Interview, Fireside Chat, Podcast
The Breakthroughs Needed for AGI Have Already Been Made: OpenAI Former Research Head Bob McGrew
- Capabilities are expected to feel "super fast" and "super exciting" for the next five years, with 2025 identified as the "year of reasoning" before progress slows as technical overhang is exhausted.
- As reasoning technology matures, the overhang of compute, data, and algorithmic efficiency will be consumed, while pre-training shifts toward improving architectures for inference efficiency and context usage, a process requiring years and new data centers.
- By 2030, fundamental concepts for intelligence may be fully discovered, including transformers, scaling, reasoning, and multimodal capabilities, though this prediction is noted with uncertainty and the hope for a different trajectory by 2035.
- Commercial AI agents are forecast to be priced at the cost of compute due to competition, with foundation models holding options on valuable application layer locations while application companies survive only with moats like network effects or deep domain expertise.
- Robotics is expected to be commercialized within "months or a small digit year" as research challenges regarding tasks like laundry folding and packing are solved, transitioning from specific puzzles to practical applications.
- A long-term development mix is anticipated for the next "year or two," featuring a combination of user-in-IDE coding and background agentic coding, with "vibe coding" projects likely being rebuilt by professionals as codebases grow.
- Within the next couple of years, answers are expected regarding whether agents can pass work between one another without human intervention based on detailed specifications, alongside a shift where AI assistants handle calendaring to make such tasks more engaging for individuals.
- Educational approaches are expected to prioritize teaching learning processes and agency over specific skills like coding, allowing individuals to become experts in desired areas while AI removes the burden of mundane tasks.
- The shift to agentic AI is predicted to disrupt businesses similarly to the transition from web to mobile, where replicating technology occurs faster than changing internal processes, and AI will increasingly recreate proprietary data, reducing its value.