The Inside View

Victoria Krakovna is a Research Scientist at DeepMind working on AGI safety and a co-founder of the Future of Life Institute, a non-profit organization working to mitigate technological risks to humanity and increase the chances of a positive future. In this interview we discuss three of her recent LW posts, namely DeepMind Alignment Team Opinions On AGI Ruin Arguments, Refining The Sharp Left Turn Threat Model and Paradigms of AI Alignment.

Transcript: theinsideview.ai/victoria

Youtube: https://youtu.be/ZpwSNiLV-nw

OUTLINE

(00:00) Intro

(00:48) DeepMind Alignment Team Opinions On AGI Ruin Arguments

(05:13) On The Possibility Of Iterating On Dangerous Domains and Pivotal acts

(14:14) Alignment and Interpretability

(18:14) Deciding Not To Build AGI And Stricted Publication norms

(27:18) Specification Gaming And Goal Misgeneralization

(33:02) Alignment Optimism And Probability Of Dying Before 2100 From unaligned AI

(37:52) Refining The Sharp Left Turn Threat Model

(48:15) A 'Move 37' Might Disempower Humanity

(59:59) Finding An Aligned Model Before A Sharp Left Turn

(01:13:33) Detecting Situational Awarareness

(01:19:40) How This Could Fail, Deception After One SGD Step

(01:25:09) Paradigms of AI Alignment

(01:38:04) Language Models Simulating Agency And Goals

(01:45:40) Twitter Questions

(01:48:30) Last Message For The ML Community

David Krueger is an assistant professor at the University of Cambridge and got his PhD from Mila. His research group focuses on aligning deep learning systems, but he is also interested in governance and global coordination. He is famous in Cambridge for not having an AI alignment research agenda per se, and instead he tries to enable his seven PhD students to drive their own research. In this episode we discuss AI Takeoff scenarios, research going on at David's lab, Coordination, Governance, Causality, the public perception of AI Alignment research and how to change it.

Youtube: https://youtu.be/bDMqo7BpNbk

Transcript: https://theinsideview.ai/david

OUTLINE

(00:00) Highlights

(01:06) Incentivized Behaviors and Takeoff Speeds

(17:53) Building Models That Understand Causality

(31:04) Agency, Acausal Trade And Causality in LLMs

(40:44) Recursive Self Improvement, Bitter Lesson And Alignment

(01:03:17) AI Governance And Coordination

(01:13:26) David’s AI Alignment Research Lab and the Existential Safety Community

(01:24:13) On The Public Perception of AI Alignment

(01:35:58) How To Get People In Academia To Work on Alignment

(02:00:19) Decomposing Learning Curves, Latest Research From David Krueger’s Lab

(02:20:06) Safety-Performance Trade-Offs

(02:30:20) Defining And Characterizing Reward Hacking

(02:40:51) Playing Poker With Ethan Caballero, Timelines

Ethan Caballero is a PhD student at Mila interested in how to best scale Deep Learning models according to all downstream evaluations that matter. He is known as the fearless leader of the "Scale Is All You Need" movement and the edgiest person at MILA. His first interview is the second most popular interview on the channel and today he's back to talk about Broken Neural Scaling Laws and how to use them to superforecast AGI.

Youtube: https://youtu.be/SV87S38M1J4

Transcript: https://theinsideview.ai/ethan2

OUTLINE

(00:00) The Albert Einstein Of Scaling

(00:50) The Fearless Leader Of The Scale Is All You Need Movement

(01:07) A Functional Form Predicting Every Scaling Behavior

(01:40) A Break Between Two Straight Lines On A Log Log Plot

(02:32) The Broken Neural Scaling Laws Equation

(04:04) Extrapolating A Ton Of Large Scale Vision And Language Tasks

(04:49) Upstream And Downstream Have Different Breaks

(05:22) Extrapolating Four Digit Addition Performance

(06:11) On The Feasability Of Running Enough Training Runs

(06:31) Predicting Sharp Left Turns

(07:51) Modeling Double Descent

(08:41) Forecasting Interpretability And Controllability

(09:33) How Deception Might Happen In Practice

(10:24) Sinister Stumbles And Treacherous Turns

(11:18) Recursive Self Improvement Precedes Sinister Stumbles

(11:51) Humans In The Loop For The Very First Deception

(12:32) The Hardware Stuff Is Going To Come After The Software Stuff

(12:57) Distributing Your Training By Copy-Pasting Yourself Into Different Servers

(13:42) Automating The Entire Hardware Pipeline

(14:47) Having Text AGI Spit Out New Robotics Design

(16:33) The Case For Existential Risk From AI

(18:32) Git Re-basin

(18:54) Is Chain-Of-Thoughts Enough For Complex Reasoning In LMs?

