Owen Lynch

Grad student at Oxford, Research Software Engineer at the Topos Institute.

Currently working on the programming language side of systems theory.

Trump is a fascist, fascists are bad.

Website: https://owenlynch.org

Last Notes

2026-06-07 11:14:57 GMT

I really feel like generational garbage collection is underappreciated, and not enough part of people's mental models when they compare GC languages to manual/Rust-style memory languages.
In a GC language, "heap objects" are *much cheaper* than "heap objects" in a language with manual memory management, because they are first allocated on what is essentially a bump allocator, and then if they survive long enough, are allocated as part of a *bulk* allocation.
Casey Muratori talks about "n+1 programming" which is when you shift from thinking about individual objects in your program to collections of objects. A well-tuned runtime system for a GC language is *precisely* this; it handles memory management in batch-mode. https://www.youtube.com/watch?v=xt1KNDmOYqA
In order to recover the performance of a GC language doing tree manipulation in a manually managed language, you have to completely rearchitect your code to think in batches, use custom allocators, etc. Which is possible, it just means that the idea of "writing functional-style code in Rust" is a lie. You have to work at a fundamentally lower level of abstraction.
A runtime system is *not* just some cope that managed languages need in order to run, it's a tool that you actually may have to reimplement when you don't have it.

2026-05-22 14:42:49 GMT

I'm trying to stick to my "ask mastodon rather than chatgpt" habit: does anyone have recommendations for interesting places to visit in Paris?

2025-09-08 17:46:04 GMT

- reply

My objection to this kind of article is that it is very, very rare for a category to actually well-capture the idea of "relationships" (whatever that means) in the way that people intuitively mean the word.
That is, when people say "relationship" they often mean some subtle combination of causal influences. But morphisms in a category usually:
1. Are theoretical entities: the presence or absence of a morphism tells you how systems can be compared, not how systems are actually related.
2. Are necessarily simply compositional: there is a well-defined way to compose any two morphisms.
And this makes a category kind of an awkward entity for modeling "relationships."
And this plays into a larger problem, which is that the application of mathematics presupposes a great deal of work in the informal sphere to distill real-world stuff into formalities. Category theory doesn't solve this problem, and a blind faith in category theory can indeed exacerbate this problem by ignoring it.
I think that considering homomorphisms is quite important, because it reveals what is inherent and what is accidental in a presentation. But it's important in the same way that being dimensionally correct is important: it's a good sanity check but it's not at all sufficient for something that is well-suited as a world model.

2025-03-02 15:17:13 GMT

- reply

Weirdly enough, the fastest way to do this is to go through Bakersfield (there's the San Joaquin train from Bakersfield to Emeryville). It's not as scenic as the coast starlight, but it's about 4 hours faster, including the bus from LA to Bakersfield. But if you are starting from Riverside rather than LA, maybe the easiest way to do this would just be to drive to Bakersfield and park there (3 hours), and then take the train from Bakersfield to Emeryville (6 hours).

2025-02-22 22:25:01 GMT

- reply

Ohad Kammar, Robert Atkey, Justus Matthiesen, possibly others

2025-02-22 09:26:52 GMT

- reply

Excellent! And I hope to get up to Edinburgh at some point after May 1st, to visit you and also perhaps some of the people on the SGAI project who live in Edinburgh.

2025-02-21 23:35:22 GMT

- reply

I have no permanent residence other than Oxford, and not too many travel plans, so I should be there most of the time! And I am planning on going to act2025.

2025-02-21 19:00:57 GMT

- reply

It does have a physical office, and we have 6 people in it! You should come visit us!

2025-02-21 09:54:34 GMT

- reply

By the way, back in 2023 Davidad and I wrote up a blog post about goals for an expansive dynamical systems theory, which I think is still interesting and maybe gets at some of the really tough math that we hope to do in this program https://topos.institute/blog/2023-06-15-compositional-world-modeling/.