Shubh Srivastava

student @bitspilaniindia

[Shubh Srivastava @Shubh_2595]

@latentbrain_uc @askalphaxiv completely free mate

[Shubh Srivastava @Shubh_2595]

@akanyaani @FrontiersMind Currently working on a cool project at my college BITS Pilani- post training a reasoning model that reasons in an indian language , with every COT token would be generated in hindi , bangla or any other indian language

[Shubh Srivastava @Shubh_2595]

@akanyaani @FrontiersMind i built an entire transformer from scratch - https://t.co/1evwLvwHEA , no external libraries except pytorch , try the model and a finetuned qwen model on the same data here - https://t.co/uhEOlX8t9o

[Shubh Srivastava @Shubh_2595]

@SarvamForDevs Sarvam Prism?

[Shubh Srivastava @Shubh_2595]

@RomitMirchi is cursor hiring interns in india? not there on the careers page

[Shubh Srivastava @Shubh_2595]

i've read it so many times , i still i cant actually get what harness engineering actually is , like hows it different from an agentic system

[Shubh Srivastava @Shubh_2595]

that's insane , starting to build - Oct 1, 14:53
@GnaniAi was one of the most fascinating startup I saw at the AI Summit with a really cool team behind it !
#GnaniAI #GreatIndianAIInternshipChallenge

[Gnani AI @GnaniAi]

We’re opening up @GnaniAi to student builders.
The Great Indian AI Internship Challenge is now live.

The idea is simple:

Build something useful with Gnani AI’s models and APIs, show us how it works, and put your work in front of the team here.

There are up to 100 paid internships available, with individual stipends of up to ₹1 lakh per month for selected interns.

You’ll also get:
1. 5,000 programme credits to build with.
2. Access to Prisma, Timbre and Evon.
3. 10 award categories across language, creativity, real-world impact, demos and more.
4. A chance to be shortlisted for an internship with Gnani AI.

You can build solo or with one teammate.

What you need to do:
Step 1: Register at https://t.co/fxIMcdsqtt.
Step 2: Pick a unique problem.
Step 3: Build using Gnani AI models or APIs.
Step 4: Record a working demo.
Step 5: Post it on LinkedIn or X.
Step 6: Submit before November 10, 2026.

That’s it.
No fixed idea list. No single “right” project. We want to see what you choose to build when the tools are in your hands.

Submissions close November 10, 2026.

Start Building!!

#GnaniAI #GreatIndianAIInternshipChallenge #AIInternship #StudentDevelopers #BuildInPublic #DevelopersIndia

[Shubh Srivastava @Shubh_2595]

@goodfellow_ian @Yoshua_Bengio @AaronCourville for reference :
overflow: Value exceeds the max capacity, causing it to "wrap around" to an incorrect number.
underflow: Value is too tiny to store, so it just gets rounded to zero.

[Shubh Srivastava @Shubh_2595]

@rasbt in his book recommended to use the native Pytorch softmax implementation rather than :

def softmax_native(x):
return torch.exp / torch.exp(x).sum(dim = 0)

the reason was numerical instability - such as overflow and underflow

[Shubh Srivastava @Shubh_2595]

@goodfellow_ian @Yoshua_Bengio @AaronCourville how this works is that when the observation c is very negative - exp(c) underflows , making denominator 0
and when x is very positive , exp(c) will overflow causing the whole expression to be undefined

[Shubh Srivastava @Shubh_2595]

@rasbt Hey @rasbt , loved build a LLM from scratch , even completed my implementation of it ...
One suggestion from side - Could you work on getting an India Specific version of it - it really helps bring the price down
Would be really useful and would spread even more