The Gradient: Perspectives on AI

Episode 124

You may think you’re doing a priori reasoning, but actually you’re just over-generalizing from your current experience of technology.

I spoke with Professor Seth Lazar about:

* Why managing near-term and long-term risks isn’t always zero-sum

* How to think through axioms and systems in political philosphy

* Coordination problems, economic incentives, and other difficulties in developing publicly beneficial AI

Seth is Professor of Philosophy at the Australian National University, an Australian Research Council (ARC) Future Fellow, and a Distinguished Research Fellow of the University of Oxford Institute for Ethics in AI. He has worked on the ethics of war, self-defense, and risk, and now leads the Machine Intelligence and Normative Theory (MINT) Lab, where he directs research projects on the moral and political philosophy of AI.

Reach me at [email protected] for feedback, ideas, guest suggestions.

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

Outline:

* (00:00) Intro

* (00:54) Ad read — MLOps conference

* (01:32) The allocation of attention — attention, moral skill, and algorithmic recommendation

* (03:53) Attention allocation as an independent good (or bad)

* (08:22) Axioms in political philosophy

* (11:55) Explaining judgments, multiplying entities, parsimony, intuitive disgust

* (15:05) AI safety / catastrophic risk concerns

* (22:10) Superintelligence arguments, reasoning about technology

* (28:42) Attacking current and future harms from AI systems — does one draw resources from the other?

* (35:55) GPT-2, model weights, related debates

* (39:11) Power and economics—coordination problems, company incentives

* (50:42) Morality tales, relationship between safety and capabilities

* (55:44) Feasibility horizons, prediction uncertainty, and doing moral philosophy

* (1:02:28) What is a feasibility horizon?

* (1:08:36) Safety guarantees, speed of improvements, the “Pause AI” letter

* (1:14:25) Sociotechnical lenses, narrowly technical solutions

* (1:19:47) Experiments for responsibly integrating AI systems into society

* (1:26:53) Helpful/honest/harmless and antagonistic AI systems

* (1:33:35) Managing incentives conducive to developing technology in the public interest

* (1:40:27) Interdisciplinary academic work, disciplinary purity, power in academia

* (1:46:54) How we can help legitimize and support interdisciplinary work

* (1:50:07) Outro

Links:

* Seth’s Linktree and Twitter

* Resources

* Attention, moral skill, and algorithmic recommendation

* Catastrophic AI Risk slides

Episode 123

I spoke with Suhail Doshi about:

* Why benchmarks aren’t prepared for tomorrow’s AI models

* How he thinks about artists in a world with advanced AI tools

* Building a unified computer vision model that can generate, edit, and understand pixels.

Suhail is a software engineer and entrepreneur known for founding Mixpanel, Mighty Computing, and Playground AI (they’re hiring!).

Reach me at [email protected] for feedback, ideas, guest suggestions.

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

Outline:

* (00:00) Intro

* (00:54) Ad read — MLOps conference

* (01:30) Suhail is *not* in pivot hell but he *is* all-in on 50% AI-generated music

* (03:45) AI and music, similarities to Playground

* (07:50) Skill vs. creative capacity in art

* (12:43) What we look for in music and art

* (15:30) Enabling creative expression

* (18:22) Building a unified computer vision model, underinvestment in computer vision

* (23:14) Enhancing the aesthetic quality of images: color and contrast, benchmarks vs user desires

* (29:05) “Benchmarks are not prepared for how powerful these models will become”

* (31:56) Personalized models and personalized benchmarks

* (36:39) Engaging users and benchmark development

* (39:27) What a foundation model for graphics requires

* (45:33) Text-to-image is insufficient

* (46:38) DALL-E 2 and Imagen comparisons, FID

* (49:40) Compositionality

* (50:37) Why Playground focuses on images vs. 3d, video, etc.

* (54:11) Open source and Playground’s strategy

* (57:18) When to stop open-sourcing?

* (1:03:38) Suhail’s thoughts on AGI discourse

* (1:07:56) Outro

Links:

* Playground homepage

* Suhail on Twitter

Episode 122

I spoke with Azeem Azhar about:

* The speed of progress in AI

* Historical context for some of the terminology we use and how we think about technology

* What we might want our future to look like

Azeem is an entrepreneur, investor, and adviser. He is the creator of Exponential View, a global platform for in-depth technology analysis, and the host of the Bloomberg Original series Exponentially.

Reach me at [email protected] for feedback, ideas, guest suggestions.

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

Outline:

* (00:00) Intro

* (00:32) Ad read — MLOps conference

* (01:05) Problematizing the term “exponential”

* (07:35) Moore’s Law as social contract, speed of technological growth and impedances

* (14:45) Academic incentives, interdisciplinary work, rational agents and historical context

* (21:24) Monolithic scaling

* (26:38) Investment in scaling

* (31:22) On Sam Altman

* (36:25) Uses of “AGI,” “intelligence”

* (41:32) Historical context for terminology

* (48:58) AI and teaching

* (53:51) On the technology-human divide

* (1:06:26) New technologies and the futures we want

* (1:10:50) Inevitability narratives

* (1:17:01) Rationality and objectivity

* (1:21:13) Cultural affordances and intellectual history

* (1:26:15) Centralized and decentralized AI systems

* (1:32:54) Instruction tuning and helpful/honest/harmless

* (1:39:18) Azeem’s future outlook

* (1:46:15) Outro

Links:

* Azeem’s website and Twitter

* Exponential View

Episode 122

I spoke with Professor David Thorstad about:

* The practical difficulties of doing interdisciplinary work

* Why theories of human rationality should account for boundedness, heuristics, and other cognitive limitations

* why EA epistemics suck (ok, it’s a little more nuanced than that)

Professor Thorstad is an Assistant Professor of Philosophy at Vanderbilt University, a Senior Research Affiliate at the Global Priorities Institute at Oxford, and a Research Affiliate at the MINT Lab at Australian National University. One strand of his research asks how cognitively limited agents should decide what to do and believe. A second strand asks how altruists should use limited funds to do good effectively.

Reach me at [email protected] for feedback, ideas, guest suggestions.

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

Outline:

* (00:00) Intro

* (01:15) David’s interest in rationality

* (02:45) David’s crisis of confidence, models abstracted from psychology

* (05:00) Blending formal models with studies of the mind

* (06:25) Interaction between academic communities

* (08:24) Recognition of and incentives for interdisciplinary work

* (09:40) Movement towards interdisciplinary work

* (12:10) The Standard Picture of rationality

* (14:11) Why the Standard Picture was attractive

* (16:30) Violations of and rebellion against the Standard Picture

* (19:32) Mistakes made by critics of the Standard Picture

* (22:35) Other competing programs vs Standard Picture

* (26:27) Characterizing Bounded Rationality

* (27:00) A worry: faculties criticizing themselves

* (29:28) Self-improving critique and longtermism

* (30:25) Central claims in bounded rationality and controversies

* (32:33) Heuristics and formal theorizing

* (35:02) Violations of Standard Picture, vindicatory epistemology

* (37:03) The Reason Responsive Consequentialist View (RRCV)

* (38:30) Objective and subjective pictures

* (41:35) Reason responsiveness

* (43:37) There are no epistemic norms for inquiry

* (44:00) Norms vs reasons

* (45:15) Arguments against epistemic nihilism for belief

* (47:30) Norms and self-delusion

* (49:55) Difficulty of holding beliefs for pragmatic reasons

* (50:50) The Gibbardian picture, inquiry as an action

* (52:15) Thinking how to act and thinking how to live — the power of inquiry

* (53:55) Overthinking and conducting inquiry

* (56:30) Is thinking how to inquire as an all-things-considered matter?

* (58:00) Arguments for the RRCV

* (1:00:40) Deciding on minimal criteria for the view, stereotyping

* (1:02:15) Eliminating stereotypes from the theory

* (1:04:20) Theory construction in epistemology and moral intuition

* (1:08:20) Refusing theories for moral reasons and disciplinary boundaries

* (1:10:30) The argument from minimal criteria, evaluating against competing views

* (1:13:45) Comparing to other theories

* (1:15:00) The explanatory argument

* (1:17:53) Parfit and Railton, norms of friendship vs utility

* (1:20:00) Should you call out your friend for being a womanizer

* (1:22:00) Vindicatory Epistemology

* (1:23:05) Panglossianism and meliorative epistemology

* (1:24:42) Heuristics and recognition-driven investigation

* (1:26:33) Rational inquiry leading to irrational beliefs — metacognitive processing

* (1:29:08) Stakes of inquiry and costs of metacognitive processing

* (1:30:00) When agents are incoherent, focuses on inquiry

* (1:32:05) Indirect normative assessment and its consequences

* (1:37:47) Against the Singularity Hypothesis

* (1:39:00) Superintelligence and the ontological argument

* (1:41:50) Hardware growth and general intelligence growth, AGI definitions

* (1:43:55) Difficulties in arguing for hyperbolic growth

* (1:46:07) Chalmers and the proportionality argument

* (1:47:53) Arguments for/against diminishing growth, research productivity, Moore’s Law

* (1:50:08) On progress studies

* (1:52:40) Improving research productivity and technology growth

* (1:54:00) Mistakes in the moral mathematics of existential risk, longtermist epistemics

* (1:55:30) Cumulative and per-unit risk

* (1:57:37) Back and forth with longtermists, time of perils

* (1:59:05) Background risk — risks we can and can’t intervene on, total existential risk

* (2:00:56) The case for longtermism is inflated

* (2:01:40) Epistemic humility and longtermism

* (2:03:15) Knowledge production — reliable sources, blog posts vs peer review

* (2:04:50) Compounding potential errors in knowledge

* (2:06:38) Group deliberation dynamics, academic consensus

* (2:08:30) The scope of longtermism

* (2:08:30) Money in effective altruism and processes of inquiry

* (2:10:15) Swamping longtermist options

* (2:12:00) Washing out arguments and justified belief

* (2:13:50) The difficulty of long-term forecasting and interventions

* (2:15:50) Theory of change in the bounded rationality program

* (2:18:45) Outro

Links:

* David’s homepage and Twitter and blog

* Papers mentioned/read

* Bounded rationality and inquiry

* Why bounded rationality (in epistemology)?

* Against the newer evidentialists

* The accuracy-coherence tradeoff in cognition

* There are no epistemic norms of inquiry

* Permissive metaepistemology

* Global priorities and effective altruism

* What David likes about EA

* Against the singularity hypothesis (+ blog posts)

* Three mistakes in the moral mathematics of existential risk (+ blog posts)

* The scope of longtermism

* Epistemics

Episode 121

I spoke with Professor Ryan Tibshirani about:

* Differences between the ML and statistics communities in scholarship, terminology, and other areas.

* Trend filtering

* Why you can’t just use garbage prediction functions when doing conformal prediction

Ryan is a Professor in the Department of Statistics at UC Berkeley. He is also a Principal Investigator in the Delphi group. From 2011-2022, he was a faculty member in Statistics and Machine Learning at Carnegie Mellon University. From 2007-2011, he did his Ph.D. in Statistics at Stanford University.

Reach me at [email protected] for feedback, ideas, guest suggestions.

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

Outline:

* (00:00) Intro

* (01:10) Ryan’s background and path into statistics

* (07:00) Cultivating taste as a researcher

* (11:00) Conversations within the statistics community

* (18:30) Use of terms, disagreements over stability and definitions

* (23:05) Nonparametric Regression

* (23:55) Background on trend filtering

* (33:48) Analysis and synthesis frameworks in problem formulation

* (39:45) Neural networks as a specific take on synthesis

* (40:55) Divided differences, falling factorials, and discrete splines

* (41:55) Motivations and background

* (48:07) Divided differences vs. derivatives, approximation and efficiency

* (51:40) Conformal prediction

* (52:40) Motivations

* (1:10:20) Probabilistic guarantees in conformal prediction, choice of predictors

* (1:14:25) Assumptions: i.i.d. and exchangeability — conformal prediction beyond exchangeability

* (1:25:00) Next directions

* (1:28:12) Epidemic forecasting — COVID-19 impact and trends survey

* (1:29:10) Survey methodology

* (1:38:20) Data defect correlation and its limitations for characterizing datasets

* (1:46:14) Outro

Links:

* Ryan’s homepage

* Works read/mentioned

* Nonparametric Regression

* Adaptive Piecewise Polynomial Estimation via Trend Filtering (2014) 

* Divided Differences, Falling Factorials, and Discrete Splines: Another Look at Trend Filtering and Related Problems (2020)

* Distribution-free Inference

* Distribution-Free Predictive Inference for Regression (2017)

* Conformal Prediction Under Covariate Shift (2019)

* Conformal Prediction Beyond Exchangeability (2023)

* Delphi and COVID-19 research

* Flexible Modeling of Epidemics

* Real-Time Estimation of COVID-19 Infections

* The US COVID-19 Trends and Impact Survey and Big data, big problems: Responding to “Are we there yet?”

In episode 120 of The Gradient Podcast, Daniel Bashir speaks to Sasha Luccioni.

Sasha is the AI and Climate Lead at HuggingFace, where she spearheads research, consulting, and capacity-building to elevate the sustainability of AI systems. A founding member of Climate Change AI (CCAI) and a board member of Women in Machine Learning (WiML), Sasha is passionate about catalyzing impactful change, organizing events and serving as a mentor to under-represented minorities within the AI community.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach Daniel at [email protected]

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

Outline:

* (00:00) Intro

* (00:43) Sasha’s background

* (01:52) How Sasha became interested in sociotechnical work

* (03:08) Larger models and theory of change for AI/climate work

* (07:18) Quantifying emissions for ML systems

* (09:40) Aggregate inference vs training costs

* (10:22) Hardware and data center locations

* (15:10) More efficient hardware vs. bigger models — Jevons paradox

* (17:55) Uninformative experiments, takeaways for individual scientists, knowledge sharing, failure reports

* (27:10) Power Hungry Processing: systematic comparisons of ongoing inference costs

* (28:22) General vs. task-specific models

* (31:20) Architectures and efficiency

* (33:45) Sequence-to-sequence architectures vs. decoder-only

* (36:35) Hardware efficiency/utilization

* (37:52) Estimating the carbon footprint of Bloom and lifecycle assessment

* (40:50) Stable Bias

* (46:45) Understanding model biases and representations

* (52:07) Future work

* (53:45) Metaethical perspectives on benchmarking for AI ethics

* (54:30) “Moral benchmarks”

* (56:50) Reflecting on “ethicality” of systems

* (59:00) Transparency and ethics

* (1:00:05) Advice for picking research directions

* (1:02:58) Outro

Links:

* Sasha’s homepage and Twitter

* Papers read/discussed

* Climate Change / Carbon Emissions of AI Models

* Quantifying the Carbon Emissions of Machine Learning

* Power Hungry Processing: Watts Driving the Cost of AI Deployment?

* Tackling Climate Change with Machine Learning

* CodeCarbon

* Responsible AI

* Stable Bias: Analyzing Societal Representations in Diffusion Models

* Metaethical Perspectives on ‘Benchmarking’ AI Ethics

* Measuring Data

* Mind your Language (Model): Fact-Checking LLMs and their Role in NLP Research and Practice

In episode 119 of The Gradient Podcast, Daniel Bashir speaks to Professor Michael Sipser.

Professor Sipser is the Donner Professor of Mathematics and member of the Computer Science and Artificial Intelligence Laboratory at MIT.

He received his PhD from UC Berkeley in 1980 and joined the MIT faculty that same year. He was Chairman of Applied Mathematics from 1998 to 2000 and served as Head of the Mathematics Department 2004-2014. He served as interim Dean of Science 2013-2014 and then as Dean of Science 2014-2020.

He was a research staff member at IBM Research in 1980, spent the 1985-86 academic year on the faculty of the EECS department at Berkeley and at MSRI, and was a Lady Davis Fellow at Hebrew University in 1988. His research areas are in algorithms and complexity theory, specifically efficient error correcting codes, interactive proof systems, randomness, quantum computation, and establishing the inherent computational difficulty of problems. He is the author of the widely used textbook, Introduction to the Theory of Computation (Third Edition, Cengage, 2012).

