The transformative potential of AI search
If current AI models are imbued with the ability to think deeply for an extended period of time, it may revolutionize today's AI tech
This post is my attempt to summarize and understand Aidan McLaughlin’s post about imbuing large language models like GPT-4 with the ability to “search,” meaning the ability to think deeply for extended periods of time, in a manner similar to how a person might puzzle over something. Any errors are mine.
There has been remarkable progress in AI in recent years, driven by advances in deep learning and the ability to train ever-larger models on massive datasets. However, an equally important concept that has been somewhat overlooked is “search”—the ability for AI systems to think deeply and strategically over extended periods, similar to how a person might puzzle over something complex. Integrating search capabilities into AI models could be the key to unlocking the next major leap forward.
Imagine if AI systems could autonomously conduct research and make scientific discoveries. It might dramatically accelerate progress in fields like medicine and technology. This hinges on equipping AI models with the ability to perform extensive search. Current foundation models like GPT-4 lack this capability, limiting their potential impact. By integrating search, we could achieve a step-change improvement in AI performance, potentially upending current scaling laws.
The story of Leela Chess Zero illustrates the critical importance of search. Developed in 2019, Leela started with no chess knowledge beyond the basic rules. Through self-play and deep learning, it discovered innovative strategies that defied centuries of human chess wisdom. Leela even won the world computer chess championship.
However, its reign was short-lived. It was soon defeated by Stockfish, an older chess engine that combined deep learning with a superior search algorithm. This allowed Stockfish to efficiently evaluate many potential future positions. It wasn’t raw model size that ultimately triumphed, but the capacity for deep, strategic thinking.
Most foundation models today, like GPT-4, do not have search capabilities. They can’t “think ahead” and reason deeply about complex, multi-step problems. However, equipping these models with search could be utterly transformative. It could enable more efficient compute to solve challenging problems with unprecedented depth and accuracy.
Evidence suggests that effective search may not require drastically larger models. Well-designed search algorithms could allow current models to exhibit more intelligent, strategic behavior, suggesting we may not need to wait for future 100 trillion parameter models—the age of highly capable AI could arrive much sooner than anticipated.
A pharmaceutical company could use search-enabled AI to discover new drugs today, without massive investments in scaling. More broadly, search could drive an intelligence explosion as AI systems rapidly uncover improved algorithms and architectures.
A vision of search-first AI
Here is some speculation about how integrating search into current AI models could reshape the AI landscape:
Immediate real-world impact: Search capabilities could drive major advances in AI applications immediately, without waiting for future breakthroughs.
Accelerated AI progress: Search-enabled AI systems could autonomously conduct AI research, leading to recursive self-improvement and rapid capability gains.
Overcoming the data wall: Search relies more on reasoning than on large datasets, offering a path to advanced AI even if data becomes limited.
Accessible advanced AI: Organizations could obtain cutting-edge AI capabilities for specific high-value applications without massive investments in large clusters and infrastructure.
Conclusion
Integrating search into AI systems could herald a new era of strategically-minded AI, driving immense progress across various fields. While raw scale matters, the ability to think deeply and search through vast possibilities may ultimately lead to truly transformative artificial intelligence.
Good search/retrieval fells like it’s here.