🎉 Our work "Random Representations Outperform Online Continually Learned Representations" accepted at #NeurIPS2024! W/ Ameya P. @sinha_shiven @oxfordtvg Ozan Sener Puneet Dokania Pre-print📜 precog.iiit.ac.in/pubs/NeurIPS-R… Code 🧑‍💻️ github.com/drimpossible/R… #ProfGiri Findings 🧵 👇

🥳 4 papers accepted at #NeurIPS2024 Workshops 🥳 #MathAI #MachineLearning #Compression #Representation #NeuralNetwork #UniReps UniReps #Algorithmic #Fairness #Evaluation AFME 2024 @ NeurIPS NeurIPS Conference Stay tuned for X 🧵 thread / 📜 pre-print / data / code! #ProfGiri

fund.blender.org

Looking at ICLR submissions with the lowest score - What a work of art! 🧵

academics are so cute. look how they write little papers for each other to read. they even format it all nice. aww

Duolingo is not a language-learning app, it’s an obsessive streak-creation app to make you feel happy for simply showing up. I haven’t met a single person who has learnt how to speak or even understand a language through Duolingo.

🤝🏾 Connections in #Bengaluru, please welcome six Precogs visiting Microsoft Research for next 10 days! 👋🏾 If you’re in Bengaluru feel free to connect 👋🏾 #IIITHStudentsRock 🤩 Fun fact: Students rarely dress this well when they are on campus; I almost didn't recognize them 😄

Actually so proud to see lab juniors doing so well! Amazing work

If you're attending please say hi, let's catch up!!

How do we remove the effect of incorrect training data on a trained model? Is retraining from scratch the best we can do? We show no! Cognac 🍷 is 8x faster than retraining from scratch, and works even when you discover as little as 5% of such data. How? 🧵👇

wait I love this

📢 Excited to share that COBIAS has been accepted at #WebSci25! 🎉 Our work aims to quantify the contextual #quality of LLM-bias #benchmarks. w/ Priyanshul Govil Hemang Jain Vamshi Krishna Aman Chadha Manas Gaur Sanorita Dey 📜 arxiv.org/abs/2402.14889 Findings 🧵⬇️

AI can generate correct-seeming hypotheses (and papers!). Brandolini's law states BS is harder to refute than generate. Can LMs falsify incorrect solutions? o3-mini (high) scores just 9% on our new benchmark REFUTE. Verification is not necessarily easier than generation 🧵

I worked a lot on my mental health and now I am no longer ambitious.

You can use 4o to generate fake receipts. There are too many real world verification flows that rely on “real images” as proof. That era is over.

Big update! I'm thrilled to share that I've joined the Center for Human-Compatible AI as a research intern. Looking forward to connecting with others interested in AI Safety around here!

The origin story of “AI as Normal Technology”, and lessons learned Many people have asked how the “AI as Normal Technology” paper came to be. This paper has been an (ongoing) journey for me and Sayash Kapoor in developing not just the substance of our arguments but also learning how

WHAT: Vamshi Krishna #MSThesis Defense, Towards Trustworthy AI: Frameworks for Evaluating Consistency in Language Models. 📜 Thesis & Papers: Links & details in thread ⏰ When 8:45 PM IST, 26th June 📍WHERE: In-person Precog Research group at IIIT Hyderabad Online QR Code #ProfGiri #IIITHStudentsRock

People are reporting that Gemini 2.5 keeps threatening to kill itself after being unsuccessful in debugging your code ☠️

✍🏽 “Dear Diary: I’m glad I joined Precog Research group at IIIT Hyderabad” Vamshi Krishna 🎓 #DualDegree | #Classof2025 “The cool thing about Precog is that we, as a group, also strive to improve — instead of focusing only on ourselves.” Read the full blog 👉 Link in comments #PrecogsRock #ProfGiri