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach Daniel at [email protected]

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

Outline:

* (00:00) Intro

* (01:40) Professor Sipser’s background

* (04:35) On interesting questions

* (09:00) Different kinds of research problems

* (13:00) What makes certain problems difficult

* (18:48) Nature of the P vs NP problem

* (24:42) Identifying interesting problems

* (28:50) Lower bounds on the size of sweeping automata

* (29:50) Why sweeping automata + headway to P vs. NP

* (36:40) Insights from sweeping automata, infinite analogues to finite automata problems

* (40:45) Parity circuits

* (43:20) Probabilistic restriction method

* (47:20) Relativization and the polynomial time hierarchy

* (55:10) P vs. NP

* (57:23) The non-connection between GO’s polynomial space hardness and AlphaGo

* (1:00:40) On handicapping Turing Machines vs. oracle strategies

* (1:04:25) The Natural Proofs Barrier and approaches to P vs. NP

* (1:11:05) Debates on methods for P vs. NP

* (1:15:04) On the possibility of solving P vs. NP

* (1:18:20) On academia and its role

* (1:27:51) Outro

Links:

* Professor Sipser’s homepage

* Papers discussed/read

* Halting space-bounded computations (1978)

* Lower bounds on the size of sweeping automata (1979)

* GO is Polynomial-Space Hard (1980)

* A complexity theoretic approach to randomness (1983)

* Parity, circuits, and the polynomial-time hierarchy (1984)

* A follow-up to Furst-Saxe-Sipser

* The Complexity of Finite Functions (1991)

In episode 118 of The Gradient Podcast, Daniel Bashir speaks to Andrew Lee.

Andrew is co-founder and CEO of Shortwave, a company dedicated to building a better product experience for email, particularly by leveraging AI. He previously co-founded and was CTO at Firebase.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach Daniel at [email protected]

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

Outline:

* (00:00) Intro

* (01:43) Andrew’s previous work, Firebase

* (04:48) Benefits of lacking experience in building Firebase

* (08:55) On “abstract reasoning” vs empirical capabilities

* (10:30) Shortwave’s AI system as a black box

* (11:55) Motivations for Shortwave

* (17:10) Why is Google not innovating on email?

* (21:53) Shortwave’s overarching product vision and pivots

* (27:40) Shortwave AI features

* (33:20) AI features for email and security concerns

* (35:45) Shortwave’s AI Email Assistant + architecture

* (43:40) Issues with chaining LLM calls together

* (45:25) Understanding implicit context in utterances, modularization without loss of context

* (48:56) Performance for AI assistant, batching and pipelining

* (55:10) Prompt length

* (57:00) On shipping fast

* (1:00:15) AI improvements that Andrew is following

* (1:03:10) Outro

Links:

* Andrew’s blog and Twitter

* Shortwave

* Introducing Ghostwriter

* Everything we shipped for AI Launch Week

* A deep dive into the world’s smartest email AI

Episode 117

“You get more of what you engage with. Everyone who complains about coverage should understand that every click, every quote tweet, every argument is registered by these publications as engagement. If what you want is really meaty, dispassionate, balanced, and fair explainers, you need to click on that, you need to read the whole thing, you need to share it, talk about it, comment on it. We get the media that we deserve.”

I spoke with Joss Fong.

Joss is a producer focused on science and technology, and was a founding member of the Vox video team. Her work has been recognized by the AAAS Kavli Science Journalism Awards, the Online Journalism Awards, and the News & Documentary Emmys. She holds a master's degree in science, health, and environmental reporting from NYU.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:32) Joss’s path into videomaking, J-school

* (07:45) Consumption and creation in explainer journalism

* (10:45) Finding clarity in information

* (13:15) Communication of ML research

* (15:55) Video journalism and science communication as separate and overlapping disciplines

* (19:41) Evolution of videos and videomaking

* (26:33) Explaining AI and communicating mental models

* (30:47) Meeting viewers in the middle, competing for attention

* (34:07) Explanatory techniques in Glad You Asked

* (37:10) Storytelling and communicating scientific information

* (40:57) “Is Beauty Culture Hurting Us?” and participating in video narratives

* (46:37) AI beauty filters

* (52:59) Obvious bias in generative AI

* (59:31) Definitions and ideas of progress, humanities and technology

* (1:05:08) “Iterative development” and outsourcing quality control to the public

* (1:07:10) Disagreement about (tech) journalism’s purpose

* (1:08:51) Incentives in newsrooms and journalistic organizations

* (1:12:04) AI for video generation and implications, limits of creativity

* (1:17:20) Skill and creativity

* (1:22:35) Joss’s new YouTube channel!

* (1:23:29) Outro

Links:

* Joss’s website and playlist of selected work

* AI-focused videos

* AI Art, Explained (2022)

* AI can do your homework. Now what? (2023)

* Computers just got a lot better at writing (2020)

* Facebook showed this ad to 95% women. Is that a problem? (2020)

* What facial recognition steals from us (2019)

* The big debate about the future of work (2017)

* AI and Creativity short film for Runway’s AIFF (2023)

* Others

* Is Beauty Culture Hurting Us? from Glad You Asked (2020)

* Joss’s Scientific American videos :)

In episode 116 of The Gradient Podcast, Daniel Bashir speaks to Kate Park.

Kate is the Director of Product at Scale AI. Prior to joining Scale, Kate worked on Tesla Autopilot as the AI team’s first and lead product manager building the industry’s first data engine. She has also published research on spoken natural language processing and a travel memoir.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:11) Kate’s background

* (03:22) Tesla and cameras vs. Lidar, importance of data

* (05:12) “Data is key”

* (07:35) Data vs. architectural improvements

* (09:36) Effort for data scaling

* (10:55) Transfer of capabilities in self-driving

* (13:44) Data flywheels and edge cases, deployment

* (15:48) Transition to Scale

* (18:52) Perspectives on shifting to transformers and data

* (21:00) Data engines for NLP vs. for vision

* (25:32) Model evaluation for LLMs in data engines

* (27:15) InstructGPT and data for RLHF

* (29:15) Benchmark tasks for assessing potential labelers

* (32:07) Biggest challenges for data engines

* (33:40) Expert AI trainers

* (36:22) Future work in data engines

* (38:25) Need for human labeling when bootstrapping new domains or tasks

* (41:05) Outro

Links:

* Scale Data Engine

* OpenAI case study

In episode 115 of The Gradient Podcast, Daniel Bashir speaks to Ben Wellington.

Ben is the Deputy Head of Feature Forecasting at Two Sigma, a financial sciences company. Ben has been at Two Sigma for more than 15 years, and currently leads efforts focused on natural language processing and feature forecasting. He is also the author of data science blog I Quant NY, which has influenced local government policy, including changes in NYC street infrastructure and the design of NYC subway vending machines. Ben is a Visiting Assistant Professor in the Urban and Community Planning program at the Pratt Institute in Brooklyn where he teaches statistics using urban open data. He holds a Ph.D. in Computer Science from New York University.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:30) Ben’s background

* (04:30) Why Ben was interested in NLP

* (05:48) Ben’s work on translational equivalence, dominant techniques

* (10:14) Scaling, large datasets at Two Sigma

* (12:50) Applying ML techniques to quantitative finance, features in financial ML systems

* (17:27) Baselines and time-dependence in constructing features, human knowledge

* (19:23) Black box models in finance

* (24:00) Two Sigma’s presence in the AI research community

* (26:55) Short- and long-term research initiatives at Two Sigma

* (30:42) How ML fits into Two Sigma’s investment strategy

* (34:05) Alpha and competition in investing

* (36:13) Temporality in data

* (40:38) Challenges for finance/AI and beating the market

* (44:36) Reproducibility

* (49:47) I Quant NY and storytelling with data

* (56:43) Descriptive statistics and stories

* (1:01:05) Benefits of simple methods

* (1:07:11) Outro

Links:

* Ben’s work on translational equivalence and scalable discriminative learning

* Two Sigma Insights

* Storytelling with data and I Quant NY

“There is this move from generality in a relative sense of ‘we are not as specialized as insects’ to generality in the sense of omnipotent, omniscient, godlike capabilities. And I think there's something very dangerous that happens there, which is you start thinking of the word ‘general’ in completely unhinged ways.”

In episode 114 of The Gradient Podcast, Daniel Bashir speaks to Venkatesh Rao.

Venkatesh is a writer and consultant. He has been writing the widely read Ribbonfarm blog since 2007, and more recently, the popular Ribbonfarm Studio Substack newsletter. He is the author of Tempo, a book on timing and decision-making, and is currently working on his second book, on the foundations of temporality. He has been an independent consultant since 2011, supporting senior executives in the technology industry. His work in recent years has focused on AI, semiconductor, sustainability, and protocol technology sectors. He holds a PhD in control theory (2003) from the University of Michigan. He is currently based in the Seattle area, and enjoys dabbling in robotics in his spare time. You can learn more about his work at venkateshrao.com

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:38) Origins of Ribbonfarm and Venkat’s academic background

* (04:23) Voice and recurring themes in Venkat’s work

* (11:45) Patch models and multi-agent systems: integrating philosophy of language, balancing realism with tractability

* (21:00) More on abstractions vs. tractability in Venkat’s work

* (29:07) Scaling of industrial value systems, characterizing AI as a discipline

* (39:25) Emergent science, intelligence and abstractions, presuppositions in science, generality and universality, cameras and engines

* (55:05) Psychometric terms

* (1:09:07) Inductive biases (yes I mentioned the No Free Lunch Theorem and then just talked about the definition of inductive bias and not the actual theorem 🤡)

* (1:18:13) LLM training and efficiency, comparing LLMs to humans

* (1:23:35) Experiential age, analogies for knowledge transfer

* (1:30:50) More clarification on the analogy

* (1:37:20) Massed Muddler Intelligence and protocols

* (1:38:40) Introducing protocols and the Summer of protocols

* (1:49:15) Evolution of protocols, hardness

* (1:54:20) LLMs, protocols, time, future visions, and progress

* (2:01:33) Protocols, drifting from value systems, friction, compiling explicit knowledge

* (2:14:23) Directions for ML people in protocols research

* (2:18:05) Outro

Links:

* Venkat’s Twitter and homepage

* Mediocre Computing

* Summer of Protocols and 2024 Call for Applications (apply!)

* Essays discussed

* Patch models and their applications to multivehicle command and control

* From Mediocre Computing

* Text is All You Need

* Magic, Mundanity, and Deep Protocolization

* A Camera, Not an Engine

* Massed Muddler Intelligence

* On protocols

* The Unreasonable Sufficiency of Protocols

* Protocols Don’t Build Pyramids

* Protocols in (Emergency) Time

* Atoms, Institutions, Blockchains

In episode 113 of The Gradient Podcast, Daniel Bashir speaks to Professor Sasha Rush.

Professor Rush is an Associate Professor at Cornell University and a Researcher at HuggingFace. His research aims to develop natural language processing systems that are safe, fast, and controllable. His group is interested primarily in tasks that involve text generation, and they study data-driven probabilistic methods that combine deep-learning based models with probabilistic controls. He is also interested in open-source NLP and deep learning, and develops projects to make deep learning systems safer, clearer, and easier to use.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:47) Professor Rush’s background

* (03:23) Professor Rush’s reflections on prior work—importance of learning and inference

* (04:58) How much engineering matters in deep learning, the Rush vs. Frankle Bet

* (07:12) On encouraging and incubating good research

* (10:50) Features of good research environments

* (12:36) 5% bets in Professor Rush’s research: State-Space Models (SSMs) as an alternative to Transformers

* (15:58) SSMs vs. Transformers

* (18:53) Probabilistic Context-Free Grammars—are (P)CFGs worth paying attention to?

* (20:53) Sequence-level knowledge distillation: approximating sequence-level distributions

* (25:08) Pruning and knowledge distillation — orthogonality of efficiency techniques

* (26:33) Broader thoughts on efficiency

* (28:31) Works on prompting

* (28:58) Prompting and In-Context Learning

* (30:05) Thoughts on mechanistic interpretability

* (31:25) Multitask prompted training enables zero-shot task generalization

* (33:48) How many data points is a prompt worth?

* (35:13) Directions for controllability in LLMs

* (39:11) Controllability and safety

* (41:23) Open-source work, deep learning libraries

* (42:08) A story about Professor Rush’s post-doc at FAIR

* (43:51) The impact of PyTorch

* (46:08) More thoughts on deep learning libraries

* (48:48) Levels of abstraction, PyTorch as an interface to motivate research

* (50:23) Empiricism and research commitments

* (53:32) Outro

Links:

* Research

* Early work / PhD

* Dual Decomposition and LP Relaxations

* Vine Pruning for Efficient Multi-Pass Dependency Parsing

* Improved Parsing and POS Tagging Using Inter-Sentence Dependency Constraints

* Research — interpretable and controllable natural language generation

* Compound Probabilistic Context-Free Grammars for Grammar Induction

* Multitask prompted training enables zero-shot task generalization

* Research — deep generative models

* A Neural Attention Model for Abstractive Sentence Summarization

* Learning Neural Templates for Text Generation

* How many data points is a prompt worth?

* Research — efficient algorithms and hardware for speech, translation, dialogue

* Sequence-Level Knowledge Distillation

* Open-source work

* NamedTensor

* Torch Struct

In episode 112 of The Gradient Podcast, Daniel Bashir speaks to Cameron Jones and Sean Trott.

Cameron is a PhD candidate in the Cognitive Science Department at the University of California, San Diego. His research compares how humans and large language models process language about world knowledge, situation models, and theory of mind.

Sean is an Assistant Teaching Professor in the Cognitive Science Department at the University of California, San Diego. His research interests include probing large language models, ambiguity in languages, how ambiguous words are represented, and pragmatic inference. He previously completed his PhD at UCSD.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:55) Cameron’s background

* (06:00) Sean’s background

* (08:15) Unexpected capabilities of language models and the need for embodiment to understand meaning

* (11:05) Interpreting results of Turing tests, separating what humans and LLMs do when behaving as though they “understand”

* (14:27) Internal mechanisms, interpretability, how we test theories

* (16:40) Languages are efficient, but for whom?

* (17:30) Initial motivations: lexical ambiguity

* (19:20) The balance of meanings across wordforms

* (22:35) Tension between speaker- and comprehender-oriented pressures in lexical ambiguity

* (25:05) Context and potential vs. realized ambiguity

* (27:15) LLM-ology

* (28:30) Studying LLMs as models of human cognition and as interesting objects of study in their own right

* (30:03) Example of explaining away effects

* (33:54) The internalist account of belief sensitivity—behavior and internal representations

* (37:43) LLMs and the False Belief Task

* (42:05) Hypothetical on observed behavior and inferences about internal representations

* (48:05) Distributional Semantics Still Can’t Account for Affordances

* (50:25) Tests of embodied theories and limitations of distributional cues

* (53:54) Multimodal models and object affordances

* (58:30) Language and grounding, other buzzwords

* (59:45) How could we know if LLMs understand language?

* (1:04:50) Reference: as a thing words do vs. ontological notion

* (1:11:38) The Role of Physical Inference in Pronoun Resolution

* (1:16:40) World models and world knowledge

* (1:19:45) EPITOME

* (1:20:20) The different tasks

* (1:26:43) Confounders / “attending” in LM performance on tasks

* (1:30:30) Another hypothetical, on theory of mind

* (1:32:26) How much information can language provide in service of mentalizing?

* (1:35:14) Convergent validity and coherence/validity of theory of mind

* (1:39:30) Interpretive questions about behavior w/r/t/ theory of mind

* (1:43:35) Does GPT-4 Pass the Turing Test?

* (1:44:00) History of the Turing Test

* (1:47:05) Interrogator strategies and the strength of the Turing Test

* (1:52:15) “Internal life” and personality

* (1:53:30) How should this research impact how we assess / think about LLM abilities?

