The Gradient: Perspectives on AI

In episode 40 of The Gradient Podcast, Andrey Kurenkov speaks to Catherine Olsson and Nelson Elhage.

Catherine and Nelson are both members of technical staff at Anthropic, which is an AI safety and research company that’s working to build reliable, interpretable, and steerable AI systems. Catherine and Nelson’s focus is on interpretability, and we will discuss several of their recent works in this interview. 
Follow The Gradient on Twitter

Outline:

(00:00) Intro
(01:10) Catherine’s Path into AI
(03:25) Nelson’s Path into AI
(05:23) Overview of Anthropic
(08:21) Mechanistic Interpretability
(15:15) Transformer Circuits 
(21:30) Toy Transformer
(27:25) Induction Heads
(31:00) In-Context Learning
(35:10) Evidence for Induction Heads Enabling In-Context Learning
(39:30) What’s Next
(43:10) Replicating Results
(46:00) Outro

Links:

Anthropic

Zoom In: An Introduction to Circuits

Mechanistic Interpretability, Variables, and the Importance of Interpretable Bases

A Mathematical Framework for Transformer Circuits

In-context Learning and Induction Heads 

PySvelte

In episode 38 of The Gradient Podcast, Daniel Bashir speaks to Been Kim.

Been is a staff research scientist at Google Brain focused on interpretability–helping humans communicate with complex machine learning models by not only building tools but also studying how humans interact with these systems. She has served with a number of conferences including ICLR, NeurIPS, ICML, and AISTATS. She gave the keynotes at ICLR 2022, ECML 2020, and the G20 meeting in Argentina in 2018. Her work TCAV received the UNESCO Netexplo award, was featured at Google I/O 2019 and in Brian Christian’s book The Alignment Problem.

Subscribe to The Gradient Podcast:  Apple Podcasts  | Spotify | Pocket Casts | RSSFollow The Gradient on Twitter

Outline:

(00:00) Intro(02:20) Path to AI/interpretability(06:10) The Progression of Been’s thinking / PhD thesis(11:30) Towards a Rigorous Science of Interpretable Machine Learning(24:52) Interpretability and Software Testing(27:00) Been’s ICLR Keynote and Human-Machine “Language”(37:30) TCAV(43:30) Mood Board Search and CAV Camera(48:00) TCAV’s Limitations and Follow-up Work(56:00) Acquisition of Chess Knowledge in AlphaZero(1:07:00) Daniel spends a very long time asking “what does it mean to you to be a researcher?”(1:09:00) The everyday drudgery, more lessons from Been(1:11:32) Outro

Links:

* Been’s website

* CAVcamera app

In episode 37 of The Gradient Podcast, Andrey Kurenkov speaks to Laura Weidinger

Laura is a senior research scientist at DeepMind, with her focus being AI ethics. Laura is also a PhD candidate at the University of Cambridge, studying philosophy of science and specifically approaches to measuring the ethics of AI systems. Previously Laura worked in technology policy at UK and EU levels, as a Policy executive at techUK. She then pivoted to cognitive science research and studied human learning at the Max Planck Institute for Human Development in Berlin, and was a Guest Lecturer at the Ada National College for Digital Skills. She received her Master's degree at the Humboldt University of Berlin, from the School of Mind and Brain, with her focus being Neuroscience/ Philosophy/ Cognitive science.

Subscribe to The Gradient Podcast:  Apple Podcasts  | Spotify | Pocket Casts | RSSFollow The Gradient on Twitter

Outline:

(00:00) Intro(01:20) Path to AI(04:25) Research in Cognitive Science(06:40) Interest in AI Ethics(14:30) Ethics Considerations for Researchers(17:38) Ethical and social risks of harm from language models (25:30) Taxonomy of Risks posed by Language Models(27:33) Characteristics of Harmful Text: Towards Rigorous Benchmarking of Language Models(33:25) Main Insight for Measuring Harm(35:40) The EU AI Act(39:10) Alignment of language agents(46:10) GPT-4Chan(53:40) Interests outside of AI(55:30) Outro

Links:

Ethical and social risks of harm from language models 

Taxonomy of Risks posed by Language Models

Characteristics of Harmful Text: Towards Rigorous Benchmarking of Language Models

Alignment of language agents

In episode 36 of The Gradient Podcast, Daniel Bashir speaks to Sebastian Raschka.

