Alex Comerford

Wanna be researcher writing about infrastructure, cryptography, and ML.

🔮

[Alex Comerford @thecmrfrd]

@mnaeraxr Many great learnings from this ^. IMO It felt very much like tensorflow v1.0 days .

[Alex Comerford @thecmrfrd]

When my grandma got her first smartphone at 91, I was blown away by how completely foreign touch gestures were to her.

“Pinch to zoom” and “double tap” might as well have been alien instructions. She had no reference frame for how these devices worked.

But after watching my brother and I use our phones, within a ~day she became just proficient enough to be dangerous.

It’s a good reminder that what's intuitive is local. What "feels" obvious to one person can be utterly opaque to another. Yet universally even a 30 second demonstration can accomplish multitudes more than just letting users "figure it out"

[Alex Comerford @thecmrfrd]

@tbpn @davidchang Non-alcoholic is the way

[Alex Comerford @thecmrfrd]

Everything that needs to be said has already been said, but no one was listening, so it all must be said again.

[Alex Comerford @thecmrfrd]

@karpathy @jenzhuscott This is one of the most sentimental gifts you can give your descendants. Many wonder what their ancestors thought, dreamt, or longed for. Growing a second brain can give them all the answers to that.

...or you can just keep a journal...

[Alex Comerford @thecmrfrd]

I used to feel a bit of shame pushing a purely vibe coded PR (and still do to some degree). But the abilities of the models to one shot maintenance fixes, misconfigurations, code owner updates, among other mundane tasks. Zero shame, just pure utility

[Alex Comerford @thecmrfrd]

@cbnsndwch @joinhello_miami @JaneStreetGroup @thelabmiami Best place to hack around in miami

[Alex Comerford @thecmrfrd]

Grateful to have finally solved one of the janestreet HF puzzles 🙇‍♂️

https://huggingface .co/spaces/jane-street/droppedaneuralnet

[Alex Comerford @thecmrfrd]

Even with the huge LLM abstraction layer, we are stuck with the forever fight against code slop. The larger the piece of code AI writes, the higher the chance slop slips in. Some AI tools try to address this, but most don't. The best known way to reduce and even negate slop are the traditional software techniques (good testing, code hygiene, etc.). The projects that will succeed will be the ones that find the right marriage of both worlds.

[Alex Comerford @thecmrfrd]

The full read is on github . com/rust-lang/rust/issues/76560

But the true unlock of this feature is being able to do type safe tensor shape transformations 👀

[Alex Comerford @thecmrfrd]

TIL if you turn on rust's `#![feature(generic_const_exprs)]` feature you can enforce compile time "inequality checking" with const generics and do stuff like a type safe suffix.

[Alex Comerford @thecmrfrd]

Also the reverse

[Alex Comerford @thecmrfrd]

There is something so intangibly satisfying about writing a bit-twiddle style function.

[Alex Comerford @thecmrfrd]

@cbnsndwch @thelabmiami Thanks for coming!

[hello_miami @joinhello_miami]

here's Hack Night in a nutshell, in case you were wondering. thank you @GeminiApp @NanoBanana for the help! 😊

[Alex Comerford @thecmrfrd]

⛓️https://t.co/nNrqlyKQnV

[Alex Comerford @thecmrfrd]

Big thanks to @AlexKontorovich for making his Intro to Real Analysis course open and lean based!

As someone who never had the chance to formally study it, it's amazing to now be able to take it and know my proofs are verified!

[Alex Comerford @thecmrfrd]

Still only up to lecture 5, but it's nice that the course revisits some classic proofs like Archimedes property :