Patrick Marangone

If you accomplished your goal perfectly, how would people be affected?

[Patrick Marangone @pmarangone_]

it's 2023 and 🦜

[Sasha Rush @srush_nlp]

Arrival is the best NLP movie.

[Jay Alto @theJayAlto]

books are the most underpriced things on earth. for less than $20, you can shift your entire perception of reality, change the trajectory of your life, and become someone completely new. all in just a few hundred pages. insane.

[Patrick Marangone @pmarangone_]

cute

[Francisco Ribeiro @fraveris]

Dmitry Oleyn - First Star, 2025.

[Patrick Marangone @pmarangone_]

@mitchellh try @realGeorgeHotz or @karpathy next

always fun to ask "refactor it like x"

way better than prompt engineering

[Patrick Marangone @pmarangone_]

feels good man

btw I should probably open a PR do add dark mode to restate UI

[khipu.ai @Khipu_AI]

SAVE THE DATE

[gabriel @gabriel1]

first profitable ai company reinvests everything into trying to save millions of lives and to make it as generally available as possible

[⿻ Andrew Trask @iamtrask]

https://t.co/FSJTYf4iwq

[Patrick Marangone @pmarangone_]

@julien_c @badlogicgames you just saved me a lot of research time :)

[Patrick Marangone @pmarangone_]

@X_ArtGallery well done

[Patrick Marangone @pmarangone_]

@badlogicgames @trq212 i mean, inst general durability the backbone of the current agentic systems? multiple workflows running in parallel with good evals?

[Natural Philosophy @Naturalphilosy]

“It is not enough to be busy. So are the ants. The question is: What are we busy about?”

— Henry David Thoreau

[Naval @naval]

What you really want is a challenge at the edge of your capability.

[Patrick Marangone @pmarangone_]

@emilheap @karpathy lmaoo

[Richard Sutton @RichardSSutton]

The bitter lesson in 26 words:

Don’t be distracted by human knowledge, as AI has been historically.
Instead focus on methods for creating knowledge that scale with computation, like search and learning.

[Patrick Marangone @pmarangone_]

lmao

[Patrick Marangone @pmarangone_]

rust btw

[zack @zack_overflow]

Unfortunately, I don't use Zig now. Every 1.5-5x human DX productivity boost from Zig features is eclipsed by the 100x boost from coding agents writing Rust:

Allocator interface:
This is my favorite Zig feature, you feel so galaxy brain using a specialized allocator to optimize a code path (e.g. arena, stack fallback etc). The problem in Rust used to be that there was no Allocator interface equivalent and if you wanted a Vec<T> that used a custom allocator you literally had to copy+paste the std version and modify it to use it (this is what Bumpalo did, look at the source). For a long while now there has been an Allocator trait in nightly, and it seems to be good now. Because it is a trait it is static dispatch, vs Zig's which is based on a vtable. Unlike Zig there isn't a community-wide convention of designing data structures to be parametric based on the allocator, but AI changes the game and makes it trivially to copy paste code and change that. I find it works well enough for my use-case.

Arbitrary bit width integers + packed structs:
Another beloved Zig feature of mine. It makes it so easy to do DOD-style CPU cache optimizations and stuff like tagged pointers, NaN boxing, etc. and even made bitflags really easy to make. You could always do this in Rust or any systems programming language but it was really ugly/unergonomic. The least worst option was using some crate like bitfield/bitflags which both rely on proc macro magic to work. Now, with coding agents I literally do not care how annoying it is to write the code by hand.

Comptime:
This is Zig's flashiest feature, no other programming language except maybe for obscure dependent-types langs have compile time evaluation as nice as Zig's. I thought I would miss it a lot, but I actually don't. For me, 95% of comptime usage is to create Zig's version of generic data structures with parametric types. Rust has a better designed type system IMO (see next section). In the remaining 5% of cases, not having comptime sucks. The only reliable way to reach an equivalent is through codegen. I'm making a game right now, and I have hardcoded hitbox geometry data generated from a tool that I want to bake into a data structure. Without comptime, I have to get Claude to write a script that generates the Rust file. However, I don't find myself needing compile time evaluation that much anyway.

Rust's type system:
I think I'd rather trade having comptime for Rust's better-designed type system, especially for bounded polymorphism (traits/typeclasses). Trying to do the equivalent in Zig is a nightmare. Also, I think that Rust's type system allows you to enforce more variants and prevent coding agents from making common mistakes. In my game I use the euclid crate which essentially allows you to not mix up coordinate spaces (very common problem in graphics programming) by creating specialized types for each coordinate space (e.g. Point<Screen> or Point<World>)

Not having to deal with memory issues:
With coding agents allowing 100x more code to be written, this also means you need to scrutinize 100x more Zig code for memory issues. Without formal verification, the surface area of the search space to enumerate to find bugs is just so much larger now. With the magnitude of code being generated now, Rust is even more attractive. Rust's tradeoff was always that it hinders developer productivity especially if you are unfamiliar with borrow checker, but this simply does not matter with coding agents anymore. And if you do use unsafe in Rust there's tools like miri which you can have the coding agent run the code against to make sure it doesn't cause UB or isn't violating Rust's aliasing rules when it comes to unsafe.

I still miss writing Zig and find it to be a great language but I like Rust more and coding agents work with better with it.

[Patrick Marangone @pmarangone_]

@JustJake lmao

[Nihal Pasham @npashi]

Finally able to talk about what I've been heads-down on for 6 months at @nvidia 🦀⚡

We just open-sourced cuda-oxide — an experimental rustc backend that lets you write CUDA kernels in pure Rust.

No DSLs. No FFI. No source-to-source step. Single source.

Short🧵👇

[Massimo @Rainmaker1973]

Scientists have created one of the most detailed 3D reconstructions of a human cell (eukaryotic cell) ever produced.

This groundbreaking model, often termed a "Cellular Landscape Cross-Section Through a Eukaryotic Cell," combines data from X-ray tomography, nuclear magnetic resonance (NMR), and cryo-electron microscopy to map molecular structures in extreme detail.