Sebastian is an Assistant Professor of Statistics at the University of Wisconsin-Madison and Lead AI Educator at Lightning AI. He has written two bestselling books: Python Machine Learning and Machine Learning with PyTorch and Scikit-Learn.

Subscribe to The Gradient Podcast:  Apple Podcasts  | Spotify | Pocket Casts | RSSFollow The Gradient on Twitter

Sections:

(00:00) Intro

(01:10) Sebastian’s intro to AI

(05:15) Sebastian’s process for learning new things

(12:15) Learning style varies with purpose

(16:10) Ordinal Regression

(31:00) Solving rank inconsistency with conditional probability

(35:00) Semi-Adversarial Networks

(44:15) Why Sebastian got into education

(52:45) Lightning AI

(1:00:00) Sebastian’s advice for educators

(1:03:30) Be cool like Sebastian and follow the Gradient

(1:03:40) Outro

Episode Links:

* Sebastian’s Homepage

* Sebastian’s Twitter

* Sebastian’s Books

In episode 35 of The Gradient Podcast, guest host Sharon Zhou speaks to Jack Shanahan.

John (Jack) Shanahan was a Lieutenant General in the United States Air Force, retired after a 36-year military career. He was the inaugural Director of the Joint Artificial Intelligence Center (JAIC) in the U.S. Department of Defense (DoD). He was also the Director of the Algorithmic Warfare Cross-Functional Team (Project Maven). Currently, he is a Special Government Employee supporting the National Security Commission on Artificial Intelligence; serves on the Board of Advisors for the Common Mission Project; is an advisor to The Changing Character of War Centre (Oxford University); is a member of the CACI Strategic Advisory Group; and serves as an Advisor to the Military Cyber Professionals Association.

Subscribe to The Gradient Podcast:  Apple Podcasts  | Spotify | Pocket Casts | RSSFollow The Gradient on Twitter

Outline:

(00:00) Intro(01:20) Introduction to Jack and Sharon(07:30) Project Maven(09:45) Relationship of Tech Sector and DoD(16:40) Need for AI in DoD(20:10) Bringing the tech-DoD divide(30:00) Conclusion

Episode Links:

* John N.T. Shanahan Wikipedia

* AI To Revolutionize U.S. Intelligence Community With General Shanahan

* Email: [email protected]

In episode 34 of The Gradient Podcast, Daniel Bashir speaks to Sara Hooker.

Sara (@sarahookr) leads Cohere for AI and is a former Research Scientist at Google. Sara founded a Bay Area non-profit called Delta Analytics, which works with non-profits and communities to build technical capacity. She is also one of the co-founders of the Trustworthy ML Initiative, an active participant of the ML Collective research group, and a host of the underrated ML podcast.

Subscribe to The Gradient Podcast:  Apple Podcasts  | Spotify | Pocket Casts | RSSFollow The Gradient on Twitter

Sections:

(00:00) Intro(02:20) Podcasting gripe-fest(06:00) Sara’s journey: from economics to AI(09:15) Economics vs. AI research(12:45) The Hardware Lottery(19:15) Towards better hardware benchmarks(26:00) Getting away from the hardware lottery(32:30) The myth of compact, interpretable, robust, performant DNNs(35:15) Top-line metrics vs. disaggregated metrics(39:20) Solving memorization in the data pipeline, noisy examples(45:35) Cohere For AI

Episode Links:

* Cohere for AI

* Sara’s Homepage

In episode 33 of The Gradient Podcast, Andrey Kurenkov speaks to Lukas Biewald.

Lukas Biewald is a co-founder of Weights and Biases, a company that creates developer tools for machine learning. Prior to that he was a co-founder and CEO of Figure Eight Inc. (formerly CrowdFlower) — an Internet company that collects training data for machine learning, which was sold for 300 million dollars.

Subscribe to The Gradient Podcast:  Apple Podcasts  | Spotify | Pocket Casts | RSSFollow The Gradient on Twitter

Outline:

* (00:00) Intro

* (01:18) Start in AI

* (06:17) CrowdFlower / Crowdsourcing

* (21:06) Discovering Deep Learning 

* (25:10) Learning Deep Learning

* (32:50) Weights and Biases

* (37:30) State of Tooling for ML 

* (41:20) Exciting AI Trends

* (44:42) Interests outside of AI

* (45:40) Outro

Links:

* Lukas’s website

* Lukas’s GitHub

* Starting a Second Machine Learning Tools Company, Ten Years Later

* Confession of a so-called AI expert

* What I learned from looking at 200 machine learning tools

* CS 329S: Machine Learning Systems Design

* Designing Machine Learning Systems

Opportunity at Weights & Biases:

In episode 32 of The Gradient Podcast, Andrey Kurenkov speaks to Chip Huyen.

