Aryan

A open source contributor trying random (sometimes new?) things

[Aryan @aryunsaini]

@Arpitaaa01 I averaged to 30k steps on weekends when I was in germany this summer 😭

[Aryan @aryunsaini]

What if LLMs also dream and learn while they sleep?

[Aryan @aryunsaini]

@jiyajain_23 Open-source has become a race like JEE, I feel bad that nowadays we don't look at open source how our seniors used to - something to learn from and contribute to.

[Aryan @aryunsaini]

@kuppiesss Koi na, roadside bhik maangna is always an option... We all can make websites and apps to automate that 💪😀

[Aryan @aryunsaini]

@vaivaswat2244 Escaping the matrix....

[Aryan @aryunsaini]

@cmchelsimehta The physics of this site is crazy, you can break the ball by pushing against the edges hard. And if the edge is not ground, the pieces fall down.

This is really nice!!!!

[Aryan @aryunsaini]

Why does @claudeai always breaking the rules we have set for it?
Always saying sorry afterwards.
AGI is here - not sure, but Rule-breaker is here already - that too which breaks it without you knowing unless you ask it specifically

[Aryan @aryunsaini]

This is inspired from the fantastic paper on Tversky Neural Networks by @mdoumbouya, @jurafsky and @chrmanning

Repo with the notebooks: https://t.co/aNNX1dgWJw

[Aryan @aryunsaini]

What happens if you put Tversky similarity inside an RNN cell?

I added a Tversky branch to mLSTM in Vision-LSTM and xLSTM. At equal params, half the depth matched or beat full depth, with ~20% fewer FLOPs and up to 2x faster training.

Read the Article: https://t.co/CrbuIY5VG9

[Aryan @aryunsaini]

@Avishka82 @grok what do you think? Should we just remove asking DSA in OA and embrace agentic coding in OAs as well like the new Amazon OAn format?

[Aryan @aryunsaini]

It's early and the API will change, so if you try it and something feels off I want to hear it. Full writeup and the notebook: https://t.co/fjyYk0aDAQ
pip install pyqit
https://t.co/YLkKNWPSnF
@PennyLaneAI

[Aryan @aryunsaini]

Every QML experiment I ran started with the same 60 lines of PennyLane boilerplate. Split the data, scale it onto angles, shape the weights, write the loop. I got sick of it and turned it into a package. PyQit: pick a model, give it your data, call https://t.co/DuBDmCqGVr().

[Aryan @aryunsaini]

I trained a 4-qubit classifier twice, once in PyQit and once by hand in plain PennyLane. Both runs agree on every epoch's loss and accuracy and on the final weights, to the last digit. The difference is 10 lines instead of 60.