Interview, Fireside Chat
Where We Go From Here with OpenAI's Mira Murati
- Building effective products on top of AI models and managing the systems and engineering problems involved in scaling, efficiency, and accessibility are expected to be incredibly difficult and a primary challenge.
- Future AI systems are predicted to evolve from current "intern-level" reliability into agents capable of autonomous operation, with a trajectory toward humans providing direction and oversight while reducing repetitive work.
- Programming is expected to become more accessible through natural language interfaces, enabling a shift where users collaborate with models as partners or co-workers rather than interacting with finite state machines.
- The industry anticipates a range of specialized models for specific use cases alongside general systems designed to know user context, goals, and life to guide and coach users.
- Foundation models are expected to grow with bigger models, incorporating additional modalities like images and video to achieve a more comprehensive understanding of the world.
- Current AI limitations regarding reliability and hallucinations are addressed through strategies such as reinforcement learning with human feedback, increased pre-training, and the introduction of browsing to cite fresh information.
- Strategic decisions to deploy models publicly for real-world feedback are considered crucial for aligning, safety, robustness, and reliability, making development in a vacuum impossible.
- The speaker plans to shift focus from theoretical mathematics to building practical things, expecting interest in mathematics to grow over time while acknowledging the difficulty of creating good products.
- Making technology widely available is expected to lead to emergent use cases, while the speaker notes that many other breakthroughs and advancements are still needed to reach AGI.
- As model power increases, the challenge of super alignment is identified as a significant technical difficulty requiring dedicated team focus to manage fears of misalignment with human intentions.