28 ms·

There's a model of computation called 'interaction nets' / 'interaction calculus', which reduces in a more physically-meaningful, local, topologically-smooth wa

There's a model of computation called 'interaction nets' / 'interaction calculus', which reduces in a more physically-meaningful, local, topologically-smooth way. I.e. you can see from these animations that LC reductions have some "jumping" parts. And that does reflect LC nature, as a reduction 'updates' many places at once. IN basically fixes this problem. And this locality can enable parallelism. And there's an easy way to translate LC to IN, as far as I understand. I'm a noob, but I feel like INs are severely under-rated. I dunno if there's any good interaction net animations. I know only one person who's doing some serious R&D with interaction nets - that's Victor Taelin.

hmokiguess 11mo ago
Speaking of Victor Taelin, what's the latest on https://higherorderco.com/ https://higherorderco.com/ ? His work is really inspiring and amazing
- killerstorm 11mo ago
  He shares the progress on Twitter quite often. In the last year they shifted the focus away from raw performance (as beating existing stuff is rather daunting) and into rather unique stuff with code synthesis, perhaps relevant to formal verification of vibe-coded code, etc.
tromp 11mo ago
> there's an easy way to translate LC to IN While easy, it sadly doesn't preserve semantics. Specifically, when you duplicate a term that ends up duplicating itself, results will diverge. There exist more involved semantics preserving translations, using so-called croissants and brackets, or with the recent rephrased approach of [1]. [1] https://arxiv.org/abs/2505.20314 https://arxiv.org/abs/2505.20314
- marvinborner 11mo ago
  > While easy, it sadly doesn't preserve semantics. There is actually an easy way that does preserve semantics at least to WHNF - it's called closed reduction. Mackie has worked on it a bunch (see some resources [1]). An even simpler implementation is Sinot's token passing. The problem with both of these approaches is the decreased amount of sharing and potential for parallelism, which is typically the reason for using interaction nets in the first place. [1] https://github.com/marvinborner/interaction-net-resources?tab=readme-ov-file#non-optimality https://github.com/marvinborner/interaction-net-resources?ta...

hmokiguess 11mo ago

Speaking of Victor Taelin, what's the latest on https://higherorderco.com/ https://higherorderco.com/ ? His work is really inspiring and amazing

killerstorm 11mo ago
He shares the progress on Twitter quite often. In the last year they shifted the focus away from raw performance (as beating existing stuff is rather daunting) and into rather unique stuff with code synthesis, perhaps relevant to formal verification of vibe-coded code, etc.

tromp 11mo ago

> there's an easy way to translate LC to IN While easy, it sadly doesn't preserve semantics. Specifically, when you duplicate a term that ends up duplicating itself, results will diverge. There exist more involved semantics preserving translations, using so-called croissants and brackets, or with the recent rephrased approach of [1]. [1] https://arxiv.org/abs/2505.20314 https://arxiv.org/abs/2505.20314

marvinborner 11mo ago
> While easy, it sadly doesn't preserve semantics. There is actually an easy way that does preserve semantics at least to WHNF - it's called closed reduction. Mackie has worked on it a bunch (see some resources [1]). An even simpler implementation is Sinot's token passing. The problem with both of these approaches is the decreased amount of sharing and potential for parallelism, which is typically the reason for using interaction nets in the first place. [1] https://github.com/marvinborner/interaction-net-resources?tab=readme-ov-file#non-optimality https://github.com/marvinborner/interaction-net-resources?ta...