Logan Weber

PhD student in CS

[Logan Weber @DoobsParn]

@thsottiaux Unread status of threads not preserved on app update

[Logan Weber @DoobsParn]

This has been a really fun project to be involved with, and it's the beginning of a big change in how we write programs. It's the first work I'm aware of that moves beyond the paradigm of isolating LLMs from program state and communicating purely through serialization.

[Ellie Cheng @ellieyhc]

🦜✨Nightjar ✨makes your prompts first-class code in your Python programs: Your prompts read and write program variables, compute with program objects, and implement program control flow. 🧵(1/n)

[Tian Jin @jintian]

⚡️Come check out how we scale LLM decoding parallelism! Excited to present learned asynchronous decoding with @ellieyhc for DLCT @ml_collective tomorrow at 10am PST! Thanks to @jasonyo @savvyRL for organizing.

[Logan Weber @DoobsParn]

If you're at ICML, come check out our poster from 11:30am to 1:30pm today! We're poster #1206 in Hall C.

[Logan Weber @DoobsParn]

Just landed in Vienna for ICML and can’t wait to share what @JesseMMichel, @alex_renda_ , @mcarbin, and I have been working on! The main idea is to use hypernetworks to compile programs to neural networks. (1/N)

[Logan Weber @DoobsParn]

You can find the paper here! https://t.co/NOUVTOfKJQ

[Logan Weber @DoobsParn]

Just landed in Vienna for ICML and can’t wait to share what @JesseMMichel, @alex_renda_ , @mcarbin, and I have been working on! The main idea is to use hypernetworks to compile programs to neural networks. (1/N)

[Logan Weber @DoobsParn]

Let me know if you’re at ICML and would be interested in talking about deep learning for code! @JesseMMichel and I will be presenting our poster on Tuesday at 11:30am-1pm. (N/N)

[Logan Weber @DoobsParn]

Just landed in Vienna for ICML and can’t wait to share what @JesseMMichel, @alex_renda_ , @mcarbin, and I have been working on! The main idea is to use hypernetworks to compile programs to neural networks. (1/N)

[Logan Weber @DoobsParn]

We instantiate a neural surrogate compiler as a hypernetwork, CompNet, that is 1.9-9.5x as data-efficient, trains 4.3-7.3x faster, and produces visual results that are 1.0-1.3x more similar to ground truth. (5/N)

[Logan Weber @DoobsParn]

That is all. I hope I've ruined your day🙂

[Logan Weber @DoobsParn]

If you compile this function with clang, it will round the input to the nearest integer:

double nearbyint(double x){ }

[Logan Weber @DoobsParn]

And if you change the name to, say,

double nearbyant(double x){ }

it will no longer do so.

[Logan Weber @DoobsParn]

Imagining the timeline where we call it "Schönfinkeling".

[Logan Weber @DoobsParn]

If you have a natural transformation between two interpretation functors, does that signal anything of interest? For the example I'm thinking of, it seems to show the source functor is "finer-grained" (separates more programs), but not sure in general.

[Logan Weber @DoobsParn]

@annie_littler Now u listen here u lil shit

[Logan Weber @DoobsParn]

This proof 👆 straight-up 👆 ruined my day

[Logan Weber @DoobsParn]

For anyone who has experience with theorem provers and is interested in learning Lean 4, I'd recommend the slides and associated exercises here: https://t.co/B2Q6EUgfWR