* (1:58:56) Outro

Links:

* Cameron’s homepage and Twitter

* Sean’s homepage and Twitter

* Research — Language and NLP

* Languages are efficient, but for whom?

* Research — LLM-ology

* Do LLMs know what humans know?

* Distributional Semantics Still Can’t Account for Affordances

* In Cautious Defense of LLM-ology

* Should Psycholinguists use LLMs as “model organisms”?

* (Re)construing Meaning in NLP

* Research — language and grounding, theory of mind, reference [insert other buzzwords here]

* Do LLMs have a “theory of mind”?

* How could we know if LLMs understand language?

* Does GPT-4 Pass the Turing Test?

* Could LMs change language?

* The extended mind and why it matters for cognitive science research

* EPITOME

* The Role of Physical Inference in Pronoun Resolution

In episode 111 of The Gradient Podcast, Daniel Bashir speaks to Nicholas Thompson.

Nicholas is the CEO of The Atlantic. Previously, he served as editor-in-chief of Wired and editor of Newyorker.com. Nick also cofounded Atavist, which sold to Automattic in 2018. Publications under Nick’s leadership have won numerous National Magazine Awards and Pulitzer Prizes, and one WIRED story he edited was the basis for the movie Argo. Nick is also the co-founder of Speakeasy AI, a software platform designed to foster constructive online conversations about the world’s most pressing problems.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:12) Nick’s path into journalism

* (03:25) The Washington Monthly — a turning point

* (05:09) Perspectives from different positions in the journalism industry

* (08:16) What is great journalism?

* (09:42) Example from The Atlantic

* (11:00) Other examples/pieces of good journalism

* (12:20) Pieces on aging

* (12:56) Mortality and life-force associated with running — Nick’s piece in WIRED

* (15:30) On urgency

* (18:20) The job of an editor

* (22:23) AI in journalism — benefits and limitations

* (26:45) How AI can help writers, experimentation

* (28:40) Examples of AI in journalism and issues: CNET, Sports Illustrated, Nick’s thoughts on how AI should be used in journalism

* (32:20) Speakeasy AI and creating healthy conversation spaces

* (34:00) Details about Speakeasy

* (35:12) Business pivots and business model trouble

* (35:37) Remaining gaps in fixing conversational spaces

* (38:27) Lessons learned

* (40:00) Nick’s optimism about Speakeasy-like projects

* (43:14) Social simulacra, a “Troll WestWorld,” algorithmic adjustments in social media

* (46:11) Lessons and wisdom from journalism about engagement, more on engagement in social media

* (50:27) Successful and unsuccessful futures for AI in journalism

* (54:17) Previous warnings about synthetic media, Nick’s perspective on risks from synthetic media in journalism

* (57:00) Stop trying to build AGI

(59:13) Outro

Links:

* Nicholas’s Twitter and website

* Speakeasy AI

* Writing

* “To Run My Best Marathon at Age 44, I Had to Outrun My Past” in WIRED

* “The year AI actually changes the media business” in NiemanLab’s Predictions for Journalism 2023

In episode 110 of The Gradient Podcast, Daniel Bashir speaks to Professor Subbarao Kambhampati.

Professor Kambhampati is a professor of computer science at Arizona State University. He studies fundamental problems in planning and decision making, motivated by the challenges of human-aware AI systems. He is a fellow of the Association for the Advancement of Artificial Intelligence, American Association for the Advancement of Science, and Association for Computing machinery, and was an NSF Young Investigator. He was the president of the Association for the Advancement of Artificial Intelligence, trustee of the International Joint Conference on Artificial Intelligence, and a founding board member of Partnership on AI.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:11) Professor Kambhampati’s background

* (06:07) Explanation in AI

* (18:08) What people want from explanations—vocabulary and symbolic explanations

* (21:23) The realization of new concepts in explanation—analogy and grounding

* (30:36) Thinking and language

* (31:48) Conscious and subconscious mental activity

* (36:58) Tacit and explicit knowledge

* (42:09) The development of planning as a research area

* (46:12) RL and planning

* (47:47) What makes a planning problem hard?

* (51:23) Scalability in planning

* (54:48) LLMs do not perform reasoning

* (56:51) How to show LLMs aren’t reasoning

* (59:38) External verifiers and backprompting LLMs

* (1:07:51) LLMs as cognitive orthotics, language and representations

* (1:16:45) Finding out what kinds of representations an AI system uses

* (1:31:08) “Compiling” system 2 knowledge into system 1 knowledge in LLMs

* (1:39:53) The Generative AI Paradox, reasoning and retrieval

* (1:43:48) AI as an ersatz natural science

* (1:44:03) Why AI is straying away from its engineering roots, and what constitutes engineering

* (1:58:33) Outro

Links:

* Professor Kambhampati’s Twitter and homepage

* Research and Writing — Planning and Human-Aware AI Systems

* A Validation-structure-based theory of plan modification and reuse (1990)

* Challenges of Human-Aware AI Systems (2020)

* Polanyi vs. Planning (2021)

* LLMs and Planning

* Can LLMs Really Reason and Plan? (2023)

* On the Planning Abilities of LLMs (2023)

* Other

* Changing the nature of AI research

In episode 109 of The Gradient Podcast, Daniel Bashir speaks to Russ Maschmeyer.

Russ is the Product Lead for AI and Spatial Commerce at Shopify. At Shopify, he leads a team that looks at how AI can better empower entrepreneurs, with a particular interest in how image generation can help make the lives of business owners and merchants more productive. He previously led design for multiple services at Facebook and co-founded Primer, an AR-enabled interior design marketplace.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:50) Russ’s background and a hacked Kinect sensor

* (06:00) Instruments and emotion, embodiment and accessibility

* (08:45) Natural language as input and generative AI in creating emotive experiences

* (10:55) Work on search queries and recommendations at Facebook, designing for search

* (16:35) AI in the retail and entrepreneurial landscape

* (19:15) Shopify and AI for business owners

* (22:10) Vision and directions for AI in commerce

* (25:01) Personalized experiences for shopping

* (28:45) Challenges for creating personalized experiences

* (31:49) Intro to spatial commerce

* (34:48) AR/VR devices and spatial commerce

* (37:30) MR and AI for immersive product search

* (41:35) Implementation details

* (48:05) WonkaVision and difficulties for immersive web experiences

* (52:10) Future projects and directions for spatial commerce

* (55:10) Outro

Links:

* Russ’s Twitter and homepage

* With a Wave of the Hand, Improvising on Kinect in The New York Times

* Shopify Spatial Commerce Projects

* MR and AI for immersive product search

* A more immersive web with a simple optical illusion

* What if your room had a reset button?

In episode 108 of The Gradient Podcast, Daniel Bashir speaks to Professor Benjamin Breen.

Professor Breen is an associate professor of history at UC Santa Cruz specializing in the history of science, medicine, globalization, and the impacts of technological change. He is the author of multiple books including The Age of Intoxication: Origins of the Global Drug Trade and the more recent Tripping on Utopia: Margaret Mead, the Cold War, and the Troubled Birth of Psychedelic Science, which you can pre-order now.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:05) Professor Breen’s background

* (04:47) End of history narratives / millenarian thinking in AI/technology

* (09:53) Transformative technological change and societal change

* (16:45) AI and psychedelics

* (17:23) Techno-utopianism

* (26:08) Technologies as metaphors for humanity

* (32:34) McLuhanist thinking / brain as a computational machine, Prof. Breen’s skepticism

* (37:13) Issues with overblown narratives about technology

* (42:46) Narratives about transformation and their impacts on progress

* (45:23) The historical importance of today’s AI landscape

* (50:05) International aspects of the history of technology

* (53:13) Doomerism vs optimism, why doomerism is appealing

* (57:58) Automation, meta-skills, jobs — advice for early career

* (1:01:08) LLMs and (history) education

* (1:07:10) Outro

Links:

* Professor Breen’s Twitter and homepage

* Books

* Tripping on Utopia: Margaret Mead, the Cold War, and the Troubled Birth of Psychedelic Science

* The Age of Intoxication: Origins of the Global Drug Trade

* Writings

* Into the mystic

* ‘Alien Jesus’

* Simulating History with ChatGPT

In episode 107 of The Gradient Podcast, Daniel Bashir speaks to Professor Ted Gibson.

Ted is a Professor of Cognitive Science at MIT. He leads the TedLab, which investigates why languages look the way they do; the relationship between culture and cognition, including language; and how people learn, represent, and process language.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:13) Prof Gibson’s background

* (05:33) The computational linguistics community and NLP, engineering focus

* (10:48) Models of brains

* (12:03) Prof Gibson’s focus on behavioral work

* (12:53) How dependency distances impact language processing

* (14:03) Dependency distances and the origin of the problem

* (18:53) Dependency locality theory

* (21:38) The structures languages tend to use

* (24:58) Sentence parsing: structural integrations and memory costs

* (36:53) Reading strategies vs. ordinary language processing

* (40:23) Legalese

* (46:18) Cross-dependencies

* (50:11) Number as a cognitive technology

* (54:48) Experiments

* (1:03:53) Why counting is useful for Western societies

* (1:05:53) The Whorf hypothesis

* (1:13:05) Language as Communication

* (1:13:28) The noisy channel perspective on language processing

* (1:27:08) Fedorenko lab experiments—language for thought vs. communication and Chomsky’s claims

* (1:43:53) Thinking without language, inner voices, language processing vs. language as an aid for other mental processing

* (1:53:01) Dependency grammars and a critique of Chomsky’s grammar proposals, LLMs

* (2:08:48) LLM behavior and internal representations

* (2:12:53) Outro

Links:

* Ted’s lab page and Twitter

* Re-imagining our theories of language

* Research — linguistic complexity and dependency locality theory

* Linguistic complexity: locality of syntactic dependencies (1998)

* The Dependency Locality Theory: A Distance-Based Theory of Linguistic Complexity (2000)

* Consequences of the Serial Nature of Linguistic Input for Sentential Complexity (2005)

* Large-scale evidence of dependency length minimization in 37 languages (2015)

* Dependency locality as an explanatory principle for word order (2020)

* Robust effects of working memory demand during naturalistic language comprehension in language-selective cortex (2022)

* A resource-rational model of human processing of recursive linguistic structure (2022)

* Research — language processing / communication and cross-linguistic universals

* Number as a cognitive technology: Evidence from Pirahã language and cognition (2008)

* The communicative function of ambiguity in language (2012)

* The rational integration of noisy evidence and prior semantic expectations in sentence interpretation (2013)

* Color naming across languages reflects color use (2017)

* How Efficiency Shapes Human Language (2019)

In episode 106 of The Gradient Podcast, Daniel Bashir speaks to Professor Harvey Lederman.

Professor Lederman is a professor of philosophy at UT Austin. He has broad interests in contemporary philosophy and in the history of philosophy: his areas of specialty include philosophical logic, the Ming dynasty philosopher Wang Yangming, epistemology, and philosophy of language. He has recently been working on incomplete preferences, on trying in the philosophy of language, and on Wang Yangming’s moral metaphysics.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:15) Harvey’s background

* (05:30) Higher-order metaphysics and propositional attitudes

* (06:25) Motivations

* (12:25) Setup: syntactic types and ontological categories

* (25:11) What makes higher-order languages meaningful and not vague?

* (25:57) Higher-order languages corresponding to the world

* (30:52) Extreme vagueness

* (35:32) Desirable features of languages and important questions in philosophy

* (36:42) Higher-order identity

* (40:32) Intuitions about mental content, language, context-sensitivity

* (50:42) Perspectivism

* (51:32) Co-referring names, identity statements

* (55:42) The paper’s approach, “know” as context-sensitive

* (57:24) Propositional attitude psychology and mentalese generalizations

* (59:57) The “good standing” of theorizing about propositional attitudes

* (1:02:22) Mentalese

* (1:03:32) “Does knowledge imply belief?” — when a question does not have good standing

* (1:06:17) Sense, Reference, and Substitution

* (1:07:07) Fregeans and the principle of Substitution

* (1:12:12) Follow-up work to this paper

* (1:13:39) Do Language Models Produce Reference Like Libraries or Like Librarians?

* (1:15:02) Bibliotechnism

* (1:19:08) Inscriptions and reference, what it takes for something to refer

* (1:22:37) Derivative and basic reference

* (1:24:47) Intuition: n-gram models and reference

* (1:28:22) Meaningfulness in sentences produced by n-gram models

* (1:30:40) Bibliotechnism and LLMs, disanalogies to n-grams

* (1:33:17) On other recent work (vector grounding, do LMs refer?, etc.)

* (1:40:12) Causal connections and reference, how bibliotechnism makes good on the meanings of sentences

* (1:45:46) RLHF, sensitivity to truth and meaningfulness

* (1:48:47) Intelligibility

* (1:50:52) When LLMs produce novel reference

* (1:53:37) Novel reference vs. find-replace

* (1:56:00) Directionality example

* (1:58:22) Human intentions and derivative reference

* (2:00:47) Between bibliotechnism and agency

* (2:05:32) Where do invented names / novel reference come from?

* (2:07:17) Further questions

* (2:10:04) Outro

Links:

* Harvey’s homepage and Twitter

* Papers discussed

* Higher-order metaphysics and propositional attitudes

* Perspectivism

* Sense, Reference, and Substitution

* Are Language Models More Like Libraries or Like Librarians? Bibliotechnism, the Novel Reference Problem, and the Attitudes of LLMs

In episode 105 of The Gradient Podcast, Daniel Bashir speaks to Eric Jang.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:25) Updates since Eric’s last interview

* (06:07) The problem space of humanoid robots

* (08:42) Motivations for the book “AI is Good for You”

* (12:20) Definitions of AGI

* (14:35) ~ AGI timelines ~

* (16:33) Do we have the ingredients for AGI?

* (18:58) Rediscovering old ideas in AI and robotics

* (22:13) Ingredients for AGI

* (22:13) Artificial Life

* (25:02) Selection at different levels of information—intelligence at different scales

* (32:34) AGI as a collective intelligence

* (34:53) Human in the loop learning

* (37:38) From getting correct answers to doing things correctly

* (40:20) Levels of abstraction for modeling decision-making — the neurobiological stack

* (44:22) Implementing loneliness and other details for AGI

* (47:31) Experience in AI systems

* (48:46) Asking for Generalization

* (49:25) Linguistic relativity

* (52:17) Language vs. complex thought and Fedorenko experiments

* (54:23) Efficiency in neural design

* (57:20) Generality in the human brain and evolutionary hypotheses

* (59:46) Embodiment and real-world robotics

* (1:00:10) Moravec’s Paradox and the importance of embodiment

* (1:05:33) How embodiment fits into the picture—in verification vs. in learning

* (1:10:45) Nonverbal information for training intelligent systems

* (1:11:55) AGI and humanity

* (1:12:20) The positive future with AGI

* (1:14:55) The negative future — technology as a lever

* (1:16:22) AI in the military

* (1:20:30) How AI might contribute to art

* (1:25:41) Eric’s own work and a positive future for AI

* (1:29:27) Outro

Links:

* Eric’s book

* Eric’s Twitter and homepage

In episode 104 of The Gradient Podcast, Daniel Bashir speaks to Nathan Benaich.

