Fireside Chat, Interview, Conference Presentation
OpenAI's Greg Brockman: Why Human Attention Is the New Bottleneck
- Demand for intelligence is projected to be unlimited, with compute availability expected to approach zero by 2026 regardless of ramping speed, ensuring demand will consistently outpace supply.
- Neural network capabilities are predicted to improve continuously without an apparent wall, driven by increased compute and constant architectural and algorithmic innovations.
- OpenAI estimates it is approximately 80% toward achieving functional AGI and is currently leading long-term research focused on fundamental algorithmic improvements and paradigm shifts.
- Model capabilities in software engineering are advancing rapidly, with agentic coding tools shifting from writing 20% of code to 80% within December, and software generation expected to surpass human ability in many contexts.
- The technology is evolving from a specialist tool for software engineers to a general-purpose utility for anyone performing computer work this year, with future models expected to solve harder problems and generate new knowledge.
- A significant one-time investment in providing AI with sufficient context is deemed necessary now to enable systems that can harness better capabilities and solve complex problems.
- Operational bottlenecks have shifted to sharing and governance, prompting plans to externalize governance solutions, dashboards, widgets, and bots to the public while prioritizing internal refinement before customer release.
- Organizational structures are anticipated to shift toward flatter, smaller teams within a decade, enabling solopreneurs to build impressive businesses and fundamentally changing the definition of a company.
- Work is expected to become easier and more competitive, with a mechanics of work shift comparable to the transition from quill to text messaging, while the definition of a consumer evolves toward goal-oriented and proactive interaction.
- Science is projected to experience a renaissance in the coming year or next year, with existing examples of AI deriving previously unsolvable physics formulas and similar advancements anticipated in biology and physical intelligence.
- Human attention is identified as the single most important future bottleneck as execution becomes easier, necessitating systems built to account for human factors and attention scarcity.
- Security and trust are prioritized through the expansion of trusted access programs for cybersecurity, with intent to tune mitigations upon wider release to ensure positive deployment and build enterprise sales muscle.
- The internet is moving toward a more secure regime requiring worldwide effort, while OpenAI aims to build an AGI that is trustworthy and useful for both personal and work applications.