Frontier Bits: AI Lectures and Papers 001 - Koramangala, Namma Bengaluru

Frontier Bits is Soket's curated series of deep technical conversations for people building at the frontier of AI. Each session brings together researchers, engineers, and practitioners to dissect the latest advances in foundation models, training systems, inference, multimodal…

About

Frontier Bits is Soket's curated series of deep technical conversations for people building at the frontier of AI. Each session brings together researchers, engineers, and practitioners to dissect the latest advances in foundation models, training systems, inference, multimodal AI, and emerging research. These are not introductory talks or product presentations. They are technical deep dives designed around first-principles thinking, rigorous discussion, and the exchange of ideas between people actively building the next generation of AI systems. This is our first offline edition and we are doing it in collaboration with the amazing community of Women in Product Who should attend Researchers, ML engineers, infrastructure engineers, AI practitioners, and technical leaders working on or supporting frontier AI systems. If your interests include large-scale pretraining, post-training, reasoning models, agentic systems, inference optimisation, multimodal models, model evaluation, AI infrastructure, harness engineering or cutting-edge AI research, you'll find yourself among peers. Agenda 7:00 – 8:00 PM Research Paper Deep Dive: On-Policy Distillation We will collectively try to cover the evolution of OPD from three resources. We expected to at-least cover the first in this session. Please follow Open Discussion Protocol (Link) for this session. 1. ThinkingMachine - On-Policy Distillation Blog 2025 2. HuggingFace - Unlocking On-Policy Distillation for Any Model Family 2025 3. Nvidia - X-Token: Projection-Guided Cross-Tokenizer Knowledge Distillation Paper 8:00 – 9:00 PM Technical Lecture & Discussion: Nvidia Megatron Framework for large scale training A deep-dive session on Megatron, Nvidia's large scale model training framework. Paper: https://arxiv.org/abs/1909.08053 9:00 – 10:00 PM Networking Continue the conversations over snacks and drinks with fellow researchers, engineers, founders, and builders working across the AI ecosystem. Attendance Seats are intentionally limited to keep the discussions interactive and high signal. Applications will be reviewed, and the exact venue will be shared with accepted attendees. And one final litmus test: if the phrase "research paper reading session" made you more curious than bored, you'll probably feel right at home at Frontier Bits.