Utkarsh Shukla (UT)

co-founder @ringg_ai▪️ Predicting Next tokens with a lot of bias. Learning constantly ! Ex DS @letsblinkit, Ex ds @AtlanHQ. Ferrari Tifosi.

[Utkarsh Shukla (UT) @the__UT]

I don’t remember where i read the tweet “if its worth reading read it slowly” but its so true !!!

[Utkarsh Shukla (UT) @the__UT]

Never thought when i would be typing into twitter’s search bar “SA” it would be for salman and not sam alt

[Utkarsh Shukla (UT) @the__UT]

@TRAI yes

[Utkarsh Shukla (UT) @the__UT]

Subtle brag, subtle brag

[Sam Altman @sama]

we had a significant security incident during evaluation of our models. we are sharing what we have learned so far. thanks to @huggingface for the partnership on this.

https://t.co/2o2VfR6PIa

[terminally onλine εngineer @tekbog]

*yelling at claude*
KIMI BUILT THIS IN A CAVE

[Utkarsh Shukla (UT) @the__UT]

@liquidai The models are exceptionally good but it really lacks on language coverage.

[Utkarsh Shukla (UT) @the__UT]

@reach_vb why does the new chatGPT app not have my chats on the left side. It was way more convenient to continue past chats. Also chat organised by datetime was really good.

[Utkarsh Shukla (UT) @the__UT]

GPT sol is just toooo fast. That tokens/sec is crazy crazy. @OpenAI 🫶

[Utkarsh Shukla (UT) @the__UT]

@kali_charan_v lol

[Utkarsh Shukla (UT) @the__UT]

is it me or American accented commentary for football is just unnatural.

[Utkarsh Shukla (UT) @the__UT]

@waterloo_intern What were the conversation domains, also would be able to share the data ?

[Utkarsh Shukla (UT) @the__UT]

Fred Again is cool… but you’ll find me from @ringg_ai at the #Gemma @huggingface Meetup on 15th July.

Looking forward to meeting fellow AI builders, researchers, and open-source enthusiasts.

See you there! 🤗 💻

[Utkarsh Shukla (UT) @the__UT]

PS5 sales will sky rocket because of GTA, ModernWarfare

[Utkarsh Shukla (UT) @the__UT]

Cocaine is great but have you ever locked in on writing gpu kernals to make inferencing faster on huge amount of coffee ?????

[Utkarsh Shukla (UT) @the__UT]

At @ringg_ai I was recently comparing the output of two conversational LLMs and was trying to find semantic similarity between the output of the simulations across 64
use-cases, 1614 simulations and a total of 18939 turns.

From my previous experiences I have had great results from OpenAI's text-embedding-3-large for english + hinglish usecases. To that end I wanted to test out how the open-source embedding models are performing, and came across the gem https://t.co/aejB5OyLHp

The similarity between the results is astonishing. With all of the training data that the big orgs have the much smaller Sbert models was doing better in cross lingual similarity.

Have added some plots showing the Distribution and Cumulative Distribution of the two models. Its astonishing how similar they are both in values and patterns

[Utkarsh Shukla (UT) @the__UT]

@BMGFIndia Data collection at its peak !!

[Utkarsh Shukla (UT) @the__UT]

@deepakmodidev @ringg_ai Great !!!!!!!

[Utkarsh Shukla (UT) @the__UT]

The Nuremberg (2025) is such a beautiful piece of cinema, love it !

[Utkarsh Shukla (UT) @the__UT]

Loss of the model goes up and down, up and down, up and down. Loss of the model goes up and down all through the @wandb town.

[Utkarsh Shukla (UT) @the__UT]

@liquidai it would be great if in the next iteration you folks add hindi in the pre-training phase atleast in the extraction models.
The smaller models are exceptionally good and fast but just lag a lot when it comes to hindi.

[Utkarsh Shukla (UT) @the__UT]

It amazes me that i can train half a billion model on my macbook great work @UnslothAI also it is using apple silicon .