Nathan is Founder and General Partner at Air Street Capital, a VC firm focused on investing in AI-first technology and life sciences companies. Nathan runs a number of communities focused on AI including the Research and Applied AI Summit and leads Spinout.fyi to improve the creation of university spinouts. Nathan co-authors the State of AI Report.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:00) Updates in Nathan World — Air Street’s second fund, spinouts,

* (07:30) Events: Research and Applied AI Summit, State of AI Report launches

* (09:50) The State of AI: main messages, the increasing role of subject matter experts

* Research

* (14:13) Open and closed-source

* (17:55) Benchmarking and evaluation, small/large models and industry verticals

* (21:10) “Vibes” in LLM evaluation

* (24:00) Codegen models, personalized AI, curriculum learning

* (26:20) The exhaustion of human-generated data, lukewarm content, synthetic data

* (29:50) Opportunities for AI applications in the natural sciences

* (35:15) Reinforcement Learning from Human Feedback and alternatives

* (38:30) Industry

* (39:00) ChatGPT and productivity

* (42:37) General app wars, ChatGPT competitors

* (45:50) Compute—demand, supply, competition

* (50:55) Export controls and geopolitics

* (54:45) Startup funding and compute spend

* (59:15) Politics

* (59:40) Calls for regulation, regulatory divergence

* (1:04:40) AI safety

* (1:07:30) Nathan’s perspective on regulatory approaches

* (1:12:30) The UK’s early access to frontier models, standards setting, regulation difficulties

* (1:17:20) Jailbreaking, constitutional AI, robustness

* (1:20:50) Predictions!

* (1:25:00) Generative AI misuse in elections and politics (and, this prediction coming true in Bangladesh)

* (1:26:50) Progress on AI governance

* (1:30:30) European dynamism

* (1:35:08) Outro

Links:

* Nathan’s homepage and Twitter

* The 2023 State of AI Report

* Bringing Dynamism to European Defense

* A prediction coming true: How AI is disrupting Bangladesh’s election

* Air Street Capital is hiring a full-time Community Lead!

In episode 103 of The Gradient Podcast, Daniel Bashir speaks to Dr. Kathleen Fisher.

As the director of DARPA’s Information Innovation Office (I2O), Dr. Kathleen Fisher oversees a portfolio that includes most of the agency’s AI-related research and development efforts, including the recent AI Forward initiative. AI Forward explores new directions for AI research that will result in trustworthy systems for national security missions. This summer, roughly 200 participants from the commercial sector, academia, and the U.S. government attended workshops that generated ideas to inform DARPA’s next phase of AI exploratory projects. Dr. Fisher previously served as a program manager in I2O from 2011 to 2014. As a program manager, she conceptualized, created, and executed programs in high-assurance computing and machine learning, including Probabilistic Programming for Advancing Machine Learning (PPAML), making building ML applications easier. She was also a co-author of a recent paper about the threats posed by large language models.

Since 2018, DARPA has dedicated over $2 billion in R&D funding to AI research. The agency DARPA has been generating groundbreaking research and development for 65 years – leading to game-changing military capabilities and icons of modern society, such as initiating the research field that rendered self-driving cars and developing the technology that led to Apple’s Siri.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:30) Kathleen’s background

* (05:05) Intersections between programming languages and AI

* (07:15) Neuro-symbolic AI, trade-offs between flexibility and guarantees

* (09:45) History of DARPA and the Information Innovation Office (I2O)

* (13:55) DARPA’s perspective on research

* (17:10) Galvanizing a research community

* (20:06) DARPA’s recent investments in AI and AI Forward

* (26:35) Dual-use nature of generative AI, identifying and mitigating security risks, Kathleen’s perspective on short-term and long-term risk (note: the “Gradient podcast” Kathleen mentions is from Last Week in AI)

* (30:10) Concerns about deployment and interaction

* (32:20) Outcomes from AI Forward workshops and themes

* (36:10) Incentives in building and using AI technologies, friction

* (38:40) Interactions between DARPA and other government agencies

* (40:09) Future research directions

* (44:04) Ways to stay up to date on DARPA’s work

* (45:40) Outro

Links:

* DARPA I2O website

* Probabilistic Programming for Advancing Machine Learning (PPAML) (Archived)

* Assured Neuro Symbolic Learning and Reasoning (ANSR)

* AI Cyber Challenge

* AI Forward

* Identifying and Mitigating the Security Risks of Generative AI Paper

* FoundSci Solicitation

* FACT Solicitation

* Semantic Forensics (SemaFor)

* GARD Open Source Resources

* I2O Newsletter signup

In episode 102 of The Gradient Podcast, Daniel Bashir speaks to Peter Tse.

Professor Tse is a Professor of Cognitive Neuroscience and chair of the department of Psychological and Brain Sciences at Dartmouth College. His research focuses on using brain and behavioral data to constrain models of the neural bases of attention and consciousness, unconscious processing that precedes and constructs consciousness, mental causation, and human capacities for imagination and creativity. He is especially interested in the processing that goes into the construction of conscious experience between retinal activation at time 0 and seeing an event about a third of a second later.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:45) Prof. Tse’s background

* (03:25) Early experiences in physics/math and philosophy of physics

* (06:10) Choosing to study neuroscience

* (07:15) Prof Tse’s commitments about determinism

* (10:00) Quantum theory and determinism

* (13:45) Biases/preferences in choosing theories

* (20:41) Falsifiability and scientific questions, transition from physics to neuroscience

* (30:50) How neuroscience is unusual among the sciences

* (33:20) Neuroscience and subjectivity

* (34:30) Reductionism

* (37:30) Gestalt psychology

* (41:30) Introspection in neuroscience

* (45:30) The preconscious buffer and construction of conscious experience, color constancy

* (53:00) Perceptual and cognitive inference

* (55:00) AI systems and intrinsic meaning

* (57:15) Information vs. meaning

* (1:01:45) Consciousness and representation of bodily states

* (1:05:10) Our second-order free will

* (1:07:20) Jaegwon Kim’s exclusion argument

* (1:11:45) Why Kim thought his own argument was wrong

* (1:15:00) Resistance and counterarguments to Kim

* (1:19:45) Criterial causation

* (1:23:00) How neurons evaluate inputs criterially

* (1:24:00) Concept neurons in the hippocampus

* (1:31:57) Criterial causation and physicalism, mental causation

* (1:40:10) Daniel makes another attempt to push back 🤡

* (1:45:47) More on AI

* (1:47:05) Prof Tse’s perspective on modern AI systems, differences with human cognition

* (2:17:25) Consciousness, attention, spirituality

* (2:20:10) Prof Tse’s hopes for AI

* (2:23:30) Outro

Links:

* Professor Tse’s homepage

* Papers

* Vision/Perception

* Perceptual learning based on the learning of diagnostic features

* Complete mergeability and amodal completion

* Attention

* How Attention Can Alter Appearances

* How Top-down Attention Alters Bottom-up preconscious operations

* Consciousness

* Network structure and dynamics of the mental workspace

* On free will

* NDPR review of “Neural Basis of Free Will”

* Kripke’s Category Error

* Ontological Indeterminism undermines Kim’s Exclusion Argument

In episode 101 of The Gradient Podcast, Daniel Bashir speaks to Vera Liao.

Vera is a Principal Researcher at Microsoft Research (MSR) Montréal where she is part of the FATE (Fairness, Accountability, Transparency, and Ethics) group. She is trained in human-computer interaction research and works on human-AI interaction, currently focusing on explainable AI and responsible AI. She aims to bridge emerging AI technologies and human-centered design practices, and use both qualitative and quantitative methods to generate recommendations for technology design. Before joining MSR, Vera worked at IBM TJ Watson Research Center, and her work contributed to IBM products such as AI Explainability 360, Uncertainty Quantification 360, and Watson Assistant.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:41) Vera’s background

* (07:15) The sociotechnical gap

* (09:00) UX design and toolkits for AI explainability

* (10:50) HCI, explainability, etc. as “separate concerns” from core AI reseaarch

* (15:07) Interfaces for explanation and model capabilities

* (16:55) Vera’s earlier studies of online social communities

* (22:10) Technologies and user behavior

* (23:45) Explainability vs. interpretability, transparency

* (26:25) Questioning the AI: Informing Design Practices for Explainable AI User Experiences

* (42:00) Expanding Explainability: Towards Social Transparency in AI Systems

* (50:00) Connecting Algorithmic Research and Usage Contexts

* (59:40) Pitfalls in existing explainability methods

* (1:05:35) Ideal and real users, seamful systems and slow algorithms

* (1:11:08) AI Transparency in the Age of LLMs: A Human-Centered Research Roadmap

* (1:11:35) Vera’s earlier experiences with chatbots

* (1:13:00) Need to understand pitfalls and use-cases for LLMs

* (1:13:45) Perspectives informing this paper

* (1:20:30) Transparency informing goals for LLM use

* (1:22:45) Empiricism and explainability

* (1:27:20) LLM faithfulness

* (1:32:15) Future challenges for HCI and AI

* (1:36:28) Outro

Links:

* Vera’s homepage and Twitter

* Research

* Earlier work

* Understanding Experts’ and Novices’ Expertise Judgment of Twitter Users

* Beyond the Filter Bubble

* Expert Voices in Echo Chambers

* HCI / collaboration

* Exploring AI Values and Ethics through Participatory Design Fictions

* Ways of Knowing for AI: (Chat)bots as Interfaces for ML

* Human-AI Collaboration: Towards Socially-Guided Machine Learning

* Questioning the AI: Informing Design Practices for Explainable AI User Experiences

* Rethinking Model Evaluation as Narrowing the Socio-Technical Gap

* Human-Centered XAI: From Algorithms to User Experiences

* AI Transparency in the Age of LLMs: A Human-Centered Research Roadmap

* Fairness and explainability

* Questioning the AI: Informing Design Practices for Explainable AI User Experiences

* Expanding Explainability: Towards Social Transparency in AI Systems

* Connecting Algorithmic Research and Usage Contexts

In episode 100 of The Gradient Podcast, Daniel Bashir speaks to Professor Thomas Dietterich.

Professor Dietterich is Distinguished Professor Emeritus in the School of Electrical Engineering and Computer Science at Oregon State University. He is a pioneer in the field of machine learning, and has authored more than 225 refereed publications and two books. His current research topics include robust artificial intelligence, robust human-AI systems, and applications in sustainability. He is a former President of the Association for the Advancement of Artificial Intelligence, and the founding President of the International Machine Learning Society. Other major roles include Executive Editor of the journal Machine Learning, co-founder of the Journal for Machine Learning Research, and program chair of AAAI 1990 and NIPS 2000. He currently serves as one of the moderators for the cs.LG category on arXiv.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Episode 100 Note

* (02:03) Intro

* (04:23) Prof. Dietterich’s background

* (14:20) Kuhn and theory development in AI, how Prof Dietterich thinks about the philosophy of science and AI

* (20:10) Scales of understanding and sentience, grounding, observable evidence

* (23:58) Limits of statistical learning without causal reasoning, systematic understanding

* (25:48) A challenge for the ML community: testing for systematicity

* (26:13) Forming causal understandings of the world

* (28:18) Learning at the Knowledge Level

* (29:18) Background and definitions

* (32:18) Knowledge and goals, a note on LLMs

* (33:03) What it means to learn

* (41:05) LLMs as learning results of inference without learning first principles

* (43:25) System I/II thinking in humans and LLMs

* (47:23) “Routine Science”

* (47:38) Solving multiclass learning problems via error-correcting output codes

* (52:53) Error-correcting codes and redundancy

* (54:48) Why error-correcting codes work, contra intuition

* (59:18) Bias in ML

* (1:06:23) MAXQ for hierarchical RL

* (1:15:48) Computational sustainability

* (1:19:53) Project TAHMO’s moonshot

* (1:23:28) Anomaly detection for weather stations

* (1:25:33) Robustness

* (1:27:23) Motivating The Familiarity Hypothesis

* (1:27:23) Anomaly detection and self-models of competence

* (1:29:25) Measuring the health of freshwater streams

* (1:31:55) An open set problem in species detection

* (1:33:40) Issues in anomaly detection for deep learning

* (1:37:45) The Familiarity Hypothesis

* (1:40:15) Mathematical intuitions and the Familiarity Hypothesis

* (1:44:12) What’s Wrong with LLMs and What We Should Be Building Instead

* (1:46:20) Flaws in LLMs

* (1:47:25) The systems Prof Dietterich wants to develop

* (1:49:25) Hallucination/confabulation and LLMs vs knowledge bases

* (1:54:00) World knowledge and linguistic knowledge

* (1:55:07) End-to-end learning and knowledge bases

* (1:57:42) Components of an intelligent system and separability

* (1:59:06) Thinking through external memory

* (2:01:10) Outro

Links:

* Research — Fundamentals (Philosophy of AI)

* Learning at the Knowledge Level

* What Does it Mean for a Machine to Understand?

* Research – “Routine science”

* Ensemble methods in ML and error-correcting output codes

* Solving multiclass learning problems via error-correcting output codes

* An experimental comparison of bagging, boosting, and randomization

* ML Bias, Statistical Bias, and Statistical Variance of Decision Tree Algorithms

* The definitive treatment of these questions, by Gareth James

* Discovering/Exploiting structure in MDPs:

* MAXQ for hierarchical RL

* Exogenous State MDPs (paper with George Trimponias, slides)

* Research — Ecosystem Informatics and Computational Sustainability

* Project TAHMO

* Challenges for ML in Computational Sustainability

* Research — Robustness

* Steps towards robust AI (AAAI President’s Address)

* Benchmarking NN Robustness to Common Corruptions and Perturbations with Dan Hendrycks

* The familiarity hypothesis: Explaining the behavior of deep open set methods

* Recent commentary

* Toward High-Reliability AI

* What's Wrong with Large Language Models and What We Should Be Building Instead

In episode 99 of The Gradient Podcast, Daniel Bashir speaks to Professor Martin Wattenberg.

Professor Wattenberg is a professor at Harvard and part-time member of Google Research’s People + AI Research (PAIR) initiative, which he co-founded. His work, with long-time collaborator Fernanda Viégas, focuses on making AI technology broadly accessible and reflective of human values. At Google, Professor Wattenberg, his team, and Professor Viégas have created end-user visualizations for products such as Search, YouTube, and Google Analytics. Note: Professor Wattenberg is recruiting PhD students through Harvard SEAS—info here.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (03:30) Prof. Wattenberg’s background

* (04:40) Financial journalism at SmartMoney

* (05:35) Contact with the academic visualization world, IBM

* (07:30) Transition into visualizing ML

* (08:25) Skepticism of neural networks in the 1980s

* (09:45) Work at IBM

* (10:00) Multiple scales in information graphics, organization of information

* (13:55) How much information should a graphic display to whom?

* (17:00) Progressive disclosure of complexity in interface design

* (18:45) Visualization as a rhetorical process

* (20:45) Conversation Thumbnails for Large-Scale Discussions

* (21:35) Evolution of conversation interfaces—Slack, etc.

* (24:20) Path dependence — mutual influences between user behaviors and technology, takeaways for ML interface design

* (26:30) Baby Names and Social Data Analysis — patterns of interest in baby names

* (29:50) History Flow

* (30:05) Why investigate editing dynamics on Wikipedia?

* (32:06) Implications of editing patterns for design and governance

* (33:25) The value of visualizations in this work, issues with Wikipedia editing

* (34:45) Community moderation, bureaucracy

* (36:20) Consensus and guidelines

* (37:10) “Neutral” point of view as an organizing principle

* (38:30) Takeaways

* PAIR

* (39:15) Tools for model understanding and “understanding” ML systems

* (41:10) Intro to PAIR (at Google)

* (42:00) Unpacking the word “understanding” and use cases

* (43:00) Historical comparisons for AI development

* (44:55) The birth of TensorFlow.js

* (47:52) Democratization of ML

* (48:45) Visualizing translation — uncovering and telling a story behind the findings

* (52:10) Shared representations in LLMs and their facility at translation-like tasks

* (53:50) TCAV

* (55:30) Explainability and trust

* (59:10) Writing code with LMs and metaphors for using

* More recent research

* (1:01:05) The System Model and the User Model: Exploring AI Dashboard Design

* (1:10:05) OthelloGPT and world models, causality

* (1:14:10) Dashboards and interaction design—interfaces and core capabilities

* (1:18:07) Reactions to existing LLM interfaces

* (1:21:30) Visualizing and Measuring the Geometry of BERT

* (1:26:55) Note/Correction: The “Atlas of Meaning” Prof. Wattenberg mentions is called Context Atlas

* (1:28:20) Language model tasks and internal representations/geometry

* (1:29:30) LLMs as “next word predictors” — explaining systems to people

* (1:31:15) The Shape of Song

* (1:31:55) What does music look like?

