Saksham Bassi

science @amazon music. opinions are my own.

Former NLP research @CILVRatNYU, @inmobi

[Saksham Bassi @SakshamBassi]

Activation patching confirms it: ablating the manifold directions tanks probe accuracy far more than random-direction controls, showing these subspaces really do carry the task-relevant info.

📄 https://t.co/EwC8CfSHC2

[Saksham Bassi @SakshamBassi]

Tasks where the ordinal variable is locally computable from token identity → clean 1D manifolds with place-cell-like tiling.

Tasks needing cross-position integration or semantic extraction → higher-dimensional, less coherent representations.

[Saksham Bassi @SakshamBassi]

Architecture matters too: Qwen3-4B shows much stronger geometric "twisting" than Gemma for indentation and unlike its numeric twisters, these twists preserve ordinal order.

[Saksham Bassi @SakshamBassi]

Do LLMs represent all ordinal concepts the same way?

We tested 4 tasks (bracket depth, indentation, table position, numeric magnitude) across Gemma-2-2B/9B & Qwen3-4B and found the answer is no, the geometry depends on how the model has to compute the value.

[Saksham Bassi @SakshamBassi]

@kaushik_akshat @IndiaPostOffice yes received bro

[Saksham Bassi @SakshamBassi]

https://t.co/QvmP5kWDMm is my work bff now @AnthropicAI

[Saksham Bassi @SakshamBassi]

@marikgoldstein @madiator 🤣

[Saksham Bassi @SakshamBassi]

@kaushik_akshat @IndiaPostOffice since 2 weeks. I wish tracking was across countries.

[Saksham Bassi @SakshamBassi]

This is the craziest thing I have heard this week, crazier than MosiacML news.

This reminds me of the famous dialogue from The Social Network. A Billion Dollars.

[Mustafa Suleyman @mustafasuleyman]

Excited to announce that we’ve raised $1.3B to build one of the largest clusters in the world and turbocharge the creation of Pi, your personal AI.

https://t.co/p5AfRXGPan

[Mustafa Suleyman @mustafasuleyman]

Today I’m excited to announce the first version of our new personal AI, Pi... https://t.co/wYpgcXdB1t

Pi is smart, kind and supportive. It’s designed to be better at natural, flowing conversation than lists, plans, or code.

[Saksham Bassi @SakshamBassi]

🫠 @huggingface

[Saksham Bassi @SakshamBassi]

adding the gradient of loss term to your loss function might help your model improve (:)

[Saksham Bassi @SakshamBassi]

just read code written by someone who I really looked up to and from a leading research lab.

have really mixed feelings after reading through functions of >400 lines.

[Saksham Bassi @SakshamBassi]

in-person project discussions with professors are so cool, reminds me of (pre-covid) undergrad research 🥹

[Saksham Bassi @SakshamBassi]

One of the recent works(https://t.co/anrqpifCks) shows decision boundaries being modified:

[Saksham Bassi @SakshamBassi]

Minimizing (KL) divergence between 2 output predictions of the same model on 2 very similar inputs helps improve model generalization.

[Saksham Bassi @SakshamBassi]

The longer a neural network takes to reach an optimal state, the better it generalizes

from: https://t.co/9I0NXMZ9Ht

[Saksham Bassi @SakshamBassi]

nyu library has Herman Miller chairs. underrated af.

[Saksham Bassi @SakshamBassi]

greenhouse is the best job application portal. it's so much better than its competitors.

[Saksham Bassi @SakshamBassi]

attaching from the SAM paper (https://t.co/8odwr8St2f)

[Saksham Bassi @SakshamBassi]

reading about improving model generalization has got me to question how we used to/currently train neural networks