2026, September
Links
- Lucid Planet - Lucid Planet II (Music Album)
- Haxe programming language
- A cop’s case for Flock
- Coordination, Good and Bad
- If math is more than proof, we need to better celebrate the rest of it
- Animation vs Math (Video)
- How to keep enjoying programming in a world of LLMs
Commentary
Lucid Planet - Lucid Planet II (Music Album)
Artist: Lucid Planet, Album: Lucid Planet II, released: 2020-11-14
Phenomenal “progressive rock” album I’ve listened to repeatedly this month.
Haxe programming language
I spent quite a few hours looking for a good programming language for some personal projects, and found a particularly interesting candidate for my needs in Haxe. What I was looking for at the time, in descending order of importance:
- “simple” language, letting me focus on logic over language details
- familiar syntax
- non-complex memory mechanics
- familiar programming paradigms
- static typing with useful abstractions
- type inference
- enums with associated values
- pattern matching
- generics
- graphics/rendering work as an established use case
- in this case via frameworks such as Heaps, OpenFL, HaxeFlixel
- also via raylib bindings
- straightforward project distribution
- multiple compilation targets for cross-platform distribution
- good tooling for potential collaboration with third parties
- suitable performance
- hxcpp backend provides suitable performance for substantial graphics machinery
- possibility to tune garbage collection if profiling reveals a bottleneck
- depth if desired
- powerful AST macros
- decent maturity and organizational health
- as far as I could assess from the outside
- beyond the experimental stage, with development supported by more than one person
Other languages closely considered were OCaml, Nim, Odin, Kotlin, V, Crystal, Rust, C++ (as a familiar baseline). As of now I can’t guarantee I’ll write any planned projects of mine in Haxe, but for what it delivers it seems to fly under the radar of many.
A cop’s case for Flock
Ned Donovan - Works in Progress, issue 25 - 2026-08-28
Donovan’s case for license-plate readers touches on a question I’ve been thinking about lately in another context: how can societies use data to improve decision-making without concentrating capabilities for surveillance and control?
I arrived at this question via the economic issue of internalising negative externalities through pricing. Negative externalities include the costs imposed on third parties, e.g. when you drive somewhere on public roads (noise, pollution, congestion, additional accident risk). Externality charging, though, is improved by more accurate, context-specific data aiding in estimating those costs and setting those charges accordingly.
I’d prefer systems with technically enforced limits on data collection, retention, and access, supported by formal verification where feasible, over systems that depend entirely on laws, institutional trust, and oversight. Viewed in this light, the Flock debate raises a broader systems-design question: which capabilities are necessary for the intended benefit, and which opportunities for misuse can be excluded by construction?
Coordination, Good and Bad
Vitalik Buterin - Personal Blog - 2020-09-11
A well-written essay on enabling beneficial coordination while limiting harmful collusion. The distinction between decentralization as fault tolerance and decentralization as resistance to collusion has influenced how I think about AI decentralization, governmental organization, and corporate governance.
If math is more than proof, we need to better celebrate the rest of it
Grant Sanderson - Guest article on Terence Tao’s Blog - 2026-09-18
Grant Sanderson’s guest article asks what mathematicians should value as AI systems become increasingly capable of generating proofs. He argues that proof production is a proxy for a broader goal: advancing human understanding. His proposal is to give “motivated explanations” (accounts that clarify where ideas come from and why they matter) academic recognition comparable to that given to new proofs.
The article interests me particularly because I see a parallel in software engineering. Even if an LLM can produce an implementation faster and more accurately than I can, I still need to understand its assumptions, evaluate its design, and judge whether its trade-offs suit the requirements. That includes performance, maintainability, and integration with the surrounding system. Explaining why a particular abstraction or algorithm fits the problem is the closest counterpart to mathematical exposition; deciding whether it fits is also an engineering responsibility.
A working implementation does not by itself provide a mental model of its behavior, limitations, or possible extensions. As long as engineers remain responsible for evaluating and maintaining software, that understanding remains essential. The question is therefore broader than who can produce the code: how do we make its behavior and design rationale understandable enough to assess and build upon?
Animation vs Math
Better script than Matrix 2 & 3.
How to keep enjoying programming in a world of LLMs
Manuel Bärenz - Post on the Haskell Discourse Forum - 2026-09-18
Manuel Bärenz argues that if we utilize LLMs correctly in everyday software engineering, it’s possible to increase productivity, joy and learning simultaneously. He and I probably both have considerable freedom to decide how we utilize LLMs at work, and not everyone has that freedom.
Facing similar tradeoffs myself, and using LLMs extensively to study and deepen my understanding, I largely agree with his assessment.