* (1:35:00) Levels of abstraction, emergent complexity in music and language models

* (1:37:00) What Prof. Wattenberg hopes to see in ML and interaction design

* (1:41:18) Outro

Links:

* Professor Wattenberg’s homepage and Twitter

* Harvard SEAS application info — Professor Wattenberg is recruiting students!

* Research

* Earlier work

* A Fuzzy Commitment Scheme

* Stacked Graphs—Geometry & Aesthetics

* A Multi-Scale Model of Perceptual Organization in Information Graphics

* Conversation Thumbnails for Large-Scale Discussions

* Baby Names and Social Data Analysis

* History Flow (paper)

* At Harvard and Google / PAIR

* Tools for Model Understanding: Facets, SmoothGrad, Attacking discrimination with smarter ML

* TensorFlow.js

* Visualizing translation

* TCAV

* Other ML papers:

* The System Model and the User Model: Exploring AI Dashboard Design (recent speculative essay)

* Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task

* Visualizing and Measuring the Geometry of BERT

* Artwork

* The Shape of Song

In episode 98 of The Gradient Podcast, Daniel Bashir speaks to Laurence Liew.

Laurence is the Director for AI Innovation at AI Singapore. He is driving the adoption of AI by the Singapore ecosystem through the 100 Experiments, AI Apprenticeship Programmes and the Generational AI Talent Development initiative. He is the current Co-Chair of the Innovations and Commercialisation working group and Co-Chair of the "Broad Adoption of AI by SME" committee.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:25) Laurence’s background

* (07:00) AI Singapore and Singapore’s AI Strategy

* (08:27) Awareness and adoption of AI in Singapore

* (19:45) AI Apprenticeship Program stories

* (27:35) Developing generational AI talent within Singapore, literacy

* (32:25) Singapore’s place within the global AI ecosystem

* (38:30) How the generative AI boom has affected Singapore

* (43:50) Laurence’s vision for the future of Singapore’s tech ecosystem

* (49:41) Outro

Links:

* AI Singapore

In episode 97 of The Gradient Podcast, Daniel Bashir speaks to Professor Michael Levin and Adam Goldstein.

Professor Levin is a Distinguished Professor and Vannevar Bush Chair in the Biology Department at Tufts University. He also directs the Allen Discovery Center at Tufts. His group, the Levin Lab, focuses on understanding the biophysical mechanisms that implement decision-making during complex pattern regulation, and harnessing endogenous bioelectric dynamics toward rational control of growth and form. 

Adam Goldstein was a visiting scientist at the Levin Lab, where he worked on cancer research, and is the co-founder and Chairman of Astonishing Labs. Previously Adam founded Hipmunk, wrote tech books for O'Reilly, and was a Visiting Partner at Y Combinator.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:37) Intros

* (03:20) Prof. Levin intro

* (04:26) Adam intro

* (06:25) A perspective on intelligence

* (08:40) Diverse intelligence — in unconventional embodiments and unfamiliar spaces, substrate independence

* (12:23) Failure of the life-machine distinction, text-based systems, grounding, and embodiment

* (16:12) What it is to be a Self, fluidity and persistence

* (22:45) The combination problem in cognitive function, levels and representation

* (27:10) Goals for AI / cognitive science, Prof Levin’s perspective on building intelligent systems

* (31:25) Adam’s and Prof. Levin’s recent research—regenerative medicine and cancer

* (36:25) Examples of regeneration, Adam on the right approach to the regeneration problem as generation

* (45:25) Protein engineering vs. Adam and Prof. Levin’s program, implicit assumptions underlying biology

* (48:15) Regeneration example in liver disease

* (50:50) Perspectives on AI and its goals

Links:

* Levin Lab homepage

* Forms of life, forms of mind

* Adam’s homepage

* Research

* On Having No Head: Cognition throughout Biological Systems

* Technological Approach to Mind Everywhere

* Living Things Are Not (20th Century) Machines: Updating Mechanism Metaphors in Light of the Modern Science of Machine Behavior

* Life, death, and self: Fundamental questions of primitive cognition viewed through the lens of body plasticity and synthetic organisms

* Modular cognition

* Endless Forms

* Future Medicine: from molecular pathways to the collective intelligence of the body

* Technological Approach to Mind Everywhere: an experimentally-grounded framework for understanding diverse bodies and minds

* The Computational Boundary of a “Self”: Developmental Bioelectricity Drives Multicellularity and Scale-Free Cognition

* Machine life

In episode 96 of The Gradient Podcast, Daniel Bashir speaks to Jonathan Frankle.

Jonathan is the Chief Scientist at MosaicML and (as of release). Jonathan completed his PhD at MIT, where he investigated the properties of sparse neural networks that allow them to train effectively through his lottery ticket hypothesis. He also spends a portion of his time working on technology policy, and currently works with the OECD to implement the AI principles he helped develop in 2019.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:35) Jonathan’s background and work

* (04:25) Origins of the Lottery Ticket Hypothesis

* (06:00) Jonathan’s empiricism and approach to science

* (08:25) More Karl Popper discourse + hot takes

* (09:45) Walkthrough of the Lottery Ticket Hypothesis

* (12:00) Issues with the Lottery Ticket Hypothesis as a statement

* (12:30) Jonathan’s advice for PhD students, on asking good questions

* (15:55) Strengths and Promise of the Lottery Ticket Hypothesis

* (18:55) More Lottery Ticket Hypothesis Papers

* (19:10) Comparing Rewinding and Fine-tuning

* (23:00) Care in making experimental choices

* (25:05) Linear Mode Connectivity and the Lottery Ticket Hypothesis

* (27:50) On what is being measured and how

* (28:50) “The outcome of optimization is determined to a linearly connected region”

* (31:15) On good metrics

* (32:54) On the Predictability of Pruning Across Scales — scaling laws for pruning

* (34:40) The paper’s takeaway

* (38:45) Pruning Neural Networks at Initialization — on a scientific disagreement

* (45:00) On making takedown papers useful

* (46:15) On what can be known early in training

* (49:15) Jonathan’s perspective on important research questions today

* (54:40) MosaicML

* (55:19) How Mosaic got started

* (56:17) Mosaic highlights

* (57:33) Customer stories

* (1:00:30) Jonathan’s work and perspectives on AI policy

* (1:05:45) The key question: what we want

* (1:07:35) Outro

Links:

* Jonathan’s homepage and Twitter

* Papers

* The Lottery Ticket Hypothesis and follow-up work

* Comparing Rewinding and Fine-tuning in Neural Network Pruning

* Linear Mode Connectivity and the LTH

* On the Predictability of Pruning Across Scales

* Pruning Neural Networks at Initialization: Why Are We Missing The Mark?

* Desirable Inefficiency

In episode 95 of The Gradient Podcast, Daniel Bashir speaks to Nao Tokui.

Nao Tokui is an artist/DJ and researcher based in Tokyo. While pursuing his Ph.D. at The University of Tokyo, he produced his first music album and singles using AI, including a 12-inch record with Nujabes, a legendary Japanese hip-hop producer. After completing his Ph.D. research, he founded Qosmo, AI Creativity and Music Lab, in 2009. Since then, he has been actively working at the intersection of AI technology and art. Nao and his team's works have been exhibited at renowned venues such as the New York MoMA and the Barbican Centre in London. Their performances have also been showcased at various music festivals, including MUTEK and Sonar. Additionally, he is leading the development of AI-based music instruments at his newly founded company, Neutone. In 2021, Nao received the Okawa Publishing Award for his Japanese book on art, creativity, and AI. The book is scheduled to be released in English as "Surfing human creativity with AI — A user's guide" in November 2023.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:15) Nao’s background and how he got into AI and music

* (05:10) Nao’s experiences as a DJ, collaboration with Nujabes

* (07:10) HCI and music

* (10:35) Leveraging the difference between AI systems and humans

* (12:40) Total control vs total chaos

* (13:45) Qosmo and the Neutone Project, misusable AI tools

* (17:25) On music and “creating something new”

* (21:00) Declarative and top-down vs. bottom-up creation, individual taste

* (23:50) How generative AI enables humans

* (26:25) On misusing technology and art

* (32:00) Dawn Patrol EP

* (36:00) A two-discriminator GAN for creating music in new genres

* (37:45) The AI DJ Project

* (38:20) The interactive vision of the project

* (42:10) How AI chooses music, breaking from constraints

* (43:15) Interpretability and how an AI system DJs differently

* (45:15) How the project altered Nao’s perspective on DJing, the role of humans

* (51:40) Nao’s book Creating with AI

* (55:15) Human-AI interaction as joint improvisation

* (58:10) Nao’s advice and takeaways for thinking about AI creatively

* (1:01:32) Outro

Links:

* Nao’s homepage and Twitter

* Other links:

* Neutone, AI audio plugin

* Real-time AI-generative DJ performance

* Qosmo

* Dawn Patrol EP

* Nao’s book: Surfing human creativity with AI — A user's guide

* Paper on Creative-GAN for deviating from existing music genres

In episode 94 of The Gradient Podcast, Daniel Bashir speaks to Divyansh Kaushik.

Divyansh is the Associate Director for Emerging Technologies and National Security at the Federation of American Scientists where his focus areas include, amongst other things, AI policy, STEM immigration, and US-China strategic competition. He holds a PhD from Carnegie Mellon University, where he focused on designing reliable AI systems that align with human values. In addition to his advocacy work on Capitol Hill, he also played a key role in establishing the Congressional Graduate Research and Development Caucus. He is a frequent contributor to leading publications, including Vox, National Defense Magazine, The Dispatch, Daily Caller, and Forbes.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:20) Divyansh intro/background

* (06:00) Zachary Lipton Appreciation Session ( + advice from Prof Lipton)

* (08:00) How Divyansh got involved in policy

* (11:30) What does policy work look like? Divyansh’s early experiences

* (15:42) AI policy issues, divides, party lines

* (19:15) Bringing AI talent into the US

* (26:45) US/China saber rattling, impact of Xi Jinping’s presidency

* (33:49) China’s AI regulations, CCP motivations, China’s disadvantages in AI and benefits of the US policy process

* (42:42) Trading off AI governance and stifling innovation

* (51:17) AI governance comments from Jeremy Howard / Connor Leahy / Andrew Maynard, regulating use vs basic technology, limits on scaling

* (1:01:30) Articulating and communicating the issues for AI governance

* (1:03:10) Existential risk concerns in AI governance, theories of change

* (1:10:15) How can AI researchers/practitioners better communicate with policymakers?

* (1:16:57) Outro

Links:

* Divyansh’s Twitter and FAS page

* Divyansh’s policy work:

* The impact of international scientists, engineers, and students on US research outputs and global competitiveness

* How Congress can shape AI governance without stifling innovation

* How Do OpenAI’s Efforts To Make GPT-4 “Safer” Stack Up Against The NIST AI Risk Management Framework?

* Six Policy Ideas for the National AI Strategy

* Other work mentioned/discussed:

* Jeremy Howard’s AI Safety and the Age of Dislightenment

* Proposals from Connor Leahy

* Andrew Maynard’s Regulating Frontier AI: To Open Source or Not?

In episode 93 of The Gradient Podcast, Daniel Bashir speaks to Professor Tal Linzen.

Professor Linzen is an Associate Professor of Linguistics and Data Science at New York University and a Research Scientist at Google. He directs the Computation and Psycholinguistics Lab, where he and his collaborators use behavioral experiments and computational methods to study how people learn and understand language. They also develop methods for evaluating, understanding, and improving computational systems for language processing.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:25) Prof. Linzen’s background

* (05:37) Back and forth between psycholinguistics and deep learning research, LM evaluation

* (08:40) How can deep learning successes/failures help us understand human language use, methodological concerns, comparing human representations to LM representations

* (14:22) Behavioral capacities and degrees of freedom in representations

* (16:40) How LMs are becoming less and less like humans

* (19:25) Assessing LSTMs’ ability to learn syntax-sensitive dependencies

* (22:48) Similarities between structure-sensitive dependencies, sophistication of syntactic representations

* (25:30) RNNs implicitly implement tensor-product representations—vector representations of symbolic structures

* (29:45) Representations required to solve certain tasks, difficulty of natural language

* (33:25) Accelerating progress towards human-like linguistic generalization

* (34:30) The pre-training agnostic identically distributed evaluation paradigm

* (39:50) Ways to mitigate differences in evaluation

* (44:20) Surprisal does not explain syntactic disambiguation difficulty

* (45:00) How to measure processing difficulty, predictability and processing difficulty

* (49:20) What other factors influence processing difficulty?

* (53:10) How to plant trees in language models

* (55:45) Architectural influences on generalizing knowledge of linguistic structure

* (58:20) “Cognitively relevant regimes” and speed of generalization

* (1:00:45) Acquisition of syntax and sampling simpler vs. more complex sentences

* (1:04:03) Curriculum learning for progressively more complicated syntax

* (1:05:35) Hypothesizing tree-structured representations

* (1:08:00) Reflecting on a prediction from the past

* (1:10:15) Goals and “the correct direction” in AI research

* (1:14:04) Outro

Links:

* Prof. Linzen’s Twitter and homepage

* Papers

* Assessing the Ability of LSTMs to Learn Syntax-Sensitive Dependencies

* RNNS Implicitly Implement Tensor-Product Representations

* How Can We Accelerate Progress Towards Human-like Linguistic Generalization?

* Surprisal does not explain syntactic disambiguation difficulty: evidence from a large-scale benchmark

* How to Plant Trees in LMs: Data and Architectural Effects on the Emergence of Syntactic Inductive Biases

In episode 92 of The Gradient Podcast, Daniel Bashir speaks to Kevin K. Yang.

Kevin is a senior researcher at Microsoft Research (MSR) who works on problems at the intersection of machine learning and biology, with an emphasis on protein engineering. He completed his PhD at Caltech with Frances Arnold on applying machine learning to protein engineering. Before joining MSR, he was a machine learning scientist at Generate Biomedicines, where he used machine learning to optimize proteins.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:40) Kevin’s background

* (06:00) Protein engineering early in Kevin’s career

* (12:10) From research to real-world proteins: the process

* (17:40) Generative models + pretraining for proteins

* (22:47) Folding diffusion for protein structure generation

* (30:45) Protein evolutionary dynamics and generative models of protein sequences

* (40:03) Analogies and disanalogies between protein modeling and language models

* (41:45) In representation learning

* (45:50) Convolutions vs. transformers and inductive biases

* (49:25) Pretraining tasks for protein structure

* (51:45) More on representation learning for protein structure

* (54:06) Kevin’s thoughts on interpretability in deep learning for protein engineering

* (56:50) Multimodality in protein engineering and future directions

* (59:14) Outro

Links:

* Kevin’s Twitter and homepage

* Research

* Generative models + pre-training for proteins and chemistry

* Broad intro to techniques in the space

* Protein structure generation via folding diffusion

* Protein sequence design with deep generative models (review)

* Evolutionary velocity with protein language models predicts evolutionary dynamics of diverse proteins

* Protein generation with evolutionary diffusion: sequence is all you need

* ML for protein engineering

* ML-guided directed evolution for protein engineering (review)

* Learned protein embeddings for ML

* Adaptive machine learning for protein engineering (review)

* Multimodal deep learning for protein engineering

In episode 91 of The Gradient Podcast, Daniel Bashir speaks to Arjun Ramani and Zhengdong Wang.

Arjun is the global business and economics correspondent at The Economist.

Zhengdong is a research engineer at Google DeepMind.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (03:53) Arjun intro

* (06:04) Zhengdong intro

* (09:50) How Arjun and Zhengdong met in the woods

* (11:52) Overarching narratives about technological progress and AI

* (14:20) Setting up the claim: Arjun on what “transformative” means

* (15:52) What enables transformative economic growth?

* (21:19) From GPT-3 to ChatGPT; is there something special about AI?

