Episode #403: Unlocking AI’s Brain: Knowledge Graphs, LLMs, and the Future of Reasoning

Crazy Wisdom

In this episode of the Crazy Wisdom Podcast, host Stewart Alsop welcomes Chia Yang, co-founder of whyhow.ai, a company specializing in data infrastructure and AI-powered knowledge graphs. They discuss the pivotal role of knowledge graphs in AI, particularly in enhancing structured search and reasoning, contrasting them with more stochastic systems like large language models (LLMs). Chia explains how knowledge graphs allow for more structured, reliable connections between data, and how this impacts the development of production-grade AI systems. He also touches on the limitations of LLMs, the significance of neurosymbolic approaches, and the future of AI reasoning. For further resources, Chia encourages listeners to visit whyhow.ai, check out their Medium articles, and join the discussion on their Discord channel.

Check out this GPT we trained on the conversation!

Timestamps

00:00 Introduction to the Crazy Wisdom Podcast

00:26 Understanding Knowledge Graphs

02:32 The Role of Knowledge Graphs in AI

05:08 Challenges and Limitations of LLMs

09:51 Production Grade Systems and SOPs

13:17 Competency Crisis and Real-World Problems

18:11 The Future of Human and Machine Collaboration

21:03 Exploring Social Inequality and Learning Challenges

21:57 The Importance and Complexity of Data

22:44 Understanding Knowledge Graphs and LLMs

24:29 Building Practical Systems with LLMs

25:42 The Evolution of Knowledge Graphs

29:12 Technical Aspects of the Platform

31:52 Philosophical Insights on Language and AI

36:48 Future Milestones and Beta Program

38:24 Final Thoughts on Knowledge Graphs

Key Insights

1. The Power of Knowledge Graphs: Knowledge graphs are central to creating structured representations of data, enabling more reliable and hierarchical relationships between information. They play a crucial role in enhancing AI systems' ability to retrieve relevant information and reason through complex problems, contrasting with the more flexible but less deterministic nature of LLMs.
2. Limitations of Large Language Models: While LLMs have revolutionized AI with their probabilistic approach, they often struggle with reasoning and structured outputs. Their reliance on semantic similarity leads to occasional inaccuracies, known as hallucinations, highlighting the need for more structured, rule-based systems like knowledge graphs to complement their capabilities.
3. Neurosymbolic Systems as the Future: There is growing interest in neurosymbolic systems, which combine the strengths of both neural networks and symbolic reasoning systems. This approach promises to overcome the limitations of LLMs by incorporating structured knowledge representation, improving AI’s ability to perform reliable and scalable reasoning tasks.
4. Production-Grade AI Systems Require Structure: As Chia explains, building AI systems ready for large-scale, real-world applications requires more than just probabilistic models. Introducing structured frameworks, like those provided by knowledge graphs, allows for more predictable and controlled outputs, which are essential for systems that need to operate reliably in high-stakes environments.
5. Human Expertise and AI Collaboration: Non-technical domain experts, such as doctors or engineers, are key to building effective AI systems. Their specialized knowledge helps refine how data is represented within a knowledge graph, making it more accurate and useful for specific industry applications, demonstrating the importance of human-AI collaboration.
6. The Role of SOPs in AI: Standard operating procedures (SOPs) are crucial not only for humans in high-stakes professions but also for AI systems. Chia draws parallels between the structured guidance provided by SOPs and how AI can benefit from similar rule-based structures to ensure accurate, repeatable outcomes, especially in critical fields like healthcare and engineering.
7. The Evolution of AI Reasoning: Chia highlights that while LLMs are powerful, the next big leap in AI will likely come from improving reasoning capabilities. Autonomous agents and AI systems that can follow structured paths and make decisions based on a combination of stochastic and symbolic methods will bring AI closer to AGI (artificial general intelligence).