Rishab

Data Scientist @reliancejio | ex @dailyhuntapp, health. lifting. AI. travel

[Rishab @rishabbjainn]

dump?¿

[Shubham Bansal @BakarBansal]

In the next 10 minutes, a delivery boy from Blinkit will ride in scorching heat to deliver me a refreshing drink.

All because it's hot.
The temperature is 48 outside. My AC is on 16 inside.
I still feel discomfort.
So, I buy comfort at someone else's discomfort.

That's the game of life.

In the game of life, morality has no space. Ethics don't exist.

But remember -

The same delivery boy fills up petrol at a station, where a handicapped person is sitting to check the air in his bike's tyre, somehow managing 2 meals a day.

And next to that handicapped person, there is a dog, somehow managing 1 meal in 2 days!

That's the game of life.

[Gaurav Aggarwal @fooobar]

Dream of shaping the future of AI? Join Reliance Jio’s mission to lead with innovative AI products with focus on Indic AI! We’re seeking top talent globally—researchers, engineers, students, academics - to build groundbreaking AI.

https://t.co/0PdgtvxnRu

[Rishab @rishabbjainn]

@selfawareatom Appreciate the data curation efforts, is it possible to open source the training data set?

[michielh.eth @michieldoteth]

GPT-2 latent paths vary by text type:
• Poetry: longest (94.5), smoothest, most chaotic (entropy 1.77)
• Tech: shortest (30.6), sharpest turns (curv 2.15), high reuse
• Math/code: angular, high-curv (2.14/2.12), low similarity
• Story: arcs, moderate everything

[Gaurav Aggarwal @fooobar]

We must rethink the fundamental inputs that drive technology innovation. At the core of it lies talent—not just competent talent, but the kind of bold, restless, generative talent that thrives on creating what doesn’t yet exist.

If this is who you are in deep AI work, buzz me.

[Sudarshan Kamath @kamath_sutra]

Exactly. Instead of training another LLM we should invest in fundamental research that can solve problems that today’s LLMs can’t.

[Rishab @rishabbjainn]

Mem0 just hit SOTA on long-term conversational memory

[Taranjeet @taranjeetio]

Today, I’m excited to share our latest research on “Building production ready AI Agents with Scalable Long-Term Memory”.

We’ve achieved state-of-the-art (SOTA) performance—26% more accurate than OpenAI Memory.

We evaluated Mem0 on the LOCOMO benchmark and found that it consistently outperformed all the six baselines on all types of questions from multi-hop to temporal. Mem0 reduces latency by 91% and cuts token usage by over 90% compared to full-context methods, offering fast and cost-effective performance.

Today’s AI agents quickly forget important information once it moves beyond their context window, leading to broken conversations, repeated mistakes, and lost user trust. Larger context windows only delay the issue - making systems slower, more expensive, and harder to scale.

Mem0 was built to solve this head on - giving AI Agents a scalable memory layer that remembers what matters, reasons faster and adapts over time.

Check out the full paper below 👇🏻

[Gaurav Aggarwal @fooobar]

Leaving everything else aside, it's perplexing to see govt using public money to fund the most well funded AI startup in India? VCs who have put 41M dollars would have taken care of it - it's their money at stake, no?

What were they thinking? 🤔

I guess I am getting old!

[Rishab @rishabbjainn]

@upperwal @deepseek_ai Unable to dm!

[Rishab @rishabbjainn]

@monali_dambre @AgnoAgi discord link?

[Rishabh Agarwal @agarwl_]

Gemma 3 made me appreciate how far small models can be pushed by scaling post training, particularly distillation.

[Rishab @rishabbjainn]

@sgdescent https://t.co/wIWABNbCSN

[Rishab @rishabbjainn]

@dwijaajaana where do you get those smoothie bowls?

[Gaurav Aggarwal @fooobar]

Calling exceptional graduate students in the US working on hard problems in AI - Join us as a Research Interns to build the Next-Generation AI Models for India and the World! India has fallen behind in the race to develop its own cutting-edge LLMs—but we are changing that.

[Sudarshan Kamath @kamath_sutra]

“Deepseek makes it cheaper to train on India context”

Man stfu!

What’s up with taking new innovations and “making it for India”?

Does that even mean anything? Is that the extent our ambition goes to?

Why don’t we focus on what works in Deepseek or Llama or every frontier model out there and what doesn’t?

Why is it that these models perform amazingly well on benchmarks but they can’t solve simple problems any human can solve?

Let’s talk about research driven fundamental innovation? Online learning? Queuing systems? High frequency action reaction cycles? Long term thinking?

Let’s invest in companies that truly are trying to achieve AGI/ASI than “building for Indian context”?

Who’s up for thinking from first principles?

[Desi Batman @desibatman_27]

Watching test cricket and using Twitter together is an elite combo

[John Nguyen @__JohnNguyen__]

🥪New Paper! 🥪Introducing Byte Latent Transformer (BLT) - A tokenizer free model scales better than BPE based models with better inference efficiency and robustness. 🧵

[Gaurav Aggarwal @fooobar]

Back with another superlative result we just achieved at Reliance AI: SOTA Text-to-Speech (TTS) models for English and several Indic languages (work still on to cover as many as we can).

[Gaurav Aggarwal @fooobar]

Ours is a scalable and effective way to unlock the hidden reasoning potential of existing models. We are 23-30% (absolute) better than the instruction-tuned models of Qwen 0.5B and Llama 3.2 1B on GSM8K benchmark.

[Gaurav Aggarwal @fooobar]

𝗨𝗻𝗹𝗼𝗰𝗸𝗶𝗻𝗴 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴 𝗰𝗵𝗮𝗸𝗿𝗮𝘀 𝗼𝗳 𝗦𝗟𝗠𝘀 (𝗦𝗺𝗮𝗹𝗹 𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹𝘀)

Super pumped to share a "fresh off the oven" result in our quest to bring reasoning capabilities to the underdogs, aka SLMs - as small as 0.5B params.