* (24:15) Zhengdong on “real AI” and divisiveness

* (27:00) Arjun on the independence of bottlenecks to progress/growth

* (29:05) Zhengdong on bottleneck independence

* (32:45) More examples on bottlenecks and surplus wealth

* (37:06) Technical arguments—what are the hardest problems in AI?

* (38:00) Robotics

* (40:41) Challenges of deployment in high-stakes settings and data sources / synthetic data, self-driving

* (45:13) When synthetic data works

* (49:06) Harder tasks, process knowledge

* (51:45) Performance art as a critical bottleneck

* (53:45) Obligatory Taylor Swift Discourse

* (54:45) AI Taylor Swift???

* (54:50) The social arguments

* (55:20) Speed of technology diffusion — “diffusion lags” and dynamics of trust with AI

* (1:00:55) ChatGPT adoption, where major productivity gains come from

* (1:03:50) Timescales of transformation

* (1:10:22) Unpredictability in human affairs

* (1:14:07) The economic arguments

* (1:14:35) Key themes — diffusion lags, different sectors

* (1:21:15) More on bottlenecks, AI trust, premiums on human workers

* (1:22:30) Automated systems and human interaction

* (1:25:45) Campaign text reachouts

* (1:30:00) Counterarguments

* (1:30:18) Solving intelligence and solving science/innovation

* (1:34:07) Strengths and weaknesses of the broad applicability of Arjun and Zhengdong’s argument

* (1:35:34) The “proves too much” worry — how could any innovation have ever happened?

* (1:37:25) Examples of bringing down barriers to innovation/transformation

* (1:43:45) What to do with all of this information?

* (1:48:45) Outro

Links:

* Zhengdong’s homepage and Twitter

* Arjun’s homepage and Twitter

* Why transformative artificial intelligence is really, really hard to achieve

* Other resources and links mentioned:

* Allan-Feuer and Sanders: Transformative AGI by 2043 is <1% likely

* On AlphaStar Zero

* Hardmaru on AI as applied philosophy

* Robotics Transformer 2

* Davis Blalock on synthetic data

* Matt Clancy on automating invention and bottlenecks

* Michael Webb on 80,000 Hours Podcast

* Bob Gordon: The Rise and Fall of American Growth

* OpenAI economic impact paper

* David Autor: new work paper

* Baumol effect paper

* Pew research centre poll, public concern on AI

* Human premium Economist piece

* Callum Williams — London tube and AI/jobs

* Culture Series book 1, Iain Banks

In episode 90 of The Gradient Podcast, Daniel Bashir speaks to Miles Grimshaw.

Miles is General Partner at Benchmark. He was previously a General Partner at Thrive Capital, where he helped the firm raise its fourth and fifth funds, and sourced deals in Lattice, Mapbox, Benchling, and Airtable, among others.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:48) Miles’ background (note: Miles is now the second newest GP at Benchmark)

* (06:07) Miles’ investment philosophy and previous investments

* (12:25) Investing in the “decade of deep learning” and how Miles became interested in AI

* (18:53) Miles’ / Benchmark’s investment in Langchain

* (24:29) On AI advances and adoption

* (39:25) Hardware shortages, radically changing UX for LLMs

* (48:12) Opportunities for AI applications in new domains

* (50:15) Miles’ advice for potential founders in AI

* (1:00:00) Outro

Links:

* Miles’ Twitter

* Benchmark homepage

* LangChain homepage

In episode 89 of The Gradient Podcast, Daniel Bashir speaks to Shreya Shankar.

Shreya is a computer scientist pursuing her PhD in databases at UC Berkeley. Her research interest is in building end-to-end systems for people to develop production-grade machine learning applications. She was previously the first ML engineer at Viaduct, did research at Google Brain, and software engineering at Facebook. She graduated from Stanford with a B.S. and M.S. in computer science with concentrations in systems and artificial intelligence. At Stanford, helped run SHE++, an organization that helps empower underrepresented minorities in technology.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:22) Shreya’s background and journey into ML / MLOps

* (04:51) ML advances in 2013-2016

* (05:45) Shift in Stanford undergrad class ecosystems, accessibility of deep learning research

* (09:10) Why Shreya left her job as an ML engineer

* (13:30) How Shreya became interested in databases, data quality in ML

* (14:50) Daniel complains about things

* (16:00) What makes ML engineering uniquely difficult

* (16:50) Being a “historian of the craft” of ML engineering

* (22:25) Levels of abstraction, what ML engineers do/don’t have to think about

* (24:16) Observability for Production ML Pipelines

* (28:30) Metrics for real-time ML systems

* (31:20) Proposed solutions

* (34:00) Moving Fast with Broken Data

* (34:25) Existing data validation measures and where they fall short

* (36:31) Partition summarization for data validation

* (38:30) Small data and quantitative statistics for data cleaning

* (40:25) Streaming ML Evaluation

* (40:45) What makes a metric actionable

* (42:15) Differences in streaming ML vs. batch ML

* (45:45) Delayed and incomplete labels

* (49:23) Operationalizing Machine Learning

* (49:55) The difficult life of an ML engineer

* (53:00) Best practices, tools, pain points

* (55:56) Pitfalls in current MLOps tools

* (1:00:30) LLMOps / FMOps

* (1:07:10) Thoughts on ML Engineering, MLE through the lens of data engineering

* (1:10:42) Building products, user expectations for AI products

* (1:15:50) Outro

Links:

* Papers

* Towards Observability for Production Machine Learning Pipelines

* Rethinking Streaming ML Evaluation

* Operationalizing Machine Learning

* Moving Fast With Broken Data

* Blog posts

* The Modern ML Monitoring Mess

* Thoughts on ML Engineering After a Year of my PhD

In episode 88 of The Gradient Podcast, Daniel Bashir speaks to Professor Stevan Harnad.

Stevan Harnad is professor of psychology and cognitive science at Université du Québec à Montréal, adjunct professor of cognitive science at McGill University, and professor emeritus of cognitive science at the University of Southampton. His research is on category learning, categorical perception, symbol grounding, the evolution of language, and animal and human sentience (otherwise known as “consciousness”). He is also an advocate for open access and an activist for animal rights.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (05:20) Professor Harnad’s background: interests in cognitive psychobiology, editing Behavioral and Brain Sciences

* (07:40) John Searle submits the Chinese Room article

* (09:20) Early reactions to Searle and Prof. Harnad’s role

* (13:38) The core of Searle’s argument and the generator of the Symbol Grounding Problem, “strong AI”

* (19:00) Ways to ground symbols

* (20:26) The acquisition of categories

* (25:00) Pantomiming, non-linguistic category formation

* (27:45) Mathematics, abstraction, and grounding

* (36:20) Symbol manipulation and interpretation language

* (40:40) On the Whorf Hypothesis

* (48:39) Defining “grounding” and introducing the “T3” Turing Test

* (53:22) Turing’s concerns, AI and reverse-engineering cognition

* (59:25) Other Minds, T4 and zombies

* (1:05:48) Degrees of freedom in solutions to the Turing Test, the easy and hard problems of cognition

* (1:14:33) Over-interepretation of AI systems’ behavior, sentience concerns, T3 and evidence sentience

* (1:24:35) Prof. Harnad’s commentary on claims in The Vector Grounding Problem

* (1:28:05) RLHF and grounding, LLMs’ (ungrounded) capabilities, syntactic structure and propositions

* (1:35:30) Multimodal AI systems (image-text and robotic) and grounding, compositionality

* (1:42:50) Chomsky’s Universal Grammar, LLMs and T2

* (1:50:55) T3 and cognitive simulation

* (1:57:34) Outro

Links:

* Professor Harnad’s webpage and skywritings

* Papers:

* Category Induction and Representation

* Categorical Perception

* From Sensorimotor Categories to Grounded Symbols

* Minds, machines and Searle 2

* The Latent Structure of Dictionaries

In episode 87 of The Gradient Podcast, Daniel Bashir speaks to Professor Terry Winograd.

Professor Winograd is Professor Emeritus of Computer Science at Stanford University. His research focuses on human-computer interaction design and the design of technologies for development. He founded the Stanford Human-Computer Interaction Group, where he directed the teaching programs and HCI research. He is also a founding faculty member of the Stanford d.school and a founding member and past president of Computer Professionals for Social Responsibility.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (03:00) Professor Winograd’s background

* (05:10) At the MIT AI Lab

* (05:45) The atmosphere in the MIT AI Lab, Minsky/Chomsky debates

* (06:20) Blue-sky research, government funding for academic research

* (10:10) Isolation and collaboration between research groups

* (11:45) Phases in the development of ideas and how cross-disciplinary work fits in

* (12:26) SHRDLU and the MIT AI Lab’s intellectual roots

* (17:20) Early responses to SHRDLU: Minsky, Dreyfus, others

* (20:55) How Prof. Winograd’s thinking about AI’s abilities and limitations evolved

* (22:25) How this relates to current AI systems and discussions of intelligence

* (23:47) Repetitive debates in AI, semantics and grounding

* (27:00) The concept of investment, care, trust in human communication vs machine communication

* (28:53) Projecting human-ness onto AI systems and non-human things and what this means for society

* (31:30) Time after leaving MIT in 1973, time at Xerox PARC, how Winograd’s thinking evolved during this time

* (38:28) What Does It Mean to Understand Language? Speech acts, commitments, and the grounding of language

* (42:40) Reification of representations in science and ML

* (46:15) LLMs, their training processes, and their behavior

* (49:40) How do we coexist with systems that we don’t understand?

* (51:20) Progress narratives in AI and human agency

* (53:30) Transitioning to intelligence augmentation, founding the Stanford HCI group and d.school, advising Larry Page and Sergey Brin

* (1:01:25) Chatbots and how we consume information

* (1:06:52) Evolutions in journalism, progress in trust for modern AI systems

* (1:09:18) Shifts in the social contract, from institutions to personalities

* (1:12:05) AI and HCI in recent years

* (1:17:05) Philosophy of design and the d.school

* (1:21:20) Designing AI systems for people

* (1:25:10) Prof. Winograd’s perspective on watermarking for detecting GPT outputs

* (1:25:55) The politics of being a technologist

* (1:30:10) Echos of the past in AI regulation and competition and learning from history

* (1:32:34) Outro

Links:

* Professor Winograd’s Homepage

* Papers/topics discussed:

* SHRDLU

* Beyond Programming Languages

* What Does It Mean to Understand Language?

* The PageRank Citation Ranking

* Stanford Digital Libraries project

* Talk: My Politics as a Technologist

In episode 86 of The Gradient Podcast, Daniel Bashir speaks to Professor Gil Strang.

Professor Strang is one of the world’s foremost mathematics educators and a mathematician with contributions to finite element theory, the calculus of variations, wavelet analysis, and linear algebra. He has spent six decades teaching mathematics at MIT, where he was the MathWorks Professor of Mathematics. He was among the first MIT faculty members to publish a course on MIT’s OpenCourseware and has since championed both linear algebra education and open courseware.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:00) Professor Strang’s background and journey into teaching linear algebra

* (04:55) Undergrad interests

* (07:10) Writing textbooks

* (10:20) Prof. Strang’s interests in deep learning

* (11:00) How Professor Strang thought about teaching early on

* (16:20) MIT OpenCourseWare and education accessibility

* (19:50) Prof Strang’s applied/example-based approach to teaching linear algebra and closing the theory-practice gap

* (22:00) Examples!

* (27:20) Orthogonality

* (29:15) Singular values

* (34:40) Professor Strang’s favorite topics in linear algebra

* (37:55) Pedagogical approaches to deep learning, mathematical ingredients of deep learning’s complexity

* (42:04) Generalization and double descent in deep learning, powers and limitations

* (46:20) Did deep learning have to evolve as it did?

* (48:30) Teaching deep learning to younger students

* (50:50) How Prof. Strang’s approach to teaching linear algebra has evolved over time

* (53:00) The Four Fundamental Subspaces

* (56:15) Reflections on a career in teaching

* (59:49) Outro

Links:

* Professor Strang’s homepage

In episode 85 of The Gradient Podcast, Andrey Kurenkov speaks to Anant Agarwal

Anant Agarwal is the chief platform officer of 2U, and founder of edX. Anant taught the first edX course on circuits and electronics from MIT, which drew 155,000 students from 162 countries. He has served as the director of CSAIL, MIT's Computer Science and Artificial Intelligence Laboratory, and is a professor of electrical engineering and computer science at MIT. He is a successful serial entrepreneur, having co-founded several companies including Tilera Corporation, which created the Tile multicore processor, and Virtual Machine Works.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:30) History with research

* (05:56) Founding EdX

* (13:05) AI at EdX

* (18:40) Reaction to AI as a teacher

* (25:00) Student interest in AI

* (32:20) AI’s impact on academia

* (35:00) Future of AI in education

* (38:25) AI writing essays

* (43:38) Experiences playing with ChatGPT

In episode 84 of The Gradient Podcast, Daniel Bashir speaks to Professor Raphaël Millière.

Professor Millière is a Lecturer (Assistant Professor) in the Philosophy of Artificial Intelligence at Macquarie University in Sydney, Australia. Previously, he was the 2020 Robert A. Burt Presidential Scholar in Society and Neuroscience in Columbia University’s Center for Science and Society, and completed his DPhil in philosophy at the University of Oxford, where he focused on self-consciousness.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:20) Prof. Millière’s background

* (08:07) AI + philosophy questions and the human side / empiricism

* (18:38) Putting aside metaphysical issues

* (20:28) Prof. Millière’s work on self-consciousness, does consciousness constitutively involve self-consciousness?

* (32:05) Relationship to recent pronouncements about AI sentience

* (41:54) Chatbots’ self-presentation as having a “self”

* (51:05) Intro to grounding and related concepts

* (1:00:06) The different types of grounding

* (1:08:48) Lexical representations and things in the world, distributional hypothesis, concepts in LLMs

* (1:21:40) Representational content and overcoming the vector grounding problem

* (1:32:01) Causal-informational relations and teleology

* (1:43:45) Levels of grounding, extralinguistic aspects of meaning

* (1:52:12) Future problems and ongoing projects

* (2:04:05) Outro

Links:

* Professor Millière’s homepage and Twitter

* Research

* Are There Degrees of Self-Consciousness?

* The Varieties of Selflessness

* Selfless Memories

* The Vector Grounding Problem

In episode 83 of The Gradient Podcast, Daniel Bashir speaks to Peli Grietzer.

Peli is a scholar whose work borrows mathematical ideas from machine learning theory to think through “ambient” and ineffable phenomena like moods, vibes, cultural logics, and structures of feeling. He is working on a book titled Big Mood: A Transcendental-Computational Essay in Art and contributes to the experimental literature collective Gauss PDF. Peli has a PhD in mathematically informed literary theory from Harvard Comparative Literature in collaboration with the HUJI Einstein Institute of Mathematics.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:17) Peli’s background

* (10:40) Daniel takes 2 entire minutes to ask how Peli thinks about ~ Art ~

* (26:10) Idealism and art as revealing the nature of reality, extralinguistic experiences of truth through literature

* (52:05) The autoencoder as a way to understand Romantic theories of art

* (1:14:55) More on how Peli thinks about autoencoders

* (1:18:05) Connections to ambient meaning, stimmung/mood

* (1:37:18) Examples of poetry/literature as mathematical experience, aesthetic unity and totalizing worldviews

* (1:51:15) Moods clashing within a single work

* (2:10:14) Modernist writers

* (2:32:46) Outro

Links:

* Peli’s Twitter

* A Theory of Vibe

* Why poetry is a variety of mathematical experience

* Peli’s thesis

In episode 82 of The Gradient Podcast, Daniel Bashir speaks to Ryan Drapeau.

Ryan is a Staff Software Engineer at Stripe and technical lead for Stripe’s Payment Fraud organization, which uses machine learning to help prevent billions of dollars of credit card and payments fraud for business every year.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:15) Ryan’s background

* (05:28) Differences between adversarial problems (fraud, content moderation, etc.)

