@maxliani It also has diminishing returns; as code grows, even the AIs become bad at understanding it. I learned this the hard way in 2 projects. I tried letting AI do the work, but theyve grown to the point they cant be maintained & the only way to move forward is to rewrite it all myself
@maxliani This is basically how I’ve come to feel about AI usage as well. It’s powerful, but for any sufficiently complex project agentic coding feels unsustainable. If we don’t at the very least design & write the architecture of a program ourselves, we will never understand how it works.
@molecularmusing Every API I write requires the user to pass an allocator argument to any potentially allocating function. This includes container types as well, which either need an allocator pointer stored inline or to have the allocator passed to individual functions
@dcolascione@valigo Not disagreeing. But I do have an idea about using an ECS-like architecture in part of my compiler for recording facts during sema that can later be exposed to LSP. Kind of like a knowledge base in a Prolog program, so that tools can query it
.@thsottiaux my dot is ghosting me on both desktop and mobile. “Couldn’t connect the call. Please try again.” Surely this constitutes a reset-worthy offense.
@dcolascione@steveklabnik My language isn’t an array lang (and I don’t have experience with them). That is neat, but it feels overall less flexible. Structural matching works for any types with similar layout.
I like that I don’t have to reach for a concept or other generic, so it compiles fast
@steveklabnik@dcolascione Interesting. I’ve been working on a little programming language recently, and I permit opt-in width subtyping on fn params with the syntax:
`fn foo(v : ~vec2)`
I generally agree that hierarchical refinement types are a mess, but this kind of subtyping I like
@steveklabnik@dcolascione I’m curious if you feel this way about width subtyping as well.
E.g. where you might have something like `fn foo(v : vec2)`, but you could pass a `vec3` or a `vec4` as arg just as well, since their first 2 members are laid out the same.
@thsottiaux When the systemd OOM killer kills VS Code (‘code’) it also kills all instances of codex open across terminals. It’s a mystery to me, but that probably needs fixing
@miketuritzin This is why GPU programming is just the worst. I’m always splitting apart larger shaders because a single branch can blow up the registers and make the whole thing slow. Put the same code in 2 shaders and now it’s fast… It feels truly silly compared to writing code for the CPU.
@AmarthiaAuthor@SandyofCthulhu I grew up on a cliff overlooking the Chesapeake Bay. Lots of the land was graded sometime over the past century, but there are still plenty of cliffs on the Bay. In the past there would have been even more.