Vaibhav Bhatia

Moving through life using epsilon-greedy policy | ML and Caffiene aficionado | @iitroorkee

[Vaibhav Bhatia @Point_Dexxter]

Tortoise Better

[Vaibhav Bhatia @Point_Dexxter]

Dayum, this is the core work of many startups

[Claude @claudeai]

Introducing Cowork: Claude Code for the rest of your work.

Cowork lets you complete non-technical tasks much like how developers use Claude Code.

[Vaibhav Bhatia @Point_Dexxter]

Wanted to build something. Got inspired after an interview.

> First RAG Project
> First full-fledged working frontend+backend+retrieval

Fun way to practice for interviews. Just working for now as v1. Will improve v2, one-upping ChatGPT with persistent memory and personalization

[Vaibhav Bhatia @Point_Dexxter]

Fuck blr traffic, how can 1.9 km take 50 mins

[Vaibhav Bhatia @Point_Dexxter]

Premium Apartment in Bangalore.
Bangalore apartment hunt arc going crazy.
#Bangalore

[Vaibhav Bhatia @Point_Dexxter]

@cneuralnetwork Nahhh, there are multiple videos online of students of bits goa.

[Vaibhav Bhatia @Point_Dexxter]

This is a cool site to visualise tokenizers used in gpt models

[Vaibhav Bhatia @Point_Dexxter]

🫠 CP algorithms being used in NLP. DP is nostalgia, although 1 mahine pehle hi kri thi.
Btw this algorithm is a question of Leetcode.

[Vaibhav Bhatia @Point_Dexxter]

Ambani bhau upload speed is not enough 🫠

[Vaibhav Bhatia @Point_Dexxter]

Got Born
Got Married
Achieved Death

2024 was such a big year for me
Ending the year with a selfie

[neural nets. @cneuralnetwork]

got married
got 10 cgpa
landed 1.5cr pa job on campus
and found this cool top on my table

2024 was so good

[Vaibhav Bhatia @Point_Dexxter]

@pr0b0t1sc00l Nope, online

[Vaibhav Bhatia @Point_Dexxter]

First time with a good gpu in my own hands :) Lessgooo

[Vaibhav Bhatia @Point_Dexxter]

@cneuralnetwork @NoobCoderVibes Below the tree

[Vaibhav Bhatia @Point_Dexxter]

Some people are overachieving in 2024, like chill out how are you completing 2-3 different phases of life in one year

[Vaibhav Bhatia @Point_Dexxter]

Update:
DL Comms Project - Results of encoder-decoder were getting skewed towards 0 as the training dataset was mostly skewed to 0, so doing standardization/normalization helped.

[Vaibhav Bhatia @Point_Dexxter]

@siddharthb_ @TurboML Interested for ML Engineer role

[Vaibhav Bhatia @Point_Dexxter]

People talking about moving on from token based - LLM, let me reach there first bro.🫠

[Tanishq Mathew Abraham, Ph.D. @iScienceLuvr]

Large Concept Models: Language Modeling in a Sentence Representation Space

This new paper from Meta introduces a very interesting, novel approach to language modeling.

Rather than doing prediction of the next subword token, focus on generating the next concept, which is represented by a sentence.

Effectively, the model operates over sentence embeddings. Here, the pretrained SONAR sentence embeddings are used. Text is encoded by SONAR, passed into the model, the next SONAR sentence embeddings are generated, and are decoded back to text.

Various variants for the model itself are explored. This includes a standard decoder-only autoregressive Transformer, different diffusion model archs, and autoregressive quantized archs. The diffusion model archs seem to perform best and were scaled up to 7B and were comparable to Llama-3.1-8B and other ~7B LLMs for summarization and summary expansion tasks.

[Vaibhav Bhatia @Point_Dexxter]

Update:
1. Been working on an encoder-decoder model for the above project which was not training at all.
Problem found : Wrong activation function.
2.The results got better with changing the batch size.
Still improvement is needed.

[Vaibhav Bhatia @Point_Dexxter]

@yash87845 @cneuralnetwork @Cobratate Taklu?

[Vaibhav Bhatia @Point_Dexxter]

Recently I saw a tweet about reverse training the model. Today I am proud to say I have also achieved it.