lukemerrick

I am not super familiar with C and CUDA, so I read solely for the README and enjoyed it supremely. The blend of cheerful walking through instructive examples and your philosophical takes on how to approach the exercise to get the most out of it put me in a great mood. You captured that special upbeat attitude that comes about when you're doing something as well as you can just because it's so legitimately interesting to you.

Meta is a very large organization, and I'm willing to believe that a good chunk of Meta FAIR (the lab releasing all of this stuff) truly do care about innovations for advancing AI safety and are doing great work along these lines. I'm not disagreeing with your point about the company being led by its financial incentives as a unit, but let's also allow ourselves permission to celebrate this work by this group of people.

Looks like there is both an ARM and x86 version according to the docs. Probably need two different binaries, but you still get cross-OS for each architecture.

Just spent two hours on this. Would happily spend the time again (and probably will soon!). Awesome work.

Related to the idea of "no one trains on data they own, they shouldn't own the resulting model": since big public datasets like The Pile have CC-SA items in them, is anyone considering bringing the argument that model weights are derivative work that must be "shared alike"?

As you guessed, the history tracking is one of the killer features. Imagine it being super easy to edit the history of a REPL session (delete, reorder, merge, and edit contents of each command) and rerun... That's a notebook! Notebooks also allow for markdown input and rich HTML output (which is killer for plotting) making it possible to polish your REPL history into a document you'd actually want to share with a colleague to explain something like a data analysis workflow.

I actually started in notebooks and then learned to love the REPL as a simplified "scratchpad notebook." I'd say in many ways notebooks are an improvement that cater heavily to REPL-lovers, but that for some quick tasks, the extra complexity isn't always worth it.

Thank you for the tip! Getting to implement a different language from Lox feels like a nice way to cut down on the tedium.

Wow, the whole concept of a peephole optimizer is a bit mind blowing to me. I'm appreciating all the reasons to power through to writing a bytecode VM as the next step.

I'm not sure how far down the compiler I actually will enjoy going vs. exploring ideas around type systems, linters, etc. up near the AST level, but if I do venture down this advice will certainly come in handy!

Thank you for the suggestions! I was actually just searching about MLIR today after reading some Julia language community discussions on the new Mojo language that uses MLIR.

I actually played with Unityper.jl and SumTypes.jl, but my conclusion was that if I was going to depart from dispatch on Julia types in my code, I might as well just stick to an untyped tree, since either way I'd have to have a single `evaluate` function for interpreting any kind of node.

Reconsidering now, it seems that there might be benefits beyond type dispatch to having a typed syntax tree, so maybe I'll give that a shot as a next step!

In general, if you keep the source code positions of every nontrivial token, and you keep the raw source code, then yeah you can print out those pretty specific point error messages regardless of whether you keep your trees lossless. Also, if you want to include filename in your messages (perhaps because unlike Lox your language supports imports), then you'll need more than just lossless trees to store the necessary information.

I'm not sure exactly how Rust and rust-analyzer keep track of the info necessary to their excellent error messages and diagnostics, but I wouldn't be surprised if pinpoint messages were not the primary motivation for rust-analyzer to do lossless parsing.