Chip Huyen is a co-founder of Claypot AI, a platform for real-time machine learning. Previously, she was with Snorkel AI and NVIDIA. She teaches CS 329S: Machine Learning Systems Design at Stanford. She has also written four bestselling Vietnamese books, and more recently her new O’Reilly book Designing Machine Learning Systems has just come out! 

Subscribe to The Gradient Podcast:  Apple Podcasts  | Spotify | Pocket Casts | RSSFollow The Gradient on Twitter

She also maintains a Discord server with a focus on Machine Learning Systems.

Outline:

* (00:00) Intro

* (01:30) 3-year trip through Asia, Africa, and South America

* (04:00) Getting into AI at Stanford

* (11:30) Confession of a so-called AI expert

* (16:40) Academia vs Industry

* (17:40) Focus on ML Systems

* (20:00) ML in Academia vs Industry

* (28:15) Maturity of AI in Industry

* (31:45) ML Tools

* (37:20) Real Time ML

* (43:00) ML Systems Class and Book

Links:

* Chip’s website

* MLOps Discord server

* Confession of a so-called AI expert

* What I learned from looking at 200 machine learning tools

* CS 329S: Machine Learning Systems Design

* Designing Machine Learning Systems

In episode 31 of The Gradient Podcast, Daniel Bashir speaks to Preetum Nakkiran.

Preetum is a Research Scientist at Apple, a Visiting Researcher at UCSD, and part of the NSF/Simons Collaboration on the Theoretical Foundations of Deep Learning. He completed his PhD at Harvard, where he co-founded the ML Foundations Group. Preetum’s research focuses on building conceptual tools for understanding learning systems.

Subscribe to The Gradient Podcast:  Apple Podcasts  | Spotify | Pocket Casts | RSSFollow The Gradient on Twitter

Sections:

(00:00) Intro

(01:25) Getting into AI through Theoretical Computer Science (TCS)

(09:08) Lack of Motivation in TCS and Learning What Research Is

(12:12) Foundational vs Problem-Solving Research, Antipatterns in TCS

(16:30) Theory and Empirics in Deep Learning

(18:30) What is an Empirical Theory of Deep Learning

(28:21) Deep Double Descent

(40:00) Inductive Biases in SGD, epoch-wise double descent

(45:25) Inductive Biases Stick Around

(47:12) Deep Bootstrap

(59:40) Distributional Generalization - Paper Rejections

(1:02:30) Classical Generalization and Distributional Generalization

(1:16:46) Future Work: Studying Structure in Data

(1:20:51) The Tweets^TM

(1:37:00) Outro

Episode Links:

* Preetum’s Homepage

* Preetum’s PhD Thesis

In episode 30 of The Gradient Podcast, Daniel Bashir speaks to Max Woolf.

Max Woolf (@minimaxir) is currently a Data Scientist at BuzzFeed in San Francisco. Some work he’s done for BuzzFeed includes using StyleGAN to create AI-generated fake boyfriends and AI-generated art quizzes. In his free time, Max creates open source Python and R software on his GitHub. More recently, Max has been developing tooling for AI content generation, such as aitextgen for easy AI text generation.

Max’s projects are funded by his Patreon. If you have found anything on his website helpful, please help contribute!

Subscribe to The Gradient Podcast:  Apple Podcasts  | Spotify | Pocket Casts | RSSFollow The Gradient on Twitter

Sections:

(00:00) Intro

(01:20) Max’s Intro to Data Science and AI

(07:00) Software Engineering in Data Science, Max’s Perspectives

(09:00) Max’s Work at BuzzFeed

(23:10) Scaling, Inference, Large Models

(27:00) AI Content Generation

(30:45) Discourse About GPT-3

(34:30) AI Inventors

(38:35) Fun Projects and One-Offs: AI-generated Pokémon

(43:35) GPT-3-generated Discussion Topics

(46:30) Advice for Data Scientists

(48:10) BuzzFeed is Hiring :)

(48:20) Outro

Episode Links:

* Max’s Homepage

* Real-World Data Science