* (08:50) How fraud manifests for businesses

* (11:07) Types of fraud

* (15:49) Fraud as an industry

* (19:05) Information asymmetries between fraudsters and defenders

* (22:40) Fraud as an ML problem and Stripe Radar

* (25:45) Evolution of Stripe Radar

* (31:38) Architectural evolution

* (41:38) Why ResNets for Stripe Radar?

* (44:15) Future architectures for Stripe Radar and the explainability/performance tradeoff

* (48:58) War stories

* (52:55) Federated learning opportunities for Stripe Radar

* (55:50) Vectors for improvement in Stripe’s fraud detection systems

* (59:22) More ways of thinking about the fraud problem, multiclass models

* (1:03:30) Lessons Ryan has picked up from working on fraud

* (1:05:44) Outro

Links:

* How We Built It: Stripe Radar

* Stripe 2022 Update

In episode 81 of The Gradient Podcast, Daniel Bashir speaks to Shiv Rao.

Shiv Rao, MD is the co-founder and CEO of Abridge, a healthcare conversation company that uses cutting-edge NLP and generative AI to bring context and understanding to every medical conversation. Shiv previously served as an Executive Vice President at UPMC Enterprises, managing the provider-facing portfolio of technology investments and R&D. He is a practicing cardiologist in UPMC's Heart and Vascular Institute.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:34) Shiv’s medicine/technology/VC background

* (05:45) Difficulties for tech in healthcare and how this informs Shiv’s approach

* (10:52) “Productivity with a smile” and how AI can make medicine feel more human

* (12:35) Shiv’s experiences in medicine and how Abridge’s product helps doctors

* (16:10) How the role of a clinical team could evolve

* (19:30) Abridge’s partnerships and real-life use cases

* (23:00) Shiv’s perspectives on concerns about bias/trust/privacy

* (25:25) Clinical decision support vs “automating doctors”

* (29:07) Transparency and Abridge’s user experience

* (35:20) Algorithmic solutionism vs human-focused approaches to technology development

* (38:50) Ways AI might impact healthcare

* (41:10) Generative AI applications

* (45:00) Generative AI opportunities beyond documentation

* (49:25) Innovation and reducing friction, UX

* (50:56) Why people make wild predictions about AI

* (54:25) What it means to “automate away” a doctor, how we’re misusing the medical workforce

* (56:10) Shiv’s advice for people interested in AI + healthcare

* (1:00:04) Outro

Links:

* Abridge Homepage

In episode 80 of The Gradient Podcast, Daniel Bashir speaks to Professor Hugo Larochelle.

Professor Larochelle leads the Montreal Google DeepMind team and is adjunct professor at Université de Montréal and a Canada CIFAR Chair. His research focuses on the study and development of deep learning algorithms.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:38) Prof. Larochelle’s background, working in Bengio’s lab

* (04:53) Prof. Larochelle’s work and connectionism

* (08:20) 2004-2009, work with Bengio

* (08:40) Nonlocal Estimation of Manifold Structure, manifolds and deep learning

* (13:58) Manifold learning in vision and language

* (16:00) Relationship to Denoising Autoencoders and greedy layer-wise pretraining

* (21:00) From input copying to learning about local distribution structure

* (22:30) Zero-Data Learning of New Tasks

* (22:45) The phrase “extend machine learning towards AI” and terminology

* (26:55) Prescient hints of prompt engineering

* (29:10) Daniel goes on totally unnecessary tangent

* (30:00) Methods for training deep networks (strategies and robust interdependent codes)

* (33:45) Motivations for layer-wise pretraining

* (35:15) Robust Interdependent Codes and interactions between neurons in a single network layer

* (39:00) 2009-2011, postdoc in Geoff Hinton’s lab

* (40:00) Reflections on the AlexNet moment

* (41:45) Frustration with methods for evaluating unsupervised methods, NADE

* (44:45) How researchers thought about representation learning, toying with objectives instead of architectures

* (47:40) The Restricted Boltzmann Forest

* (50:45) Imposing structure for tractable learning of distributions

* (53:11) 2011-2016 at U Sherbooke (and Twitter)

* (53:45) How Prof. Larochelle approached research problems

* (56:00) How Domain Adversarial Networks came about

* (57:12) Can we still learn from Restricted Boltzmann Machines?

* (1:02:20) The ~ Infinite ~ Restricted Boltzmann Machine

* (1:06:55) The need for researchers doing different sorts of work

* (1:08:58) 2017-present, at MILA (and Google)

* (1:09:30) Modulating Early Visual Processing by Language, neuroscientific inspiration

* (1:13:22) Representation learning and generalization, what is a good representation (Meta-Dataset, Universal representation transformer layer, universal template, Head2Toe)

* (1:15:10) Meta-Dataset motivation

* (1:18:00) Shifting focus to the problem—good practices for “recycling deep learning”

* (1:19:15) Head2Toe intuitions

* (1:21:40) What are “universal representations” and manifold perspective on datasets, what is the right pretraining dataset

* (1:26:02) Prof. Larochelle’s takeaways from Fortuitous Forgetting in Connectionist Networks (led by Hattie Zhou)

* (1:32:15) Obligatory commentary on The Present Moment and current directions in ML

* (1:36:18) The creation and motivations of the TMLR journal

* (1:41:48) Prof. Larochelle’s takeaways about doing good science, building research groups, and nurturing a research environment

* (1:44:05) Prof. Larochelle’s advice for aspiring researchers today

* (1:47:41) Outro

Links:

* Professor Larochelle’s homepage and Twitter

* Transactions on Machine Learning Research

* Papers

* 2004-2009

* Nonlocal Estimation of Manifold Structure

* Classification using Discriminative Restricted Boltzmann Machines

* Zero-data learning of new tasks

* Exploring Strategies for Training Deep Neural Networks

* Deep Learning using Robust Interdependent Codes

* 2009-2011

* Stacked Denoising Autoencoders

* Tractable multivariate binary density estimation and the restricted Boltzmann forest

* The Neural Autoregressive Distribution Estimator

* Learning Attentional Policies for Tracking and Recognition in Video with Deep Networks

* 2011-2016

* Practical Bayesian Optimization of Machine Learning Algorithms

* Learning Algorithms for the Classification Restricted Boltzmann Machine

* A neural autoregressive topic model

* Domain-Adversarial Training of Neural Networks

* NADE

* An Infinite Restricted Boltzmann Machine

* 2017-present

* Modulating early visual processing by language

* Meta-Dataset

* A Universal Representation Transformer Layer for Few-Shot Image Classification

* Learning a universal template for few-shot dataset generalization

* Impact of aliasing on generalization in deep convolutional networks

* Head2Toe: Utilizing Intermediate Representations for Better Transfer Learning

* Fortuitous Forgetting in Connectionist Networks

In episode 79 of The Gradient Podcast, Daniel Bashir speaks to Jeremie Harris.

Jeremie is co-founder of Gladstone AI, author of the book Quantum Physics Made Me Do It, and co-host of the Last Week in AI Podcast. Jeremy previously hosted the Towards Data Science podcast and worked on a number of other startups after leaving a PhD in physics.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:37) Jeremie’s physics background and transition to ML

* (05:19) The physicist-to-AI person pipeline, how Jeremie’s background impacts his approach to AI

* (08:20) A tangent on inflationism/deflationism about natural laws (I promise this applies to AI)

* (11:45) How ML implies a particular viewpoint on the above question

* (13:20) Jeremie’s first (recommendation systems) company, how startup founders can make mistakes even when they’ve read Paul Graham essays

* (17:30) Classic startup wisdom, different sorts of startups

* (19:35) OpenAI’s approach in shipping features for DALL-E 2 and generation vs. discrimination as an approach to product

* (24:55) Capabilities and risk

* (26:43) Commentary on fundamental limitations of alignment in LLMs

* (30:45) Intrinsic difficulties in alignment problems

* (41:15) Daniel tries to steel man / defend anti-longtermist arguments (nicely :) )

* (46:23) Anthropic’s paper on asking models to be less biased

* (47:20) Why Jeremie is excited about Anthropic’s Constitutional AI scheme

* (51:05) Jeremie’s thoughts on recent Eliezer discourse

* (56:50) Cheese / task vectors and steerability/controllability in LLMs

* (59:50) Difficulty of one-shot solutions in alignment work, better strategies

* (1:02:00) Lack of theoretical understanding of deep learning systems / alignment

* (1:04:50) Jeremie’s work and perspectives on AI policy

* (1:10:00) Incrementality in convincing policymakers

* (1:14:00) How recent developments impact policy efforts

* (1:16:20) Benefits and drawbacks of open source

* (1:19:30) Arguments in favor of (limited) open source

* (1:20:35) Quantum Physics (not Mechanics) Made Me Do It

* (1:24:10) Some theories of consciousness and corresponding physics

* (1:29:49) Outro

Links:

* Jeremie’s Twitter

* Quantum Physics Made Me Do It

* Gladstone AI

In episode 78 of The Gradient Podcast, Daniel Bashir speaks to Antoine Blondeau.

Antoine is a serial AI entrepreneur and Co-Founder and Managing Partner of Alpha Intelligence Capital. He was chief executive at Dejima when the firm worked on CALO, one of the biggest AI projects in US history and precursor to Apple’s Siri. Later, he co-founded Sentient Technologies, which boasted the title of world’s highest funded AI company in 2016. In 2018, he founded Alpha Intelligence Capital to support future AI unicorns, and has raised more than $300 million, which has been deployed into 31 companies.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:30) Antoine’s background

* (04:00) Dejima and the CALO cognitive assistant (the precursor to Siri)

* (07:35) More detail on CALO

* (10:10) Sentient Technologies and entrepreneurship during the AlexNet moment

* (14:35) Early predictions on scale

* (17:15) Role of evolutionary computation and neuroevolution

* (20:00) Antoine’s motivations for becoming an investor

* (22:30) Alpha Intelligence Capital’s investment focus

* (27:40) Safety and trust issues in fully automated systems

* (37:00) Models of culture, discernment in the use of AI systems

* (39:30) Antoine’s experience as an investor in today’s AI environment

* (44:43) How modern LLMs impact standard advice regarding the appropriateness of cutting-edge technologies in business

* (49:00) Data (and other) moats

* (52:07) Application/research areas Antoine is watching

* (55:00) Antoine’s advice for people watching AI’s current developments

* (58:47) Outro

Links:

* Alpha Intelligence Capital Homepage

In episode 77 of The Gradient Podcast, Daniel Bashir speaks to Joon Park.

Joon is a third-year PhD student at Stanford, advised by Professors Michael Bernstein and Percy Liang. He designs, builds, and evaluates interactive systems that support new forms of human-computer interaction by leveraging state-of-the-art advances in natural language processing such as large language models. His research introduced the concept of, and the techniques for building generative agents—computational software agents that simulate believable human behavior. Joon’s work has been supported by the Microsoft Research PhD Fellowship, the Stanford School of Engineering Fellowship, and the Siebel Scholarship.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:43) Joon’s path from studio art to social computing / AI

* (05:00) Joon’s perspectives on Human-Computer Interaction (HCI) and its recent evolution

* (06:45) How foundation models enter the picture

* (10:28) On slow algorithms and technology: A Slow Algorithm Improves Users’ Assessments of the Algorithm’s Accuracy

* (12:10) Motivations

* (17:55) The jellybean-counting task, hypotheses

* (22:00) Applications and takeaways

* (28:05) Deliberate engagement in social media / computing systems, incentives

* (32:55) Daniel rants about The Social Dilemma + anti- social media rhetoric, Joon on the role of academics, framings of addiction

* (39:05) Measuring the Prevalence of Anti-Social Behavior in Online Communities

* (48:30) Statistics on anti-social behavior and anecdotal information, limitations in the paper’s measurements

* (51:45) Participatory and value-sensitive design

* (52:50) “Interaction” in On the Opportunities and Risks of Foundation Models

* (53:45) Broader insights on foundation models and emergent behavior

* (56:50) Joon’s section on interaction

* (1:01:05) Daniel’s bad segue to Social Simulacra: Creating Populated Prototypes for Social Computing Systems

* (1:02:50) Context for Social Simulacra and Generative Agents, why Social Simulacra was tackled first

* (1:24:05) The value of norms

* (1:26:20) Collaborations between designers and developers of social simulacra

* (1:30:00) Generative Agents: Interactive Simulacra of Human Behavior

* (1:30:30) Context / intro

* (1:45:10) On (too much) coherence in generative agents and believability

* (1:52:02) Instruction tuning’s impact on generative agents, model alignment w/ believability goals, desirability of agent conflict / toxic LLMs

* (1:56:55) Release strategies and toxicity in LLMs

* (2:03:05) On designing interfaces and responsible use

* (2:09:05) Capability advances and the capability-safety research gap

* (2:14:12) Worries about LLM integration, human-centered framework for technology release / LLM incorporation

* (2:18:00) Joon’s philosophy as an HCI researcher

* (2:20:39) Outro

Links:

* Joon’s homepage and Twitter

* Research

* A Slow Algorithm Improves Users’ Assessments of the Algorithm’s Accuracy

* Measuring the Prevalence of Anti-Social Behavior in Online Communities

* On the Opportunities and Risks of Foundation Models

* Social Simulacra: Creating Populated Prototypes for Social Computing Systems

* Generative Agents: Interactive Simulacra of Human Behavior

In episode 76 of The Gradient Podcast, Andrey Kurenkov speaks to Dr Christoffer Holmgård

Dr. Holmgård is a co-founder and the CEO of Modl.ai, which is building AI Engine for game development. Before starting the company, Christoffer was director of the indie game studio Die Gute Fabrik (which is German for "The Good Factory"), and has also done extensive research as an assistant professor in AI and Machine Learning for Games at Northeastern University. 

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

(00:00) Intro(01:30) History with video games (06:30) History with AI(09:40) Modeling stress responses in virtual environments(13:30) Play style personas from empirical data(17:15) Automating video game testing(21:00) Video game development(28:15) modl.ai(33:45) Automated playtesting with procedural personas through MCTS with evolved heuristics(35:40) Thoughts on recent AI progress(40:50) RL for game testing(44:40) AI in Minecraft(47:50) Impact of AI on video game development(01:00:00) Ethics of Gen AI (01:06:20) Hobbies / Interests (01:08:30) Outro

In episode 75 of The Gradient Podcast, Daniel Bashir speaks to Riley Goodside.

Riley is a Staff Prompt Engineer at Scale AI. Riley began posting GPT-3 prompt examples and screenshot demonstrations in 2022. He previously worked as a data scientist at OkCupid, Grindr, and CopyAI.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:37) Riley’s journey to becoming the first Staff Prompt Enginer

* (02:00) data science background in online dating industry

* (02:15) Sabbatical + catching up on LLM progress

* (04:00) AI Dungeon and first taste of GPT-3

* (05:10) Developing on codex, ideas about integrating codex with Jupyter Notebooks, start of posting on Twitter

* (08:30) “LLM ethnography”

* (09:12) The history of prompt engineering: in-context learning, Reinforcement Learning from Human Feedback (RLHF)

* (10:20) Models used to be harder to talk to

* (10:45) The three eras

* (10:45) 1 - Pre-trained LM era—simple next-word predictors

* (12:54) 2 - Instruction tuning

* (16:13) 3 - RLHF and overcoming instruction tuning’s limitations

* (19:24) Prompting as subtractive sculpting, prompting and AI safety

* (21:17) Riley on RLHF and safety

* (24:55) Riley’s most interesting experiments and observations

* (25:50) Mode collapse in RLHF models

* (29:24) Prompting models with very long instructions

* (33:13) Explorations with regular expressions, chain-of-thought prompting styles

* (36:32) Theories of in-context learning and prompting, why certain prompts work well

* (42:20) Riley’s advice for writing better prompts

* (49:02) Debates over prompt engineering as a career, relevance of prompt engineers

* (58:55) Outro

Links:

* Riley’s Twitter and LinkedIn

* Talk: LLM Prompt Engineering and RLHF: History and Techniques

In episode 74 of The Gradient Podcast, Daniel Bashir speaks to Professor Talia Ringer.

