time to cook

Thanks a lot! Really enjoyed diving into your paper 🙌

Reward model to policy model: karm karo, fal ki chinta mat karo

You training a model and it doesnt work after spending several thousand dollars. You get PTSD: Post Training Stress Disorder 🤣

LLMs is a unique experiment whose goal is to see how far people who aren't incapable of understanding but can live by following patterns can go.

and that’s how i got reward hacked

I’ll be at NeurIPS 2025 in San Diego presenting Aryabhata, a 7B reasoning model with SOTA JEE Math performance. If you: - work on reasoning LLMs - are into reinforcement learning - are trying to make machines think – or just want to chat I’d love to meet you. DMs open!

I’ll be at NeurIPS 2025 in San Diego presenting Aryabhata, a 7B reasoning model with SOTA JEE Math performance. If you: - work on reasoning LLMs - are into reinforcement learning - are trying to make machines think – or just want to chat I’d love to meet you. DMs open!

From writing my first-ever paper to presenting it at NeurIPS. Still can’t believe I’m stepping into a conference I’ve followed for years.

Fanboy moment with Daniel Han at NeurIPS Conference

Going through 100s of papers on LLM pre-training, mid-training, and post-training at NeurIPS… and this single line captures the whole conference: “To machines, human language is not language, but distribution.”

Ritvik Rastogi Damn saw Sanjeev Arora asking you questions

Bye bye, San Diego! NeurIPS week done! Papers, posters, conversations, and a lot of perspective Grateful for the learning and the people Until next time!

Came across this super-interesting paper at a poster session at NeurIPS: A benchmark that evaluates LLMs on algorithmic reasoning using OEIS integer sequences. They ask LLMs to generate Python code that computes OEIS sequences, No lookup tables allowed; runtime limits enforced.

2025 Wrapped: Grateful for a year of steady progress, a lot of learning, and slow, deliberate compounding. • Read ~234 research papers • Wrote notes and published consistently (223 days) • Started learning boxing • Stayed consistent with training (173 gym days) • First

spent the last few months building this application 👇 Inksphere — an immersive ebook reader that acts as your smart reading companion. ✨ Generates consistent illustrations 📖 Profiles characters automatically ⏱️ Tracks story timelines all designed to deepen your immersion in

we will demo SWE Agent which beats claude code and codex on a complex codebase with RL post-training and harness engineering learn more at India AI Summit. visit us at @athenaagentrl booth to talk about Agentic Engineering