Martin Keen debunks five common AI myths, covering topics from reduced AI hallucinations and the misinterpretation of AI's "thinking" process, to the rising costs of AI inference, the limitations of large context windows, and the current challenges to fully autonomous AI agents.
This session features two talks on optimizing and verifying AI reasoning. Hongxiang Fan discusses cross-stack co-design for efficient AI, focusing on Test-Time Scaling (TTS) challenges, optimal verification granularity, and system-level optimizations for edge deployments. Nagarajan Natarajan introduces 'Advancing Verified Reasoning' with the InterVent platform, aiming to ensure AI agents comply with complex policies through formal verification, dynamic steering, and leveraging verification signals for training. Both emphasize addressing the computational and reliability costs of advanced AI.
Eric Zelikman, formerly of Stanford and xAI, discusses his research on AI reasoning (STaR, Q-STaR) and introduces his new venture, humans&. He argues for a paradigm shift from building AI with pure IQ to AI with EQ, focusing on long-term memory, human collaboration, and empowering users to achieve their full potential.