Hacker News: leodavi

New comment by leodavi in "The Thinking Plant's Man (2025)"

leodavi — Wed, 06 May 2026 18:48:46 +0000

I want more of this. Trees using coordinated cell transpiration to control the weather is so fictional but so possible. It's the kind of thing that makes Earth so cool. What do you read? Anything digestible to a curious layman with ~20 minute reading sessions on the train rides?

New comment by leodavi in ""Parse, don't validate" through the years with C++"

leodavi — Thu, 30 Apr 2026 18:37:57 +0000

Well, in C++ the constructor must return a value of its class type - you can't return an Option from a constructor on T, for example, and since constructors are the canonical way to construct an object, it creates stylistic and idiomatic friction when you start using free functions to create a Maybe instead of constructors.

New comment by leodavi in "GPT-5.5"

leodavi — Fri, 24 Apr 2026 19:26:37 +0000

> in a highly controlled and predictable environment

Why this constraint? A common sentiment I see online (sorry, to group you in) is "[tool] will be capable, actually, but only in a context that trivializes its usefulness."

I think modern post-training like RLVR + inference-time output token scaling can _probably_ scale so the agents can solve any computable task, even when placed in noisy or misconfigured environments. But it won't be economical for a long while. But it already seems largely capable of that today.

New comment by leodavi in "DeepSeek v4"

leodavi — Fri, 24 Apr 2026 13:19:09 +0000

Probably not. The active parameter set may change from token to token, based on my understanding of MoE, so you'd be streaming (at the worst case, unlikely for a real scenario but frames the problem) 49B parameters from SSD for every output token...

New comment by leodavi in "Julia"

leodavi — Tue, 03 Feb 2026 02:33:48 +0000

The narrative style reminds me of the novel-game Caves of Qud. Very well done.

New comment by leodavi in "The largest number representable in 64 bits"

leodavi — Tue, 03 Feb 2026 01:58:06 +0000

To the author: this work you've published rests on some background knowledge that I'm familiar with but also a lot of background knowledge that's foreign to me. For the parts I was able to follow, I thought the mental gear-turning was enjoyable and interesting! But pursuing these intersections between algebra and computer science is often not so interesting, at least, when I research it myself. Is this something you've learned in university? I ask because I fear that without a formal learning environment, I may never pick up on the more advanced logic halfway through the article. Anyway, cheers to the good work.

New comment by leodavi in "Ask HN: What Are You Working On? (Nov 2025)"

leodavi — Mon, 10 Nov 2025 06:40:07 +0000

A visual editor for creating video games on the browser and on Linux desktop: https://stickyfingies.github.io/g2ngine

I've done this with C++ in the past, but ran into substantial friction with the CMake toolchain, specifically w.r.t:

- cross-platform compilation with large dependencies (vcpkg ports)

- running multiple compiler chains in the same build step

That second point is necessary if, for example, there's some AOT asset processing work that uses a native tool, and you're building for web. Expressing that some targets should use the emscripten toolchain while others should use the native toolchain, and interleaving between them, was a mess. TBF, I haven't done that with cargo or build.rs yet and it may prove to be equally frustrating.

Other features:

- undo/redo using a stack of swappable states

- serialization to disk (native) and LocalStorage (web) with some integration tests in progress but I am not satisfied with the correctness of my implementation: I want to *guarantee* that all information is preserved round-trip, but I also want a Patek watch.

- OBJ, GLTF, GLB models are loaded as "blueprint scenes" which are distinct from the "world scene." I made this distinction at the type-level because "scenes" are groups of entities that use newtype IDs (`LightId(u64)`, `MeshId(u64)` etc.) as primary and foreign keys to refer to each other, and I wanted to make it impossible for an entity in scene A to hold an ID for an entity in scene B. Instantiating a blueprint requires creating new IDs for every object.

- W.I.P. Alpha rendering, depth sorting, overhauling the material system to support multiple shaders (tough) that may be compiled after the engine itself (even tougher, a lot of runtime dynamic state and schema validation stuff), physics, scripting - oh yeah!

- Scripting using JS on both web (runs in browser itself) and desktop (uses a packaged JS runtime `Boa`) but Boa doesn't perform well on desktop in debug mode so I'm exploring other options.