Human Validation is Real Bottleneck to AGI
OpenAI's Head of Codex, Alexander Embiricos, argued that the primary constraint on AGI is not compute power but human bandwidth for validation and prompting. He stated that while users could benefit from AI tens of thousands of times daily, the effort required to prompt and recognize opportunities is the bottleneck. The solution, he suggests, is creating effortless, context-aware AI integrations that act preemptively rather than simply building more powerful models.
- The core of the bottleneck is not just typing speed, but the entire cycle of human-led prompt engineering, output review, and validation which is too slow to keep pace with AI's potential output. - Embiricos suggests a shift to "scalable oversight" or "agentic review," where AI agents are developed to validate the work of other AIs, thereby removing the human-in-the-loop speed constraints. - The proposed solution involves creating AI systems that are "default useful," capable of preemptively acting in a helpful way based on context, rather than waiting for explicit human prompts for every action. - This human bottleneck is a key challenge in the efficacy of Reinforcement Learning from Human Feedback (RLHF), a critical process for training models, as the quality and speed of human judgments directly limit model improvement. - The concept of agentic AI workflows, where AI can plan, act, and iterate, is a direct response to this bottleneck, aiming to make AI a collaborative partner rather than a tool that requires constant guidance. - Startups are already emerging to tackle this problem, with companies like Rapidata raising millions to create platforms for scalable, on-demand human feedback to accelerate AI development cycles. - While scaling laws have traditionally focused on compute, data, and model size, this perspective suggests that human-computer interaction bandwidth is an equally important, and currently limiting, factor in realizing AI's productivity gains. - As a practical example of overcoming this, OpenAI internally used an AI agent to help build the Sora Android app in just 18 days with a small team of engineers by having the AI analyze the existing iOS app and generate work plans.
