SC
The Breakthroughs Needed for AGI Have Already Been Made: OpenAI Former Research Head Bob McGrew
@ Sequoia Capital
06/17/25
Takeaways
BM
Bob McGrew
06/17/25
@ Sequoia Capital
The advancements in large language models (LLMs) have enabled a language interface for robots, allowing them to perform a wider range of tasks more efficiently, such as laundry folding and packing egg crates.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
In 2025, reasoning will become the defining focus in AI development, as pre-training reaches diminishing returns.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
The diminishing returns in pre-training mean that significant increases in intelligence will require exponentially more compute resources, making it harder to achieve breakthroughs.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
Post-training focuses on developing model personality, which requires substantial human effort to create appealing and effective agent behaviors.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
The concept of agents will disrupt traditional economic models, as AI will provide services at the cost of compute, leading to an abundance of AI lawyers and other professionals.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
The future of AI should prioritize making services accessible and affordable, allowing everyone to have access to legal advice and other professional services.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
Robotics is an area ripe for innovation, as it currently feels like where AI was a few years ago, presenting opportunities for new companies to emerge. The advancements in robotics are nearing commercialization, with significant research challenges being addressed in the coming months or years.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
At OpenAI, we pursued robotics not for commercial motives, but to demonstrate the power of machine learning, which later influenced large language models.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
The introduction of large language models (LLMs) provides a language interface for robots, allowing for easier task descriptions and enabling robots to perform generic tasks more effectively.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
Companies like Physical Intelligence can now tackle a variety of tasks, such as laundry folding and packing egg crates, thanks to advancements in existing technology and research.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
Startups should focus on areas where deep understanding outside of the model is necessary, as these are safer spaces to operate without being overshadowed by larger labs.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
Enterprise applications often require a deep understanding of specific business contexts, which is where startups can find opportunities that larger labs may overlook.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
The frontier labs tend to focus on training models for new tasks, while many enterprise problems are too small for them to address individually, creating opportunities for startups.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
Proprietary data can be less valuable than the insights generated by infinitely smart agents using public data, as the latter can synthesize new information more effectively.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
The value of proprietary data lies in its ability to inform decision-making, but AI can replicate insights from that data without the original labor involved in gathering it.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
Coding has rapidly evolved, with AI tools significantly impacting how software is developed, leading to nonlinear progress in the field.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
The future of coding will involve a mix of traditional coding practices and AI-assisted development, with a focus on collaboration between human programmers and AI agents.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
The distinction between engineers and researchers at OpenAI was minimized to foster collaboration and understanding of the full stack, enhancing the quality of research and implementation.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
Different generations use ChatGPT in varied ways: older users see it as a search tool, while younger users view it as a life coach or even an operating system for their tasks.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
The total addressable market for ChatGPT includes any task that requires thought or action that users prefer to delegate to AI, raising concerns about dependency on technology. AI should enhance your expertise in areas of interest and alleviate the burden of mundane tasks.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
My 8-year-old son uses ChatGPT not as a friend, but as an expert to explain things to him, which sparks his curiosity and enthusiasm.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
When I was a kid, learning about coin collecting would have taken weeks, but now my son can access information instantly through ChatGPT.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
Preparing the next generation for AI involves teaching them the process of learning and encouraging them to have ideas and projects.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
It's important for children to understand the basics before using AI tools to extend their capabilities.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
Using AI in daily life, like for scheduling, can make tasks more enjoyable and efficient.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
In an increasingly agentic world, the ability to execute offensive work has increased, leading to more security threats.
BM
Bob McGrew
06/17/25
@ Sequoia Capital
Startups have an opportunity to disrupt existing businesses by leveraging AI technologies that organizations are not yet set up to adopt.