(19:52) Why Diffusion Models Outperform Other Generative Models

(21:13) Using Whisper To Train GPT4

(22:33) Text To Video Was Only Slightly Impressive

(23:29) Last Message

Irina Rish a professor at the Université de Montréal, a core member of Mila (Quebec AI Institute), and the organizer of the neural scaling laws workshop towards maximally beneficial AGI.

In this episode we discuss Irina's definition of Artificial General Intelligence, her takes on AI Alignment, AI Progress, current research in scaling laws, the neural scaling laws workshop she has been organizing, phase transitions, continual learning, existential risk from AI and what is currently happening in AI Alignment at Mila.

Transcript: theinsideview.ai/irina

Youtube: https://youtu.be/ZwvJn4x714s

OUTLINE

(00:00) Highlights

(00:30) Introduction

(01:03) Defining AGI

(03:55) AGI means augmented human intelligence

(06:20) Solving alignment via AI parenting

(09:03) From the early days of deep learning to general agents

(13:27) How Irina updated from Gato

(17:36) Building truly general AI within Irina's lifetime

(19:38) The least impressive thing that won't happen in five years

(22:36) Scaling beyond power laws

(28:45) The neural scaling laws workshop

(35:07) Why Irina does not want to slow down AI progress

(53:52) Phase transitions and grokking

(01:02:26) Does scale solve continual learning?

(01:11:10) Irina's probability of existential risk from AGI

(01:14:53) Alignment work at Mila

(01:20:08) Where will Mila get its compute from?

(01:27:04) With Great Compute Comes Great Responsibility

(01:28:51) The Neural Scaling Laws Workshop At NeurIPS

Shahar is a senior researcher at the Center for the Study of Existential Risk in Cambridge. In his past life, he was a Google Engineer, though right now he spends most of your time thinking about how to prevent the risks that occur if companies like Google end up deploying powerful AI systems, by organizing AI Governance role-playing workshops.

In this episode, we talk about a broad variety of topics, including how we could apply the lessons from running AI Governance workshops to governing transformative AI, AI Strategy, AI Governance, Trustworthy AI Development and end up answering some Twitter questions.

Youtube: https://youtu.be/3T7Gpwhtc6Q

Transcript: https://theinsideview.ai/shahar

Host: https://twitter.com/MichaelTrazzi

Shahar: https://www.shaharavin.com

Outline

(00:00) Highlights

(01:20) Intelligence Rising

(06:07) Measuring Transformative AI By The Scale Of Its Impact

(08:09) Comprehensive AI Services

(11:38) Automating CEOs Through AI Services

(14:21) Towards A "Tech Company Singularity"

(15:58) Predicting AI Is Like Predicting The Industrial Revolution

(19:57) 50% Chance Of Human-brain Performance By 2038

(22:25) AI Alignment Is About Steering Powerful Systems Towards Valuable Worlds

(23:51) You Should Still Worry About Less Agential Systems

(28:07) AI Strategy Needs To Be Tested In The Real World To Not Become Theoretical Physics

(31:37) Playing War Games For Real-time Partial-information Advesarial Thinking

(34:50) Towards World Leaders Playing The Game Because It’s Useful

(39:31) Open Game, Cybersecurity, Government Spending, Hard And Soft Power

(45:21) How Cybersecurity, Hard-power Or Soft-power Could Lead To A Strategic Advantage

(48:58) Cybersecurity In A World Of Advanced AI Systems

(52:50) Allocating AI Talent For Positive R&D ROI

(57:25) Players Learn To Cooperate And Defect

(01:00:10) Can You Actually Tax Tech Companies?

(01:02:10) The Emergence Of Bilateral Agreements And Technology Bans

(01:03:22) AI Labs Might Not Be Showing All Of Their Cards

(01:06:34) Why Publish AI Research

(01:09:21) Should You Expect Actors To Build Safety Features Before Crunch Time

(01:12:39) Why Tech Companies And Governments Will Be The Decisive Players

(01:14:29) Regulations Need To Happen Before The Explosion, Not After

(01:16:55) Early Regulation Could Become Locked In

(01:20:00) What Incentives Do Companies Have To Regulate?

