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

Colleagues of the Applied Physics Laboratory and the Whiting School of Engineering are invited to the April talk in a speaker series co-presented by the Johns Hopkins Institute for Assured Autonomy (IAA) and the Computer Science Department, featuring national scholars presenting new research and development at the intersection of autonomy and assurance.
This talk will be “What’s Wrong with Large Language Models, and What We Should Be Building Instead” featuring Dr. Thomas Dietterich, Distinguished Professor Emeritus in the School of Electrical Engineering and Computer Science at Oregon State University, presenting in JHU’s Malone Hall 107 on Tuesday, April 16th at 10:30 a.m.
Dr. Dietterich abstract and bio are attached. This event is open to all APL and JHU staff, faculty, and students; please share!
ABSTRACT:
Large Language Models provide a pre-trained foundation for training many interesting AI systems. However, they have many shortcomings. They are expensive to train and to update, their non-linguistic knowledge is poor, they make false and self-contradictory statements, and these statements can be socially and ethically inappropriate. This talk will review these shortcomings and current efforts to address them within the existing LLM framework. It will then argue for a different, more modular architecture that decomposes the functions of existing LLMs and adds several additional components. We believe this alternative can address many of the shortcomings of LLMs.
About the Johns Hopkins Institute for Assured Autonomy: Led by APL and the Whiting School of Engineering, the IAA is becoming a nationally recognized center of excellence in autonomous systems, showcasing the robust portfolio of research and work from two premier divisions of JHU and creating strategic external partnerships. The IAA seeks to ensure the safe, secure, and reliable integration of autonomous systems and artificial intelligence (AI) in society. As autonomous systems proliferate, both physically and virtually, the institute seeks to ensure the systems will be trusted and safe in their operations, will withstand corruption by adversaries, and will integrate seamlessly into ecosystems and communities. In this burgeoning field, JHU strives to advance a clear vision for an autonomous future.

[@PaulTopping1]

A very balanced and constructive take on the current state of LLMs. Good talk.

[@mbrochh82]

Here's a ChatGPT summary:
- Dr. Dietterich is a distinguished professor emeritus at Oregon State University and a pioneer in machine learning.
- His research includes robust AI, human-AI systems, and sustainability applications.
- Dr. Dietterich has an AB from Oberlin College, an MS from the University of Illinois, and a PhD from Stanford University.
- Large language models (LLMs) like GPT-3 and GPT-4 have surprising capabilities but also many flaws.
- Industry is spending a lot to work around LLM flaws, but Dr. Dietterich suggests building modular AI systems instead.
- LLMs can answer questions and carry on dialogues, covering a wide range of human knowledge.
- LLMs can summarize documents, translate languages, write code, and learn new tasks from few examples.
- LLMs often produce incorrect or self-contradictory answers and can be socially inappropriate.
- Training and retraining LLMs are extremely expensive, with GPT-4 costing at least $100 million.
- LLMs lack easy update mechanisms and attribution for their answers.
- LLMs have poor non-linguistic knowledge and planning/reasoning skills.
- Dr. Dietterich advocates for a modular AI system with separate components for language, knowledge, and reasoning.
- He suggests using knowledge graphs to represent factual world knowledge and update it dynamically.
- The goal is to build AI systems that complement human capabilities rather than replicate them.
- Main message: Large language models are flawed and not fit for purpose out of the box; we should focus on building modular AI systems that can be updated easily and complement human skills.

[@JHUInstituteforAssuredAutonomy]

Thanks! Great use of ChatGPT!

[@someone3533]

@@JHUInstituteforAssuredAutonomy A commenter, the great meowmr9 has accused Dr. Dietterich that he is copying Dr. Yann LeCun and that there is nothing novel about what he is saying. Is this factually true? and if it is not true, then how can that be proven/demonstrated?

[@someone3533]

@@JHUInstituteforAssuredAutonomy Sy deleted my question.

[@JetBeetlePropulsion]

Thanks for the summary, this save me a lot of time.

[@kellymoses8566]

True AI must be able to continually learn.

[@NanheeByrnesPhD]

Thank you for this wonderful video. Even a non-tech person can easily understand the presentation.

[@JHUInstituteforAssuredAutonomy]

Glad it was helpful, we completely agree!

[@someone3533]

@@JHUInstituteforAssuredAutonomy A commenter, the young Titan meowmr9 says that Dr. Dietterich is copying Dr. Yann Lecun and that there is nothing novel about what he is saying. This is a serious accusation. Is it true what the young Titan says? And if it's not true, than how can that be proven/demonstrated?

[@maurobenetti6390]

I think the problem is they want to use a caffe machine to wash dishes. You should focus on reasoning capabilities rather than world knowledge.

[@AgentSmith-16384]

the audio quality is sh*t

[@garydecad6233]

Used cc