Mohammed Safi Ur Rahman Khan

PhD @ai4bharat, @iitmadras, @WSAI_IITM

[Sumanth @sumanthd17]

Happy translations to all those who celebrate! Hopefully this is the last one we ever need!!

Btw keep your eyes peeled for a key feature that’s gonna drop soon, along with the model weights :)

[Mitesh Khapra @MiteshKhapra]

No Indian idea should have to wait for translation.

Introducing Indic-Translate: built to carry meaning across India's languages.

Built by @BodhanAI & @ai4bharat.

22 languages, both directions · 32K context, whole documents at once

Live on https://t.co/qX5WB8Rkuu · 05.09.2026

[Simran Khanuja @simi_97k]

Which cultures do LLMs best align with on open-ended questions? Why? Can we change that default?

We study this in our new work: "Steering LLMs for Culturally Localized Generation" (COLM 2026), from my time at Google DeepMind! 🧵

https://t.co/yrOjhVfVUi

[Wadhwani School of Data Science & AI (WSAI), IITM @WSAI_IITM]

From being the recipient of the 2025 @Google Ph.D Fellowship to his second year at @WSAI_IITM, Mohammed Safi talks about his experience.

Watch the video to find out why @WSAI_IITM is a great place to grow as a Ph.D researcher.

[Wadhwani School of Data Science & AI (WSAI), IITM @WSAI_IITM]

Ever wished that fake news was easy to identify and avoid? Dr. Suman Kundu’s paper EA2N which takes a completely different approach to identifying fake news.

Watch the video to understand how this new approach works.

[Himansu Didwania @hdidwania1997]

We’re hiring a Product Manager for Speech at Sarvam.

Looking for someone with ML/AI experience to work with research and engineering on ASR & TTS- shaping roadmaps and turning research into real products.

DMs open.

[Andrej Karpathy @karpathy]

Three days ago I left autoresearch tuning nanochat for ~2 days on depth=12 model. It found ~20 changes that improved the validation loss. I tested these changes yesterday and all of them were additive and transferred to larger (depth=24) models. Stacking up all of these changes, today I measured that the leaderboard's "Time to GPT-2" drops from 2.02 hours to 1.80 hours (~11% improvement), this will be the new leaderboard entry. So yes, these are real improvements and they make an actual difference. I am mildly surprised that my very first naive attempt already worked this well on top of what I thought was already a fairly manually well-tuned project.

This is a first for me because I am very used to doing the iterative optimization of neural network training manually. You come up with ideas, you implement them, you check if they work (better validation loss), you come up with new ideas based on that, you read some papers for inspiration, etc etc. This is the bread and butter of what I do daily for 2 decades. Seeing the agent do this entire workflow end-to-end and all by itself as it worked through approx. 700 changes autonomously is wild. It really looked at the sequence of results of experiments and used that to plan the next ones. It's not novel, ground-breaking "research" (yet), but all the adjustments are "real", I didn't find them manually previously, and they stack up and actually improved nanochat. Among the bigger things e.g.:

- It noticed an oversight that my parameterless QKnorm didn't have a scaler multiplier attached, so my attention was too diffuse. The agent found multipliers to sharpen it, pointing to future work.
- It found that the Value Embeddings really like regularization and I wasn't applying any (oops).
- It found that my banded attention was too conservative (i forgot to tune it).
- It found that AdamW betas were all messed up.
- It tuned the weight decay schedule.
- It tuned the network initialization.

This is on top of all the tuning I've already done over a good amount of time. The exact commit is here, from this "round 1" of autoresearch. I am going to kick off "round 2", and in parallel I am looking at how multiple agents can collaborate to unlock parallelism.
https://t.co/WAz8aIztKT

All LLM frontier labs will do this. It's the final boss battle. It's a lot more complex at scale of course - you don't just have a single train. py file to tune. But doing it is "just engineering" and it's going to work. You spin up a swarm of agents, you have them collaborate to tune smaller models, you promote the most promising ideas to increasingly larger scales, and humans (optionally) contribute on the edges.

And more generally, *any* metric you care about that is reasonably efficient to evaluate (or that has more efficient proxy metrics such as training a smaller network) can be autoresearched by an agent swarm. It's worth thinking about whether your problem falls into this bucket too.

[AI4Bharat @ai4bharat]

Sarvam just released Sarvam-30B and Sarvam-105B, two large language models built in India.
Benchmarks look promising. But the real question is: How do they actually feel when you use them?

Both models are now live on the Indic-LLM Arena. You can now blindly compare them against other frontier models.

Try them out today and cast your votes to decide who leads
https://t.co/tov64KH9ZQ

@MiteshKhapra @anoopk @pratykumar
@vivek_raghavan @WSAI_IITM @OfficialINDIAai
@EkStep_Org @PeoplePlusAI @SarvamAI

[Pratyush Kumar @pratykumar]

📢 Open-sourcing the Sarvam 30B and 105B models! Trained from scratch with all data, model research and inference optimisation done in-house, these models punch above their weight in most global benchmarks plus excel in Indian languages.

