So it’s week 6 of the sabbatical. I’ve settled into an acceptable rhythm of working on personal projects. A couple weeks ago I was trying to work on three or four different projects at once. Despite leaning heavily into using coding agents, I find that I could only keep that up for a day or two. The constant context switching is exhausting.

Yet I also couldn’t just do one thing at a time. The latency is often too much. Not just because agents take time to respond and do the work I asked for, but because some of the jobs require meaningful compute and time.

And so I’ve settled into having just two active projects on the burners at the same time. As for right now, the main one continues to be the Sparse Autoencoder feature interpretability project. Hopefully more to come on that shortly. The other is this Perezzo paper craft collaboration with RJ that I’ve been tinkering with on and off for a while now.

A while back, we tried to build one of these with just slices, which … didn’t quite work the way I had hoped. It was flimsy and didn’t cut well. Then recently I came across this book Making Geometry on my shelf again, and thought “well, we could cut this out.”



Which, well, maybe it will work, maybe not. But it is a nice project to put on the other burner. Both projects requires waiting on a machine. One is running matmuls, and the other is cutting paper.

Feels appropriate and on brand somehow.