The Breakthroughs Needed for AGI Have Already Been Made: OpenAI Former Research Head Bob McGrew

Bob McGrew, OpenAI’s former chief research officer, argues the fundamental breakthroughs needed for AGI are already in place and agent services will be priced at compute cost.

• McGrew argues that by 2035, no new fundamental AI concepts beyond transformers, scaled pre-training, and reasoning will have emerged.
• Pre-training hits diminishing returns because intelligence scales log-linearly with compute, requiring exponential hardware investment for each increment.
• Agent services will be priced at compute cost due to near-infinite supply, collapsing economic moats in law, medicine, and other high-skill fields.
• The 01-to-03 gap—tool use inside chain-of-thought—took six months; McGrew expects future reasoning gains to be progressively harder.
• Frontier labs effectively hold options on the most valuable application-layer segments; safe startup territory requires deep enterprise context outside the model.
• Proprietary datasets built from human labor are at risk because infinitely patient AI agents can reconstruct equivalent knowledge from public data.
• Robotics is McGrew’s top startup bet: LLM language interfaces and strong vision encoders now let companies like Physical Intelligence solve diverse tasks in months vs. years for single tasks.
• In cybersecurity, offensive AI lowers attack costs dramatically, making agentic defense stacks (minimal human input) the necessary response.

2025-06-17 · Watch on YouTube