Thomas Dietterich, "What’s Wrong with Large Language Models, and What We Should Be Building Instead"

Step-by-Step Tutorial:

Title: What’s Wrong with Large Language Models, and What We Should Be Building Instead

1. Introduction to the Seminar Series:

   • The seminar is co-sponsored by the Computer Science Department of the Whiting School of Research topics at the intersection of assurance and autonomy.
   • Dr. Dietrich, a distinguished professor emeritus in the School of Engineering and Computer Science at Oregon State University, will be the speaker.

2. Background of Dr. Dietrich:

   • Dr. Dietrich is a pioneer in the field of machine learning with over 200 publications and two books.
   • His current research topics include robust artificial intelligence and applications in sustainability.

3. Challenges with Large Language Models (LLMs):

   • Large language models have many flaws and shortcomings out of the box, making them not fit for many purposes.
   • Industry is spending a lot of money to work around these flaws, but building a different kind of system may be more beneficial.

4. Proposed Solution: Hybrid AI System:

   • Dr. Dietrich suggests building a hybrid AI system that includes a large language model as just one component.
   • This system would address the limitations of LLMs and provide a more robust and flexible solution.

5. Encouragement for Graduate Students:

   • Dr. Dietrich emphasizes that AI is not solved and there are still many problems that need to be addressed.
   • He urges graduate students to contribute to solving these challenges in the field.

6. Capabilities of Large Language Models:

   • Large language models like Chat GPT have shown surprising capabilities in carrying out conversations, answering questions, summarizing documents, and more.
   • However, they also have limitations such as producing incorrect or self-contradictory answers.

7. Challenges and Shortcomings of Large Language Models:

   • LLMs can produce dangerous and socially unacceptable answers, leading to ethical concerns.
   • Training and retraining LLMs are extremely expensive, with some models costing millions of dollars.

8. Addressing Flaws in Large Language Models:

   • Efforts are being made to improve LLMs through retrieval augmentation, reinforcement learning from human feedback, and direct preference optimization.
   • These approaches aim to enhance the accuracy and reliability of LLM outputs.

9. Moving Towards Modular AI Systems:

   • Dr. Dietrich advocates for building modular AI systems that combine linguistic skills with world knowledge and planning capabilities.
   • By separating components and leveraging modularity, the system can achieve better performance and address specific tasks effectively.

10. Future Directions and Research Areas:

    • Research areas such as building knowledge graphs, integrating evidence across multiple sources, and improving cognitive architectures are essential for advancing AI systems.
    • Collaboration between AI researchers, engineers, and domain experts is crucial for developing more robust and reliable AI solutions.

11. Conclusion and Call to Action:

    • Dr. Dietrich highlights the importance of rethinking AI systems to address current limitations and challenges.
    • He encourages researchers and practitioners to focus on developing innovative solutions that leverage the strengths of AI while mitigating its weaknesses.

By following these steps, you can gain insights into the key points discussed in the seminar and understand the proposed solutions for improving large language models and building more effective AI systems.