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

AI Team

May 13, 2024

Here are this week's important vid from Johns Hopkins Institute for Assured Autonomy

The transcript discusses a seminar on AI and machine learning, featuring Dr. Tom Dietrich, a pioneer in the field. The talk centers around large language models (LLMs) and their limitations, emphasizing the industry’s significant investment to mitigate these flaws. Dr. Dietrich critiques current LLMs for their inability to be truly fit for purpose “out of the box,” proposing instead the development of modular AI systems that incorporate LLMs as just one component among others. He highlights the ongoing challenges with LLMs including incorrect or dangerous outputs, their expensive training processes, and their tendency to produce outdated or unverifiable information. The overarching message is that while LLMs like ChatGPT exhibit impressive abilities, they are fundamentally flawed in ways that necessitate innovative solutions, especially in how they integrate and interact with more extensive data and contextual frameworks.

Value Categories

Use Case

AI Marketing: The seminar discusses the current state and potential of LLMs, serving as a form of academic and professional engagement to promote understanding and further development in the field.
Content Creation: Provides insights into the challenges and advancements in AI, particularly around LLMs, contributing to the broader discourse and knowledge pool in AI research and application.

Value Proposition

Cut Costs: Proposes more efficient, modular AI systems that could potentially reduce the high costs associated with training and maintaining LLMs.
Increase Productivity: By addressing the shortcomings of LLMs and suggesting more robust, modular systems, there’s potential to enhance productivity in AI-driven tasks and applications.
Save Time: Modular systems might streamline processes that currently rely on LLMs, reducing time spent managing LLM limitations and improving responsiveness.

Type

AI Text Generation: Central to the discussion, with a focus on improving how LLMs manage and generate text-based content.
AI Image Generation and AI Video Generation: Not directly discussed, but the principles of modular improvement could apply to these areas as well.
AI Voice Generation: Although not explicitly covered, the concepts of modular AI systems suggest potential applications to voice-driven AI technologies as well.