Published Sep 3, 2021

Evan Hubinger on Effective Altruism and AI Safety

AI researcher Evan Hubinger delves into the complexities of AI safety, highlighting the challenges of interpretability, alignment, and optimization biases in AI systems, while emphasizing the need for transparency and rigorous safety measures to mitigate existential risks.

Episode Highlights

Topics covered

Episode Highlights

Related Episodes

Connor Leahy on EleutherAI, Replicating GPT-2/GPT-3, AI Risk and Alignment
Answers 383 questions
Scott Aaronson: Against AI Doomerism
Answers 383 questions
Davidad Dalrymple: Towards Provably Safe AI
Answers 383 questions
Laura Weidinger: Ethical Risks, Harms, and Alignment of Large Language Models
Answers 383 questions
Eric Jang: AI is Good For You
Answers 383 questions
Jeremie Harris: Realistic Alignment and AI Policy
Answers 383 questions
Upol Ehsan on Human-Centered Explainable AI and Social Transparency
Answers 383 questions
Daniel Situnayake: AI on the Edge
Answers 383 questions
Seth Lazar: Normative Philosophy of Computing
Answers 383 questions
Ben Green: "Tech for Social Good" Needs to Do More
Answers 383 questions
Sara Hooker: Cohere For AI, the Hardware Lottery, and DL Tradeoffs
Answers 383 questions
Peter Henderson on RL Benchmarking, Climate Impacts of AI, and AI for Law
Answers 383 questions
Eric Jang on Robots Learning at Google and Generalization via Language
Answers 383 questions
Catherine Olsson and Nelson Elhage: Anthropic, Understanding Transformers
Answers 383 questions
Anant Agarwal: AI for Education
Answers 383 questions

Dexa/Daniel Bashir

Evan Hubinger on Effective Altruism and AI Safety

Topics covered

Popular Clips

Ensuring Meaningful Interpretability

Deceptive Optimization in Recommendation Algorithms

Generalization and Objectives

The Importance of AI Safety

Capability Generalization and Objective Generalization

Optimization Algorithms

Risks of Learned Optimization

Overcoming Bias

Detecting Mesa Optimizers

Neural Network Inductive Biases

Non-Myopic Models

Evolutionary Proxies

Deception in Gradient Descent

Relaxed Adversarial Training

Deception vs Corrigible

Episode Highlights

Interpreting AI Systems

AI Safety Importance

Optimization and Biases

Related Episodes

Connor Leahy on EleutherAI, Replicating GPT-2/GPT-3, AI Risk and Alignment

Scott Aaronson: Against AI Doomerism

Davidad Dalrymple: Towards Provably Safe AI

Laura Weidinger: Ethical Risks, Harms, and Alignment of Large Language Models

Eric Jang: AI is Good For You

Jeremie Harris: Realistic Alignment and AI Policy

Upol Ehsan on Human-Centered Explainable AI and Social Transparency

Daniel Situnayake: AI on the Edge

Seth Lazar: Normative Philosophy of Computing

Ben Green: "Tech for Social Good" Needs to Do More

Sara Hooker: Cohere For AI, the Hardware Lottery, and DL Tradeoffs

Peter Henderson on RL Benchmarking, Climate Impacts of AI, and AI for Law

Eric Jang on Robots Learning at Google and Generalization via Language

Catherine Olsson and Nelson Elhage: Anthropic, Understanding Transformers

Anant Agarwal: AI for Education

Evan Hubinger on Effective Altruism and AI Safety

Topics covered

Popular Clips

Episode Highlights

Interpreting AI SystemsEvan Hubinger examines the limitations of AI interpretability tools and the importance of transparency in AI systems. He proposes strategies like relaxed adversarial training to ensure AI models are not only effective but also understandable and trustworthy.

Interpreting AI Systems

AI Safety Importance

Optimization and BiasesEvan Hubinger explores the intricacies of mesa optimizers and their implications for AI alignment. He discusses the challenges posed by deceptive alignments and the influence of inductive biases on AI systems.

Optimization and Biases

Related Episodes