Bob McGrew: AI Agents And The Path To AGI

Key Concepts

AGI (Artificial General Intelligence), LLMs (Large Language Models), Scaling Laws, Reasoning, Test Time Compute, Agents, Distillation, Forward Deployed Engineer, Robotics, Foundation Models, Zero to One Phase, Innovators, Pre-training, Data Wall, Reliability, UI (User Interface), Personalization, Bottlenecks, S-Curves.

Early Days at OpenAI and the Pursuit of AGI

Bob McRu, formerly Chief Research Officer at OpenAI, recounts his journey into AI, initially aiming to build a robotics startup based on deep learning. He joined OpenAI to learn from smart people and explore deep learning's potential. OpenAI's initial goal was AGI, but the early approach of focusing on research and papers felt academic. Early projects included teaching a robot hand to solve a Rubik's Cube, aiming for generalization, and solving Dota 2, which reinforced the belief that scale was key to improving AI. Alec Radford's work on language models, specifically GPT-1, involved the simple objective of predicting the next token, which surprisingly led to coherent text generation. GPT-2 and GPT-3 then scaled up these ideas with larger datasets and a focus on generalization.

The Significance of Scale and OpenAI's Culture

OpenAI pioneered the concept of scale in AI. Unlike Google Brain's "let a thousand flowers bloom" approach or DeepMind's top-down planning, OpenAI found a middle ground. There was no centralized plan, but research leadership, like Ilia and Dario, guided the direction, emphasizing scale as a way to improve ideas. This culture avoided the academic obsession with individual credit, instead fostering collaboration and focusing on internal reputation. Early on, OpenAI even cited papers as "OpenAI" to avoid authorship disputes.

Scaling Laws and Bottlenecks

Scaling laws are prevalent in AI progress. Achieving the first working version of a model, like DALL-E, is a significant challenge before scaling laws can be applied. Scaling involves both increasing the scale itself (a systems, data, and algorithmic problem) and improving the slope of the scaling law through better architectures and optimization algorithms. A "data wall" exists, limiting the effectiveness of simply scaling pre-training with larger datasets. Reasoning and test time compute are new mechanisms to overcome this bottleneck. McRu compares this to Moore's Law, where progress continues through different mechanisms as bottlenecks are encountered and solved.

Reasoning, Agents, and Reliability

Reasoning, as implemented in models like GPT-4 and Gemini, allows models to think longer and provide better answers. This unlocks the potential for agents to perform actions on users' behalf. Reliability is crucial for user trust in agents. Increasing reliability from 90% to 99% or higher requires significant increases in compute, which can now be achieved through reasoning and longer thinking times.

Distillation and AI Startups

Distillation involves training smaller models to perform almost as well as larger models on specific distributions of user input. This is becoming increasingly common, with labs focusing on creating smaller, faster models. For AI startups, McRu advises starting with the best available model to exploit frontier AI capabilities. Once the product works, distillation can be used to reduce costs.

The Future of AI and its Impact

McRu is skeptical of AI achieving deep emotional connections, like in the movie "Her." He sees more potential in AI assistants that understand user preferences and context, particularly in the workplace. He highlights the need for better UI and software to make AI more accessible and useful. Drawing parallels to Palantir's forward-deployed engineers, he emphasizes the importance of engineers working closely with customers to build tailored solutions.

AI and Education

McRu discusses the implications of AI for education, particularly coding. While AI can code, he believes it's still important for children to learn coding to develop intuition and understand what's possible. He envisions two primary roles in the future: the "lone genius" leveraging AI and the "manager" leading AI-powered teams. He draws an analogy to the automation of farming in the 19th century, where new jobs emerged that were unimaginable at the time.

Robotics and Scientific Advancement

McRu believes robotics companies are where LLM companies were five years ago, predicting a "ChatGPT moment" for robotics within five years. Companies like Skilled AI and Physical Intelligence are building foundation models for robots. He anticipates breakthroughs in robotics that will lead to increased reliability and market scope. He also believes that automating the scientist/innovator could accelerate scientific advancement, although new bottlenecks will likely emerge.

Notable Quotes

• "I think in in the big picture we're we're reaching that bottleneck for pre-training and data but now we have this new mechanism with reasoning and test time compute."
• "Scaling is not easy um it is in fact probably the Practical problem in any sort of model building and it's a systems problem it's a data problem it's an algorithmic problem even if you're just trying to scale the same architecture."
• "I think the what we're going to see out of reasoning out of long thinking is that it's really going to unlock uh the possibility of agents to do actions on your behalf which you know has sort of always been possible but it's just never been quite good enough."
• "What AI desperately needs right now is like you said the UI the software it's just building software and if you can put that in a package that a particular person really really needs I feel like that's one of the big things that we learned at palun."
• "I think those will be the two jobs of the future genius and manager."

Synthesis/Conclusion

The conversation with Bob McRu provides a deep dive into the current state and future trajectory of AI. Key takeaways include the importance of scaling laws, the emergence of reasoning as a new mechanism for improvement, the potential of AI agents, and the need for better UI and software to drive adoption. McRu emphasizes the importance of collaboration, the value of tailored solutions, and the potential for AI to transform various industries, including robotics and scientific advancement. While challenges and bottlenecks remain, he expresses optimism about the future of AI and its impact on society.