As the quantizing of local models becomes increasingly efficient, the reality of local AI coding agents feels reasonable. They're poor coders, but I watched in real time as something called Bonsai, running in Pi, built an HTML page for me of pixelated rain falling. It made the revisions I asked accurately — as accurately as a flagship model might, which is to say you get out what you actually put in, not what you think you put in.
I allowed one agent on my Mac to SSH into another one of my machines and install a much more powerful local agent. It downloaded a 48 GB model, ran tests, asked it questions, and printed the answers back onto my other machine. Meanwhile nothing on the desktop of the remote machine indicated anything was happening. It made me realize how little I knew about computers and networking before AI.
Then Claude got lost in the network and couldn't find the computer. It said things like, "wait there's multiple computers on this network, I can't find the old one.... that must be your phone...." — and it was at that point I realized I truly do not understand what I am doing.
Yet.
Shane Curry, "I Truly Do Not Understand What I Am Doing. Yet.," https://shanecurry.com/blog/i-truly-do-not-understand-what-im-doing-yet/, published April 19, 2026, updated April 29, 2026.