AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

Bob McGrew: The Breakthroughs for AGI Are Already Here

· Training Data

Bob McGrew, former Chief Research Officer at OpenAI, argues on Training Data that the fundamental conceptual breakthroughs needed for AGI—transformer-based language models, large-scale pre-training, and reasoning—have already been discovered, and that work through 2035 will be about scaling and refining these ideas rather than inventing new paradigms. He frames the current moment as a "trifecta" of pre-training (now facing diminishing returns but still critical for new architectures and inference efficiency), post-training (a "thick" problem about model personality requiring product and human-nature expertise rather than pure scaling), and reasoning (the dominant axis of 2025, with significant low-hanging fruit but increasingly hard problems once initial overhang is consumed). McGrew makes several contrarian market claims: AI agents will be priced near the cost of compute because frontier model competition commoditizes intelligence, eroding the economic scarcity that makes professions like law expensive. Proprietary data is overvalued when infinitely patient AI can synthesize equivalent knowledge through embodied labor—though data tied to specific customer relationships (e.g., a financial advisor's knowledge of your portfolio) remains genuinely useful. He sees robotics as finally reaching commercialization thanks to LLMs providing a language interface and strong vision encoders, citing Physical Intelligence as evidence that what took OpenAI years for a single task now takes months for many. On coding, he bifurcates the field into agentic background work (bug fixes, refactors, COBOL translation) where clear success criteria exist, versus human-driven design work requiring taste, architecture judgment, and non-obvious consequence handling—a divide unlikely to collapse soon. He also shares personal parenting examples showing how his eight-year-old uses ChatGPT as an expert tutor for coin collecting and Arduino projects, and offers leadership lessons from managing elite researchers at OpenAI, including using OpenAI as paper first-author to defuse academic credit disputes. He mentions investments in Distill (enterprise context extraction) and Outtake (agentic cybersecurity).

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.