Get the weights at Hugging Face and AIKosh. Thanks to the good folks at SGLang for day 0 support, vLLM support coming soon. Links, benchmark scores, examples, and more in our blog - https://t.co/DcCG3zlN8p

[Aritra 🤗 @ariG23498]

With the help of @huggingface we (/w @RisingSayak) are building a ML Club India 🇮🇳

What we want to do:
1. Online talks
2. IST compatible timing
2. Open to all

More to come in this week! Watch this space. 🤗
Special thanks to @LysandreJik who motivated me to keep working on this. 🔥

[Pratyush Kumar @pratykumar]

Drop 8/14: Introducing Saaras V3, the next iteration of our speech recognition model. We extend our lead in this space with an even more accurate model, particularly for mixed-language and noisy speech.
We have also expanded support for all the 22 scheduled languages of India.
And we now support real-time streaming! In this mode, the model delivers low latency while preserving transcription quality.
https://t.co/SQSgCPsUr5

[Pratyush Kumar @pratykumar]

Drop 7/14: Introducing Sarvam Arya - our multi-agent orchestration platform, built from the ground up with robust systems engineering principles and frontier AI-assisted developer experience.

We illustrate the Arya advantage on a common ETL task of extracting structured data from unstructured documents. Arya system with GPT 4.1 mini achieved ~5x higher accuracy at ~10x lower cost compared to Claude Code with agent swarm.

Read more about Arya in this thread and our blog: https://t.co/UVodZcgMLd

[Pratyush Kumar @pratykumar]

Drop 5/14: Introducing Bulbul V3, our latest text-to-speech model. It raises the bar for how human it sounds, while being super robust.

In an independent third-party human listening study, Bulbul V3 delivers the highest listener preference, and low error rates across use-cases and languages. See details in our blog, but first watch the video. https://t.co/HkijWrupi9

[AI4Bharat @ai4bharat]

For AI to be truly inclusive, it must understand more than just grammar—it must understand context.

@AI4Bharat at @iitmadras had launched the Indic LLM Arena. This isn't just another leaderboard; it’s a public utility for:
✅ Developers: Test your models against real-world Indian use cases.
✅ Enterprises: Find out which LLM actually resonates with your customers in rural India.
✅ Sovereignty: Building AI that respects our social fabric and safety norms.

Be a part of this movement. Try the Arena today and help us rank the models that will power India's digital future.
👉 https://t.co/xgaq3RkLjF

#GenerativeAI #DigitalIndia #IITMadras #IndicLLM #indiaaiimpactsummit2026

@MiteshKhapra @anoopk @prajdabre @ravi_iitm @partha_p_t @ManishGuptaMG1 @meghtweets @dineshteewari1 @abapna
@WSAI_IITM @OfficialINDIAai @EkStep_Org @PeoplePlusAI

[Wadhwani School of Data Science & AI (WSAI), IITM @WSAI_IITM]

🚀 We Are Hiring!
We, at @WSAI_IITM, are assembling a talented engineering team to build advanced conversational AI systems that can drive meaningful societal impact. The goal is to create AI agents that help people at scale, across many languages, regions, and public service contexts.

[AI4Bharat @ai4bharat]

Calling Product & UX/UI Designers!

AI4Bharat is looking for talented volunteers to help redesign our website and reimagine the digital front door for India's open-source language technology.

What you’ll work on:
• Redesigning the overall website experience and user journeys
• A cohesive visual design system
• Clear, compelling ways to present cutting-edge AI research

Why join?
A meaningful pro-bono opportunity to create real impact — and add a standout project to your portfolio.

Apply here: https://t.co/HK5skPLwwI

Questions? Contact [email protected]
#DesignForImpact #UXUI #ProductDesign #AI4Bharat #IndianLanguages #OpenSourceAI #SovereignAI #DesignForGood

[Sarvam @SarvamAI]

The real breakthrough in AI is not just about what it can do, but how we choose to use it.

Our co-founder @pratykumar will be at the @TheProductfolks (Un)Conference '25 in Bengaluru, joining leaders reimagining the future of products and technology.

Pratyush will share thoughts on how India's growing AI ecosystem is shaping innovation, collaboration, and accessibility at scale and what it really takes to build responsibly in this new era.

It's going to be a day filled with meaningful conversations, diverse perspectives, and ideas that push boundaries.

📅 Date: 15th November
📍 Location: Bengaluru
🔗 RSVP: https://t.co/exTncg5RFR

[Partha Talukdar @partha_p_t]

So glad to see this getting launched, congratulations @MiteshKhapra and @ai4bharat team! 🎉

This arena fills an important gap and I encourage everyone to try it out.

Glad @googlecloud and @GoogleDeepMind could provide initial support.

[Mohammed Safi Ur Rahman Khan @SafiKhan2k]

@importdhruv @ai4bharat Its an arena... Meant to compare two models.

[Mohammed Safi Ur Rahman Khan @SafiKhan2k]

@aadityansha_06 @ai4bharat We will release a sanitized version of this data in the open-source. We never sell data. Open source has always been our motto...

We are a non-profit academic research lab - not in this for making money.