks a New Yorker style cartoon looks like.)

Something I def still have to work on is not letting the posts go on painfully long. It’s difficult and I mostly fail, but I’ve picked up a few tricks to improve. One is to remind myself that I can always come back and edit them, so I don’t need to rephrase things a million times like they’re gonna get engraved in bronze. It’s better to just throw it out there warts and all, than have it sit in some random folder.

Another trick has been to just liberally split things into new posts, if I feel like there’s anything there at all. These aren’t academic papers, it’s some guy posting about stuff he’s learning on his personal blog. So there’s really no reason to not just make a post on some tiny random thing, which helps other posts stay shorter.

That said, I still want to just pump them out with less effort, basically using them as my notes.

The AI meatgrinder continues to grind

The coding agents have gotten real damn good, especially in the last year. This is actually a godsend for the blog; I think one of the better things I can offer is making cool and informative figures about a given idea, which usually means making very custom plots with matplotlib. The coding LLMs still have a few pitfalls, but they are absolutely stellar at that type of thing. So it’s waaaay easier to go from “I’d like to see this image given this data” to just having the image.

Speaking of, in the “welcome back” post before, I mentioned that even if the LLM can explain the concept pretty flawlessly, one competitive edge we might still have on the LLMs is making figures. In the one last year where I asked it to make a diagram of how Dynamic Movement Primitives (as a random example) work, it definitely kind of fudged it. It’s been a year since then, so how does it do?

I didn’t check it closely, but from what I remember, that looks mostly right. At least better than last year! We’re in trouble, aren’t we 😬

on the other hand… on closer inspection, I guess it still wants to tell us about the “leerned bogas”:

So, we can check back in a year when it’s absolutely perfect.

What else…

Definitely glad I ditched the stupid wordpress site. God, that thing was a quagmire. A github pages site + jekyll + some custom mods has some annoying bits, but it’s far better.

I really like the 3D presentation of the site I made, that was fun. I’ll def be experimenting with other versions of that in the future.

Anyway, that’s all. See you again in a year!