@LH DDR4 non-ECC. ECC is too expensive here in the EU.
I switched to Btrfs, which helps with data corruption. Claude was also able to identify the corrupted sectors and isolate them at the Linux kernel level.
@mjackson I'm building something to prove you wrong, though I'm not sure I'll pull it off 😄
The idea: pair a custom language with an advanced compiler that "compiles the framework away" and emits highly optimized vanilla JS.
https://t.co/lvqfjh4aRI
I’m not sure. Most games use off-the-shelf engines anyway, so there’s not much in their code worth protecting. Even custom engines rarely do anything Claude couldn’t reproduce from scratch.
What makes a game valuable is its assets, design, and art direction, and streaming doesn’t protect any of that.
@LH@roc_lang > Rust does contort the code in unnatural ways, requiring data copies to avoid borrow checker issues
Maybe you could instruct Claude to be more lenient with using unsafe? It doesn't really matter, specially if you are going to rewrite in Zed later on.
Nice write up! Yes, the self-hosted backend is for local development only, production builds still need to use LLVM.
The memory safety is mostly an non-issue in practice. Compilers can (and should) pre allocate most of the memory it needs ahead of time. Borrow checkers pay for their price when you need to do a lot ad-hoc allocations. I've never had a single issue related to memory in Beni.
But if you are planning to self-host Zed then it might be smarter to stick with Rust indeed, and then push toward that goal.
Check the Zig skills that I wrote in case you try that.
https://t.co/oU34Yt0j2c
It's important to make the agents write the architecture correctly from the start. Tell them to use data-oriented design, bump allocators, etc. You can also clone Zig itself locally and use it as reference implementation.
@LH You should consider Zig. Its self-hosted backend builds very fast . My compiler builds in ~2s vs 2min in LLVM mode. It's very good for agentic workloads.
@roc_lang migrated from Rust to Zig, Richard Feldman talks about it in some podcasts
@LukasHozda I understand the decision not to accept AI contributions, but I wish the core team used AI more themselves. I got tired of waiting for AArch64 support, forked it, and asked Claude to do it. Lo and behold, a few hours later:
https://t.co/zruFEwX28S
I don’t know honestly. I find it hard to believe.
I tried using Opus 5.5 to create assets in Blender and it’s not very good. It lacks spatial awareness. It’ll draw a house upside down, look at it and say “LGTM”
The agents left behind several debug pages, all accessible through URL query parameters.
They named the rifle PAC9. There’s also an obscure Polish rifle called the PCC PAC9. Maybe it’s just a coincidence. I hope these LLMs aren’t being trained on Polish military schematics lol
@W0lfByteStudios The JS is minified, so there isn’t much left in there, but they didn’t strip the comments from the CSS.
I found this in there, which indicates orchestration.
@MXRetroDev Eu tive um problema parecido fazendo engenharia reversa do Betrayal at Krondor. A IA foi incapaz de escrever código para DOS. A maioria dos bugs estava relacionada a pointers far/near, uma limitação do OS de 16 bits
https://t.co/7rXbHUigKN
@jarredsumner At this point, wouldn’t it be better to just rewrite the backend in Rust instead of trying to shoehorn JS into everything? These models can easily one-shot this kind of conversion