Increasing Entropy #4

Discoveries of the week, random links, and a few thoughts on random topics

100x more programmers

About a month ago, I showed Amp to my wife Anna, and since then she has rebuilt her business website and migrated data from an old CMS.

I remember the mix of suspicion and joy she expressed when I told her, “You’re a programmer now. You can just do things”.

I think that’s generally where the world is heading. Everyone will do more and more things with computers: create new programs and websites, fix bugs in software they use, run one-off throwaway scripts for work and fun.

How I use AI (Oct 2026) by Ben Stolovitz.

How Revolut Cut AI Coding Costs 90% With Self-Hosted Open Models — David Abitbol from Revolut in a 10-minute video on how they self-hosted GLM 5.3 for development. Interesting that they seem to be going against the industry trend. It seems they spend a lot of time optimizing inference and TTFT while everyone else is moving toward developing in parallel in cloud sandboxes. I’m curious how it’ll end up.

Mac at any scale — How and why Namespace built a datacenter with MacBooks: macOS virtualization, networking, macOS disk images, and more.

Walker playground 🚶 — Drag one to carry it, flick it to throw it, resize the window to make it bounce off the edge.

Spreading Liquid with three.js.

The Golden Age — We’re in the Golden Age. But at what cost?

“You can’t design software you don’t work on” — Sean Goedecke.

This resonated with me. I’ve been doing a lot of codebase reviews for prospective customers at datarockets. Very often I’ve seen questionable code or decisions, but I’ve always been hesitant to mark them as definitely bad. A lot of decisions make sense in context and context could be huge and versatile.

People continue to lobotomize LLMs with the help of other LLMs.

Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights — Rikiya Takehi, Jayden Teoh, Alex Zhang, and Omar Khattab used Astra to understand Qwen3.5-0.8B weights and tune some of them to achieve different goals, e.g. increase a winning rate from 0% to 100% in a game. See demos.

Hugging Face reproduced the experiment by changing the weights from the paper and rerunning the evaluation.

No training or fine-tuning is required, just observing neuron activations and direct changes to model weights.

“Qwen 3.8 27B on my laptop can interpret model weights, too” — Simon Willison.

Matthew Honnibal calls the original post misleading. He notes that the results were achieved not just by observing and understanding weights, but also by using scripts and analysis. He says that’s what ML engineers do from time to time anyway.

Sean Goedecke thinks that “a smart human ML researcher could perform all of these tasks”.

To me, it’s still impressive that it’s possible to improve model performance by changing weights directly. Even though we humans can also do this (analyze and change weights directly), we can do it at a much bigger scale with the help of other LLMs. Maybe Astra doesn’t understand weights clearly, but it can analyze and modify the relevant weights directly.