The Rundown - October 25, 2025

“[I]t’s hard to design a convincing fake study. Fake things look fake. Reality is overdetermined.”

  • Chris Kelley - “Vibes won’t cut it”:

“ The building block task of programming, going back 40+ years, is writing a block of code that takes some input, does some processing as described by some requirements, and returns some output. Code generation AIs do this pretty well, even for some extremely hard programming riddles that coders (and AI companies) like to brag about solving. However, most of the advancements in programming languages, technique, and craft in the last 40 years have been designing safer and better ways of tying these blocks together to create larger and larger programs with more complexity and functionality. Humans use these advancements to arrange these blocks in logical abstraction layers, so we can fit an understanding of the layers’ interconnections in our heads as we work, diving into blocks temporarily as needed. This is where the AIs fall down. The amount of context required to hold the interconnections between these blocks quickly grows beyond the AIs’ effective short term memory (in practice, much smaller than its advertised context window size), and the AIs lack the ability to reason about the abstractions as we do. This leads to real-word code that’s illogically layered, hard to understand, debug, and maintain.”
“So the bubble is bad. Really bad. But even so, there will be things we can salvage from it: open source models, skilled programmers, cheap GPUs bought out of bankruptcy for pennies on the dollar. It would be better if we created that stuff without burning the world’s economy to the ground and emitting a heptillion tons of CO2, but ignoring the productive residue of the AI crash won’t bring the economy back, or suck the carbon out of the atmosphere.”

Trying to Sum Up Summer Before It Ends

I felt tired and defeated after the 2nd milestone presentation of my PhD thesis. A lot more friction (external and internal) than my last go. So I'm going to have to retool the approach. Less engineering. More law. Other than some new notes, not much going on with the draft (We are currently at 63 pages, and 1,382 notes).

Ruby and I travelled deeper into Kyushu - into Saga and Arita. A lot of obvious signs (abandoned fields and houses) that the Japanese countryside is being hollowed out by the declining birth rate. We felt sorry for a while, but then exploited the economic dowturn by buying their stuff at a discount.

Got myself a mini-PC (a Minisforum UM 790 Pro) and an old ThinkPad X1. Stripped off the Windows 11 and installed Linux (Arch on the Mini PC, NixOS on the laptop). Spent much of August just configuring a workflow for the writing and programming work ahead. Everything is now built around light, distraction-free terminal-based tools. I draft my thesis as LaTex files in Neovim, using the same editor to Python code.

But before all this I got to try out the latest version of Windows - something I haven't done since version 8 of the OS. It has become so much worse in terms of bloat,of pushing things like cloud services, content partnerships, and "Artificial Intelligence", with no easy way to set these aside. Meanwhile, the Linux experience has improved a lot. Installation and documentation is better (but then again my baseline Linux is experience is early 2000's Slackware distro - we had to compile everything ourselves). Arch has a well-earned reputation among programmers because it's so customizable and bleeding edge. On the other hand, NixOS gives you an interesting vision of computing - everything in your OS, like the apps installed and their settings can be encoded in a declarative programming language, and then "built" by a compiler/package manager. So theoretically, I can instantiate my NixOS environment in any computer through one script (without having to trust a cloud provider). Interesting times ahead.