The Jim Rutt Show
EP 192 David Krakauer on Science, Complexity and AI
Episode guests
Podcast summary created with Snipd AI
Quick takeaways
- Complex systems encode a coarse-grained history and act as theories that capture aspects of reality.
- Deep learning models primarily serve as predictive engines and may have limited potential for scientific discovery.
- Large language models like GPT-4 show impressive abilities in generating creative output and analyzing information.
- Scientific breakthroughs have historically been driven by constraints and parsimony, not comprehensive knowledge of microscopic details.
Deep dives
Complex systems encoding adaptive history
Complex systems, unlike purely physical systems, encode a coarse-grained history of their adaptive behavior. What distinguishes complex systems is the presence of a schema that encapsulates this history.
Construction as a way to understand complex systems
The concept of construction, which is a way of encoding and representing reality, offers a useful lens to understand complex systems. Complex systems can be seen as theories of the world, where each element within the system, whether it's a microbe or a human brain, acts as a theorizer that captures and encodes certain aspects of reality.
The potential of deep learning models as theories
By training deep neural networks, we are essentially constructing a theory or rule system of the phenomenon being modeled. These models represent a form of schema, encoding the regularities and patterns observed in the data. However, it should be noted that these models primarily serve as predictive engines, providing outputs based on inputs, and their ability to generate new insights and discoveries may be limited.
Potential avenues for scientific progress and innovation
While deep learning models excel at prediction and providing reference material, they may not be effective at driving scientific discovery. The nature of these models as libraries of established knowledge makes them less conducive to innovative and out-of-the-box thinking. However, the interaction and composition of different models, along with human creativity and insight, may lead to new theoretical frameworks and insights in various domains.
The Power of Language Models in Manipulating Language
Large language models like GPT-4 and GPT-5 have shown impressive abilities to manipulate language and generate creative work. They can summarize complex literature, compare and contrast different concepts, and even create unique ideas. The models demonstrate remarkable compositionality and can potentially be used for generating creative output in various domains. Additionally, these models have shown potential for analyzing and synthesizing information, making them useful tools for research and analysis. However, it's important to note their limitations, as they lack the ability to understand geometry or solve complex scientific problems on their own.
The Importance of Constraints in Scientific Revolutions and Breakthroughs
Historically, scientific breakthroughs and revolutions have been driven by constraints, rather than excess power or computational resources. Examples such as Darwin's theory of evolution and Mendeleev's periodic table demonstrate how simple principles and observations led to major scientific advancements. These achievements were not dependent on comprehensive knowledge of microscopic details, but instead on recognizing patterns, observing repetition, and proposing elegant explanations. It raises questions about the role of constraints and parsimony in scientific progress, and emphasizes the importance of understanding and leveraging simple processes that can generate complex phenomena.
Assessing the Risks and Potential of AI
The discussion around the risks and potential of AI often lacks empirical precedent and is often sensationalized. While there are significant risks associated with AI, such as misuse of narrow AI or unintended negative consequences, it is crucial to approach the topic with an informed and measured perspective. Precedent from past technologies, like genetic engineering and nuclear weapons, shows that responsible regulation and self-imposed moratorium can effectively manage risks. Rather than dwelling on doomsday scenarios, a more productive approach is to identify reasonable risks, adapt regulations appropriately, and explore the potential positive applications of AI, like developing info agents to filter and curate information.
Remember Everything You Learn from Podcasts
