Descriptions:
This Y Combinator session brings together five researchers presenting papers at the frontier of AI development. The opening talk introduces asymmetric self-play for large language models — a framework inspired by AlphaZero where a “conjecturer” model generates RL training tasks and a “solver” model attempts them, with both trained simultaneously. The presenter argues this approach lets models improve beyond the ceiling imposed by human-generated training data, and draws a direct line toward significantly more capable systems.
Additional presentations cover AI for biology, real-time streaming RAG for voice agents from startup Giga, and Lean theorem proving for scientific research. A recurring theme is intelligence efficiency: “intelligence per sample” and “intelligence per watt” are named as the two central unsolved problems in current AI. The host compares in-context learning, LoRA fine-tuning at various ranks, and full SFT, showing each is optimal at a different data-volume regime — and argues that no current method matches the continuous, monotonically improving learning that characterizes human skill acquisition.
The session is a high-density research overview for practitioners and researchers interested in LLM training dynamics, self-improving AI systems, RAG infrastructure for low-latency applications, and the theoretical limits of scaling current deep learning paradigms.
📺 Source: Y Combinator · Published June 12, 2026
🏷️ Format: Keynote Launch
