Do you think that ChatGPT can reason?

This podcast episode features Subbarao Kambhampati, an experienced AI researcher, who emphasizes the limitations of large language models (LLMs) in reasoning and knowledge generation. Through a detailed exploration of LLM design as advanced n-gram models, he dissects the common misconceptions surrounding their capabilities, highlights the significance of external verifiers, and advocates for a more critical approach to AI research, urging researchers to embrace logic and skepticism in their work.

Outlines

Sign in to continue reading, translating and more.

Continue

Machine Learning Street Talk

The Limits of Reasoning in Large Language Models

Brave Search API and Retrieval Augmented Generation

Introduction to Subbarao Kambhampati and His Work

LLMs as N-gram Models on Steroids

The Factuality and Reasoning Claims of LLMs

Why Do People Believe LLMs Reason?

The Limits of Reasoning: Standardized Tests and Diagonalization Arguments

LLMs and Planning Problems

Defining Reasoning: Deductive Closure and Beyond

The Illusion of Reasoning: Web-Scale Data and the Joke-Explaining LLM

LLMs and Ciphertext Decoding: Another Diagonalization Argument

The Importance of Skepticism and Diagonalization Arguments in LLM Research

The Sociology of LLM Research: Hype and the Pursuit of Citations

The "Dumb AI" Argument and the Illusion of Anthropomorphism

LLMs and Creativity: Inductive Leaps and the Role of Verification

The Creativity Gap and Combinatorial Creativity

The Verification Gap and the Importance of Instance-Level Correctness

The Importance of Human Data and the Seed Vault Analogy

Synthetic Data and the Blind Leading the Blind

The ARC Challenge and the Python Interpreter Advantage

LLMs as Critics: The Limits of Self-Reflection

The LLM Modulo Framework: A Generate-Test Approach to Reasoning

The End-to-End Predictive Model Argument and the Cost of Verification

Fine-Tuning, Chain of Thought, and the Amortization Argument

The "Advice Taking" Problem and the Limits of Chain of Thought

Computational Complexity and the Misinterpretation of LLM Capabilities

LLMs and Turing Completeness: An Orthogonal Question

The "Later Models Will Do Everything" Argument and the Architecture Lottery

LLMs as Tools: The LLM Modulo Framework and the Need for External Verifiers

The Future of AI: Generalist vs. Specialized Models and Agentic Systems

LLM Modulo and the Future of AI: Bridging the Gap Between Generalists and Specialists

Advice for Young Researchers: Embrace Logic and Skepticism

Do you think that ChatGPT can reason?

Machine Learning Street Talk

00:00The Limits of Reasoning in Large Language Models

The Limits of Reasoning in Large Language Models

01:28Brave Search API and Retrieval Augmented Generation

Brave Search API and Retrieval Augmented Generation

02:06Introduction to Subbarao Kambhampati and His Work

Introduction to Subbarao Kambhampati and His Work

03:00LLMs as N-gram Models on Steroids

LLMs as N-gram Models on Steroids

08:19The Factuality and Reasoning Claims of LLMs

The Factuality and Reasoning Claims of LLMs

11:08Why Do People Believe LLMs Reason?

Why Do People Believe LLMs Reason?

13:18The Limits of Reasoning: Standardized Tests and Diagonalization Arguments

The Limits of Reasoning: Standardized Tests and Diagonalization Arguments

15:02LLMs and Planning Problems

LLMs and Planning Problems

19:13Defining Reasoning: Deductive Closure and Beyond

Defining Reasoning: Deductive Closure and Beyond

22:50The Illusion of Reasoning: Web-Scale Data and the Joke-Explaining LLM

The Illusion of Reasoning: Web-Scale Data and the Joke-Explaining LLM

24:51LLMs and Ciphertext Decoding: Another Diagonalization Argument

LLMs and Ciphertext Decoding: Another Diagonalization Argument

26:59The Importance of Skepticism and Diagonalization Arguments in LLM Research

The Importance of Skepticism and Diagonalization Arguments in LLM Research

28:27The Sociology of LLM Research: Hype and the Pursuit of Citations

The Sociology of LLM Research: Hype and the Pursuit of Citations

31:37The "Dumb AI" Argument and the Illusion of Anthropomorphism

The "Dumb AI" Argument and the Illusion of Anthropomorphism

32:48LLMs and Creativity: Inductive Leaps and the Role of Verification

LLMs and Creativity: Inductive Leaps and the Role of Verification

41:52The Creativity Gap and Combinatorial Creativity

The Creativity Gap and Combinatorial Creativity

45:44The Verification Gap and the Importance of Instance-Level Correctness

The Verification Gap and the Importance of Instance-Level Correctness

49:58The Importance of Human Data and the Seed Vault Analogy

The Importance of Human Data and the Seed Vault Analogy

50:27Synthetic Data and the Blind Leading the Blind

Synthetic Data and the Blind Leading the Blind

53:45The ARC Challenge and the Python Interpreter Advantage

The ARC Challenge and the Python Interpreter Advantage

56:58LLMs as Critics: The Limits of Self-Reflection

LLMs as Critics: The Limits of Self-Reflection

59:04The LLM Modulo Framework: A Generate-Test Approach to Reasoning

The LLM Modulo Framework: A Generate-Test Approach to Reasoning

1:01:33The End-to-End Predictive Model Argument and the Cost of Verification

The End-to-End Predictive Model Argument and the Cost of Verification

1:04:39Fine-Tuning, Chain of Thought, and the Amortization Argument

Fine-Tuning, Chain of Thought, and the Amortization Argument

1:07:27The "Advice Taking" Problem and the Limits of Chain of Thought

The "Advice Taking" Problem and the Limits of Chain of Thought

1:09:53Computational Complexity and the Misinterpretation of LLM Capabilities

Computational Complexity and the Misinterpretation of LLM Capabilities

1:12:21LLMs and Turing Completeness: An Orthogonal Question

LLMs and Turing Completeness: An Orthogonal Question

1:14:15The "Later Models Will Do Everything" Argument and the Architecture Lottery

The "Later Models Will Do Everything" Argument and the Architecture Lottery

1:18:52LLMs as Tools: The LLM Modulo Framework and the Need for External Verifiers

LLMs as Tools: The LLM Modulo Framework and the Need for External Verifiers

1:23:53The Future of AI: Generalist vs. Specialized Models and Agentic Systems

The Future of AI: Generalist vs. Specialized Models and Agentic Systems

1:29:53LLM Modulo and the Future of AI: Bridging the Gap Between Generalists and Specialists

LLM Modulo and the Future of AI: Bridging the Gap Between Generalists and Specialists

1:34:55Advice for Young Researchers: Embrace Logic and Skepticism

Advice for Young Researchers: Embrace Logic and Skepticism