Ishaan (@auto_grad

Ishaan

5,951 posts

Ishaan

@auto_grad_

RL @smallest_AI, prev @sarvamai

Joined November 2021

Pinned
Ishaan
@auto_grad_
Feb 25
Drop 3/4: NanoTok The fastest BPE tokenizer in written in C++ and wrapped in Python for api reference. Upto 32x faster encode, 24x faster decode vs tiktoken & HF Tokenizers (varies for different model configs). Peaks at 33M tokens/sec encode, 69M tokens/sec decode. Benchmarked
GitHub - 0xD4rky/nanotok
From github.com
18K
Ishaan
@auto_grad_
Nov 17, 2024
these are some computer vision papers that everyone must go through atleast once: 1. ResNets: arxiv.org/pdf/1512.03385… 2. YOLO: arxiv.org/abs/1506.02640 3. DeConv: lxu.me/mypapers/dcnn_… 4. GAN: arxiv.org/abs/1406.2661 5. Unet: arxiv.org/abs/1505.04597 6. Focal Loss:
72K
Ishaan
@auto_grad_
Oct 12, 2024
hey everyone, i am sharing my repo for ViT where you can find a proper path to learn vision transformers and its uses in video processings. i have also implemented ViT from scratch so if you wanna know how it works, i got you covered. more coming soon!
GitHub - 0xD4rky/Vision-Transformers: This repo has all the basic things you'll need in-order to...
From github.com
73K
Ishaan
@auto_grad_
Feb 26, 2025
a little career update: I am joining @SarvamAI as an ML intern. hoping to learn and contribute my level best!
38K
Ishaan
@auto_grad_
Aug 21, 2024
huggingface being an absolute saviour when it comes to blogs about deep concepts
20K
Ishaan
@auto_grad_
May 26, 2025
an interesting paper towards a new type of tokenization after a long time: arxiv.org/pdf/2503.13423 > super-bpe operates in two steps: (a) initally, the model learns subword units as in traditional BPE (b) then, the model identifies and learns frequent multi-word expressions,
38K
Ishaan
@auto_grad_
Jul 7, 2025
little career updates from me and @r4plh : i have joined @atomicworkhq as a ML engineer and @r4plh has joined @invideoOfficial as a CV engineer. now why is this special? > we started ml together > we failed together > we questioned this field together > opted out of campus
29K
Ishaan
@auto_grad_
Jan 16, 2025
this paper (arxiv.org/pdf/2406.16232) is what i was trying to find. the paper introduced a method to optimise the vector valued functions i.e. to optimize multiple objectives simultaneously (which often are interdependent). the jacobian matrix gives us info about how each
Ishaan
@auto_grad_
Dec 3, 2024
such a great answer on why loss functions aren't a vector valued function
18K
Ishaan
@auto_grad_
Dec 4, 2024
such a great paper on quantization of deep neural networks: arxiv.org/pdf/1911.09464
13K
Ishaan
@auto_grad_
Jun 27, 2025
my journey: > no family pressure, was hella enjoying life till 10th > heard about iits, joined resonance to prep for the same > cracked kvpy, prmo, jee advanced, got into iit > iit was not a dream place, it was basically a place for smart minds to follow norms and settle for way
Raj Dabre
@prajdabre
Jun 27, 2025
My journey: Bachelors: Debt (parents paid) Masters: Broke even (government scholarship) PhD: Broke even (government scholarship) 7 years of employment in government organization: Standard pay (slight savings) Google: What the fuck!
23K
Ishaan
@auto_grad_
Apr 20, 2025
i have written a detailed technical blog on the Q-LoRA mechanism covering technical and conceptual understandings involved in the paper and tried to simplify them. do read if you wanna dive deep into this mechanism :)
Q-LoRA Is All You Need
From medium.com
16K
Ishaan
@auto_grad_
Jan 9, 2025
i had the second round interview for a MLE role at a great us-startup today. interviewer was a phd scholar from an ivy league university. i was awestruck with the amount of indepth knowledge he had about the core deep learning concepts and the advancements in ai, yet so humble!
23K
Ishaan
@auto_grad_
Jan 29, 2025
oxford seems cool
35K
Ishaan
@auto_grad_
Jan 7, 2025
ResNets simplified: (i will only include 5 solid points, so do read) 1. problem: deeper neural networks were performing worse than shallower ones, despite having more capacity to learn (mainly due to early layers not being able to learn properly, read this:
9.9K