(01:23:06) Why Shahar Is Terrified Of AI DAOs

(01:27:33) Concrete Mechanisms To Tell Apart Who We Should Trust With Building Advanced AI Systems

(01:31:19) Increasing Privacy To Build Trust

(01:33:37) Sensibilizing To Privacy Through Federated Learning

(01:35:23) How To Motivate AI Regulations

(01:37:44) How Governments Could Start Caring About AI risk

(01:39:12) Attempts To Regulate Autonomous Weapons Have Not Resulted In A Ban

(01:40:58) We Should Start By Convincing The Department Of Defense

(01:42:08) Medical Device Regulations Might Be A Good Model Audits

(01:46:09) Alignment Red Tape And Misalignment Fines

(01:46:53) Red Teaming AI systems

(01:49:12) Red Teaming May Not Extend To Advanced AI Systems

(01:51:26) What Climate change Teaches Us About AI Strategy

(01:55:16) Can We Actually Regulate Compute

(01:57:01) How Feasible Are Shutdown Swi

Katja runs AI Impacts, a research project trying to incrementally answer decision-relevant questions about the future of AI. She is well known for a survey published in 2017 called, When Will AI Exceed Human Performance? Evidence From AI Experts and recently published a new survey of AI Experts: What do ML researchers think about AI in 2022. We start this episode by discussing what Katja is currently thinking about, namely an answer to Scott Alexander on why slowing down AI Progress is an underexplored path to impact. 

Youtube: https://youtu.be/rSw3UVDZge0

Audio & Transcript: https://theinsideview.ai/katja

Host: https://twitter.com/MichaelTrazzi

Katja: https://twitter.com/katjagrace

OUTLINE

(00:00) Highlights

(00:58) Intro

(01:33) Why Advocating For Slowing Down AI Might Be Net Bad

(04:35) Why Slowing Down AI Is Taboo

(10:14) Why Katja Is Not Currently Giving A Talk To The UN

(12:40) To Avoid An Arms Race, Do Not Accelerate Capabilities

(16:27) How To Cooperate And Implement Safety Measures

(21:26) Would AI Researchers Actually Accept Slowing Down AI?

(29:08) Common Arguments Against Slowing Down And Their Counterarguments

(36:26) To Go To The Stars, Build AGI Or Upload Your Mind

(39:46) Why Katja Thinks There Is A 7% Chance Of AI Destroys The World

(46:39) Why We Might End Up Building Agents

(51:02) AI Impacts Answer Empirical Questions To Help Solve Important Ones

(56:32) The 2022 Expert Survey on AI Progress

(58:56) High Level Machine Intelligence

(1:04:02) Running A Survey That Actually Collects Data

(1:08:38) How AI Timelines Have Become Shorter Since 2016

(1:14:35) Are AI Researchers Still Too Optimistic?

(1:18:20) AI Experts Seem To Believe In Slower Takeoffs

(1:25:11) Automation and the Unequal Distributions of Cognitive power

(1:34:59) The Least Impressive Thing that Cannot Be Done in 2 years

(1:38:17) Final thoughts

Markus Anderljung is the Head of AI Policy at the Centre for Governance of AI  in Oxford and was previously seconded to the UK government office as a senior policy specialist. In this episode we discuss Jack Clark's AI Policy takes, answer questions about AI Policy from Twitter and explore what is happening in the AI Governance landscape more broadly.

Youtube: https://youtu.be/DD303irN3ps

Transcript: https://theinsideview.ai/markus

Host: https://twitter.com/MichaelTrazzi

Markus: https://twitter.com/manderljung

OUTLINE

(00:00) Highlights & Intro

(00:57) Jack Clark’s AI Policy Takes: Agree or Disagree

(06:57) AI Governance Takes: Answering Twitter Questions

(32:07) What The Centre For the Governance Of AI Is Doing

(57:38) The AI Governance Landscape

(01:15:07) How The EU Is Regulating AI

(01:29:28) Towards An Incentive Structure For Aligned AI

Alex Lawsen is an advisor at 80,000 hours, released an Introduction to Forecasting Youtube Series and has recently been thinking about forecasting AI progress, why you cannot just "update all the way bro" (discussed in my latest episode with Connor Leahy) and how to develop inside views about AI Alignment in general.

Youtube: https://youtu.be/vLkasevJP5c

Transcript: https://theinsideview.ai/alex 

Host: https://twitter.com/MichaelTrazzi

Alex: https://twitter.com/lxrjl

OUTLINE

(00:00) Intro

(00:31) How Alex Ended Up Making Forecasting Videos

(02:43) Why You Should Try Calibration Training

(07:25) How Alex Upskilled In Forecasting

(12:25) Why A Spider Monkey Profile Picture

(13:53) Why You Cannot Just "Update All The Way Bro"

(18:50) Why The Metaculus AGI Forecasts Dropped Twice

(24:37) How Alex’s AI Timelines Differ From Metaculus

(27:11) Maximizing Your Own Impact Using Forecasting

(33:52) What Makes A Good Forecasting Question

(41:59) What Motivated Alex To Develop Inside Views About AI

(43:26) Trying To Pass AI Alignment Ideological Turing Tests

(54:52) Why Economic Growth Curve Fitting Is Not Sufficient To Forecast AGI

(01:04:10) Additional Resources

Robert Long is a research fellow at the Future of Humanity Institute. His work is at the intersection of the philosophy of AI Safety and consciousness of AI. We talk about the recent LaMDA controversy, Ilya Sutskever's slightly conscious tweet, the metaphysics and philosophy of consciousness, artificial sentience, and how a future filled with digital minds could get really weird. 

Youtube: https://youtu.be/K34AwhoQhb8

Transcript: https://theinsideview.ai/roblong

Host: https://twitter.com/MichaelTrazzi

Robert: https://twitter.com/rgblong

Robert's blog: https://experiencemachines.substack.com

OUTLINE

(00:00:00) Intro

(00:01:11) The LaMDA Controversy

(00:07:06) Defining AGI And Consciousness

(00:10:30) The Slightly Conscious Tweet

(00:13:16) Could Large Language Models Become Conscious?

(00:18:03) Blake Lemoine Does Not Negotiate With Terrorists

(00:25:58) Could We Actually Test Artificial Consciousness?

(00:29:33) From Metaphysics To Illusionism

(00:35:30) How We Could Decide On The Moral Patienthood Of Language Models

(00:42:00) Predictive Processing, Global Workspace Theories and Integrated Information Theory

(00:49:46) Have You Tried DMT?

(00:51:13) Is Valence Just The Reward in Reinforcement Learning?

(00:54:26) Are Pain And Pleasure Symetrical?

(01:04:25) From Charismatic AI Systems to Artificial Sentience

(01:15:07) Sharing The World With Digital Minds

(01:24:33) Why AI Alignment Is More Pressing Than Artificial Sentience

(01:39:48) Why Moral Personhood Could Require Memory

(01:42:41) Last thoughts And Further Readings

Ethan Perez is a research scientist at Anthropic, working on large language models. He is the second Ethan working with large language models coming on the show but, in this episode, we discuss why alignment is actually what you need, not scale. We discuss three projects he has been pursuing before joining Anthropic, namely the Inverse Scaling Prize, Red Teaming Language Models with Language Models, and Training Language Models with Language Feedback. 

Ethan Perez: https://twitter.com/EthanJPerez

Transcript: https://theinsideview.ai/perez

Host: https://twitter.com/MichaelTrazzi

OUTLINE

(00:00:00) Highlights

(00:00:20) Introduction

(00:01:41) The Inverse Scaling Prize

(00:06:20) The Inverse Scaling Hypothesis

(00:11:00) How To Submit A Solution

(00:20:00) Catastrophic Outcomes And Misalignment

(00:22:00) Submission Requirements

(00:27:16) Inner Alignment Is Not Out Of Distribution Generalization

(00:33:40) Detecting Deception With Inverse Scaling

(00:37:17) Reinforcement Learning From Human Feedback

(00:45:37) Training Language Models With Language Feedback

(00:52:38) How It Differs From InstructGPT

(00:56:57) Providing Information-Dense Feedback

(01:03:25) Why Use Language Feedback

(01:10:34) Red Teaming Language Models With Language Models

(01:20:17) The Classifier And Advesarial Training

(01:23:53) An Example Of Red-Teaming Failure

(01:27:47) Red Teaming Using Prompt Engineering

(01:32:58) Reinforcement Learning Methods

(01:41:53) Distributional Biases

(01:45:23) Chain of Thought Prompting

(01:49:52) Unlikelihood Training and KL Penalty

(01:52:50) Learning AI Alignment through the Inverse Scaling Prize

(01:59:33) Final thoughts on AI Alignment