Professor Ringer is an Assistant Professor with the Programming Languages, Formal Methods, and Software Engineering group at the University of Illinois at Urbana Champaign. Their research leverages proof engineering to allow programmers to more easily build formally verified software systems.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Daniel’s long annoying intro

* (02:15) Origin Story

* (04:30) Why / when formal verification is important

* (06:40) Concerns about ChatGPT/AutoGPT et al failures, systems for accountability

* (08:20) Difficulties in making formal verification accessible

* (11:45) Tactics and interactive theorem provers, interface issues

* (13:25) How Prof Ringer’s research first crossed paths with ML

* (16:00) Concrete problems in proof automation

* (16:15) How ML can help people verifying software systems

* (20:05) Using LLMs for understanding / reasoning about code

* (23:05) Going from tests / formal properties to code

* (31:30) Is deep learning the right paradigm for dealing with relations for theorem proving?

* (36:50) Architectural innovations, neuro-symbolic systems

* (40:00) Hazy definitions in ML

* (41:50) Baldur: Proof Generation & Repair with LLMs

* (45:55) In-context learning’s effectiveness for LLM-based theorem proving

* (47:12) LLMs without fine-tuning for proofs

* (48:45) Something ~ surprising ~ about Baldur results (maybe clickbait or maybe not)

* (49:32) Asking models to construct proofs with restrictions, translating proofs to formal proofs

* (52:07) Methods of proofs and relative difficulties

* (57:45) Verifying / providing formal guarantees on ML systems

* (1:01:15) Verifying input-output behavior and basic considerations, nature of guarantees

* (1:05:20) Certified/verifies systems vs certifying/verifying systems—getting LLMs to spit out proofs along with code

* (1:07:15) Interpretability and how much model internals matter, RLHF, mechanistic interpretability

* (1:13:50) Levels of verification for deploying ML systems, HCI problems

* (1:17:30) People (Talia) actually use Bard

* (1:20:00) Dual-use and “correct behavior”

* (1:24:30) Good uses of jailbreaking

* (1:26:30) Talia’s views on evil AI / AI safety concerns

* (1:32:00) Issues with talking about “intelligence,” assumptions about what “general intelligence” means

* (1:34:20) Difficulty in having grounded conversations about capabilities, transparency

* (1:39:20) Great quotation to steal for your next thinkpiece + intelligence as socially defined

* (1:42:45) Exciting research directions

* (1:44:48) Outro

Links:

* Talia’s Twitter and homepage

* Research

* Concrete Problems in Proof Automation

* Baldur: Whole-Proof Generation and Repair with LLMs

* Research ideas

In episode 72 of The Gradient Podcast, Daniel Bashir speaks to Brigham Hyde.

Brigham is Co-Founder and CEO of Atropos Health. Prior to Atropos, he served as President of Data and Analytics at Eversana, a life sciences commercialization service provider. He led the investment in Concert AI in the oncology real-world data space at Symphony AI. Brigham has also held research faculty positions at Tufts University and the MIT Media Lab.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:55) Brigham’s background

* (06:00) Current challenges in healthcare

* (12:33) Interpretablity and delivering positive patient outcomes

* (17:10) How Atropos surfaces relevant data for patient interventions, on personalized observational research studies

* (22:10) Quality and quantity of data for patient interventions

* (27:25) Challenges and opportunities for generative AI in healthcare

* (35:17) Database augmentation for generative models

* (36:25) Future work for Atropos

* (39:15) Future directions for AI + healthcare

* (40:56) Outro

Links:

* Atropos Health homepage

* Brigham’s Twitter and LinkedIn

In episode 72 of The Gradient Podcast, Daniel Bashir speaks to Professor Scott Aaronson.

Scott is the Schlumberger Centennial Chair of Computer Science at the University of Texas at Austin and director of its Quantum Information Center. His research interests focus on the capabilities and limits of quantum computers and computational complexity theory more broadly. He has recently been on leave to work at OpenAI, where he is researching theoretical foundations of AI safety. 

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:45) Scott’s background

* (02:50) Starting grad school in AI, transitioning to quantum computing and the AI / quantum computing intersection

* (05:30) Where quantum computers can give us exponential speedups, simulation overhead, Grover’s algorithm

* (10:50) Overselling of quantum computing applied to AI, Scott’s analysis on quantum machine learning

* (18:45) ML problems that involve quantum mechanics and Scott’s work

* (21:50) Scott’s recent work at OpenAI

* (22:30) Why Scott was skeptical of AI alignment work early on

* (26:30) Unexpected improvements in modern AI and Scott’s belief update

* (32:30) Preliminary Analysis of DALL-E 2 (Marcus & Davis)

* (34:15) Watermarking GPT outputs

* (41:00) Motivations for watermarking and language model detection

* (45:00) Ways around watermarking

* (46:40) Other aspects of Scott’s experience with OpenAI, theoretical problems

* (49:10) Thoughts on definitions for humanistic concepts in AI

* (58:45) Scott’s “reform AI alignment stance” and Eliezer Yudkowsky’s recent comments (+ Daniel pronounces Eliezer wrong), orthogonality thesis, cases for stopping scaling

* (1:08:45) Outro

Links:

* Scott’s blog

* AI-related work

* Quantum Machine Learning Algorithms: Read the Fine Print

* A very preliminary analysis of DALL-E 2 w/ Marcus and Davis

* New AI classifier for indicating AI-written text and Watermarking GPT Outputs

* Writing

* Should GPT exist?

* AI Safety Lecture

* Why I’m not terrified of AI

In episode 71 of The Gradient Podcast, Daniel Bashir speaks to Ted Underwood.

Ted is a professor in the School of Information Sciences with an appointment in the Department of English at the University of Illinois at Urbana Champaign. Trained in English literary history, he turned his research focus to applying machine learning to large digital collections. His work explores literary patterns that become visible across long timelines when we consider many works at once—often, his work involves correcting and enriching digital collections to make them more amenable to interesting literary research.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (01:42) Ted’s background / origin story

* (04:35) Context in interpreting statistics, “you need a model,” the need for data about human responses to literature and how that manifested in Ted’s work

* (07:25) The recognition that we can model literary prestige/genre because of ML

* (08:30) Distant reading and the import of statistics over large digital libraries

* (12:00) Literary prestige

* (12:45) How predictable is fiction? Scales of predictability in texts

* (13:55) Degrees of autocorrelation in biography and fiction and the structure of narrative, how LMs might offer more sophisticated analysis

* (15:15) Braided suspense / suspense at different scales of a story

* (17:05) The Literary Uses of High-Dimensional Space: how “big data” came to impact the humanities, skepticism from humanists and responses, what you can do with word count

* (20:50) Why we could use more time to digest statistical ML—how acceleration in AI advances might impact pedagogy

* (22:30) The value in explicit models

* (23:30) Poetic “revolutions” and literary prestige

* (25:53) Distant vs. close reading in poetry—follow-up work for “The Longue Durée”

* (28:20) Sophistication of NLP and approaching the human experience

* (29:20) What about poetry renders it prestigious?

* (32:20) Individualism/liberalism and evolution of poetic taste

* (33:20) Why there is resistance to quantitative approaches to literature

* (34:00) Fiction in other languages

* (37:33) The Life Cycles of Genres

* (38:00) The concept of “genre”

* (41:00) Inflationary/deflationary views on natural kinds and genre

* (44:20) Genre as a social and not a linguistic phenomenon

* (46:10) Will causal models impact the humanities?

* (48:30) (Ir)reducibility of cultural influences on authors

* (50:00) Machine Learning and Human Perspective

* (50:20) Fluent and perspectival categories—Miriam Posner on “the radical, unrealized potential of digital humanities.”

* (52:52) How ML’s vices can become virtues for humanists

* (56:05) Can We Map Culture? and The Historical Significance of Textual Distances

* (56:50) Are cultures and other social phenomena related to one another in a way we can “map”?

* (59:00) Is cultural distance Euclidean?

* (59:45) The KL Divergence’s use for humanists

* (1:03:32) We don’t already understand the broad outlines of literary history

* (1:06:55) Science Fiction Hasn’t Prepared us to Imagine Machine Learning

* (1:08:45) The latent space of language and what intelligence could mean

* (1:09:30) LLMs as models of culture

* (1:10:00) What it is to be a human in “the age of AI” and Ezra Klein’s framing

* (1:12:45) Mapping the Latent Spaces of Culture

* (1:13:10) Ted on Stochastic Parrots

* (1:15:55) The risk of AI enabling hermetically sealed cultures

* (1:17:55) “Postcards from an unmapped latent space,” more on AI systems’ limitations as virtues

* (1:20:40) Obligatory GPT-4 section

* (1:21:00) Using GPT-4 to estimate passage of time in fiction

* (1:23:39) Is deep learning more interpretable than statistical NLP?

* (1:25:17) The “self-reports” of language models: should we trust them?

* (1:26:50) University dependence on tech giants, open-source models

* (1:31:55) Reclaiming Ground for the Humanities

* (1:32:25) What scientists, alone, can contribute to the humanities

* (1:34:45) On the future of the humanities

* (1:35:55) How computing can enable humanists as humanists

* (1:37:05) Human self-understanding as a collaborative project

* (1:39:30) Is anything ineffable? On what AI systems can “grasp”

* (1:43:12) Outro

Links:

* Ted’s blog and Twitter

* Research

* The literary uses of high-dimensional space

* The Longue Durée of literary prestige

* The Historical Significance of Textual Distances

* Machine Learning and Human Perspective

* The life cycles of genres

* Can We Map Culture?

* Cohort Succession Explains Most Change in Literary Culture

* Other Writing

* Reclaiming Ground for the Humanities

* We don’t already understand the broad outlines of literary history

* Science fiction hasn’t prepared us to imagine machine learning.

* How predictable is fiction?

* Mapping the latent spaces of culture

* Using GPT-4 to measure the passage of time in fiction

In episode 70 of The Gradient Podcast, Daniel Bashir speaks to Irene Solaiman.

Irene is an expert in AI safety and policy and the Policy Director at HuggingFace, where she conducts social impact research and develops public policy. In her former role at OpenAI, she initiated and led bias and social impact research at OpenAI in addition to leading public policy. She built AI policy at Zillow group and advised poilcymakers on responsible autonomous decision-making and privacy as a fellow at Harvard’s Berkman Klein Center.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:00) Intro to Irene and her work

* (03:45) What tech people need to learn about policy, and vice versa

* (06:35) Societal impact—words and reality, Irene’s experience

* (08:30) OpenAI work on GPT-2 and release strategies (yes, this was recorded on Pi Day)

* (11:00) Open-source proponents and release

* (14:00) What does a multidisciplinary approach to working on AI look like?

* (16:30) Thinking about end users and enabling contributors with different sets of expertise

* (18:00) “Preparing for AGI” and current approaches to release

* (21:00) Who constitutes a researcher? What constitutes safety and who gets resourced? Limitations in red-teaming potentially dangerous systems.

* (22:35) PALMS and Values-Targeted Datasets

* (25:52) PALMS and RLHF

* (27:00) Homogenization in foundation models, cultural contexts

* (29:45) Anthropic’s moral self-correction paper and Irene’s concerns about marketing “de-biasing” and oversimplification

* (31:50) Data work, human systemic problems → AI bias

* (33:55) Why do language models get more toxic as they get larger? (if you have ideas, let us know!)

* (35:45) The gradient of generative AI release, Irene’s experience with the open-source world, tradeoffs along the release gradient

* (38:40) More on Irene’s orientation towards release

* (39:40) Pragmatics of keeping models closed, dealing with open-source by force

* (42:22) Norm setting for release and use, normalization of documentation on social impacts

* (46:30) Race dynamics :(

* (49:45) Resource allocation and advances in ethics/policy, conversations on integrity and disinformation

* (53:10) Organizational goals, balancing technical research with policy work

* (58:10) Thoughts on governments’ AI policies, impact of structural assumptions

* (1:04:00) Approaches to AI-generated sexual content, need for more voices represented in conversations about AI

* (1:08:25) Irene’s suggestions for AI practitioners / technologists

* (1:11:24) Outro

Links:

* Irene’s homepage and Twitter

* Papers

* Release Strategies and the Social Impacts of Language Models

* Hugh Zhang’s open letter in The Gradient from 2019

* Process for Adapting Large Models to Society (PALMS) with Values-Targeted Datasets

* The Gradient of Generative AI Release: Methods and Considerations

In episode 69 of The Gradient Podcast, Daniel Bashir speaks to Drago Anguelov.

Drago is currently a Distinguished Scientist and Head of Research at Waymo, where he joined in 2018. Earlier, he spent eight years at Google working on 3D vision and pose estimation for StreetView, then leading a research team that developed computer vision systems for annotating Google Photos. He has been involved in developing popular neural network methods such as the Inception architecture and the SSD detector. Before joining Waymo, he also led the 3D perception team at Zoox.

Have suggestions for future podcast guests (or other feedback)? Let us know here or reach us at [email protected]

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

Outline:

* (00:00) Intro

* (02:04) Drago’s background in AI and self-driving, work with Daphne Koller + Sebastian Thrun, computer vision / pose estimation

* (14:20) One- and two-stage object detectors

* (15:15) Early experiences and thoughts on self-driving and its prospects

* (21:00) An introduction to the “self-driving stack”: mapping & localization, perception, behavior modeling & planning, simulation

* (29:25) From Stuart Russell’s comments on early Waymo’s “old-fashioned” approach

* (37:34) Scaling 3D Detection: challenges and architectural innovations

* (43:20) Behavior modeling: making decisions and modeling interactions in multi-agent environments

* (52:42) Distributional RL (+ imitation learning) in self-driving?

* (54:10) The Waymo Open Dataset

* (1:01:48) Looking forward in self-driving

* (1:04:36) Outro

Links:

* Drago’s LinkedIn and Twitter

* Research

* SSD: Single-Shot Multibox Detector

* SCAPE: Shape completion and animation of people

* Behavior Models for Autonomous Driving

* Wayformer

* Symphony: Learning Realistic and Diverse Agents for Autonomous Driving Simulation

* Imitation Is Not Enough

* Scaling 3D Detection to the Long Tail

In episode 68 of The Gradient Podcast, Daniel Bashir speaks to Professor Joanna Bryson.

Professor Bryson is Professor of Ethics and Technology at the Hertie School, where her research focuses on the impact of technology on human cooperation and AI/ICT governance. Professor Bryson has advised companies, governments, transnational agencies, and NGOs, particularly in AI policy. She is one of the few people doing this sort of work who actually has a PhD and work experience in AI, but also advanced degrees in the social sciences. She started her academic career though in the liberal arts, and publishes regularly in the natural sciences.

Have suggestions for future podcast guests (or other feedback)? Let us know here!

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

Outline:

* (00:00) Intro

* (01:35) Intro to Professor Bryson’s work

* (06:37) Shifts in backgrounds expected of AI PhDs/researchers

* (09:40) Masters’ degree in Edinburgh, Behavior-Based AI

* (11:00) PhD, differences between MIT’s engineering focus and Edinburgh, systems engineering + AI

* (16:15) Comments on ways you can make contributions in AI

* (18:45) When definitions of “intelligence” are important

* (24:23) Non- and proto-linguistic aspects of intelligence, arguments about text as a description of human experience

* (31:45) Cognitive leaps in interacting with language models

* (37:00) Feelings of affiliation for robots, phenomenological experience in humans and (not) in AI systems

* (42:00) Language models and technological systems as cultural artifacts, expressing agency through machines

* (44:15) Capabilities development and moral patient status in AI systems

* (51:20) Prof. Bryson’s perspectives on recent AI regulation

* (1:00:55) Responsibility and recourse, Uber self-driving crash

* (1:07:30) “Preparing for AGI,” “Living with AGI,” how to respond to recent AI developments

* (1:12:18) Outro

Links:

* Professor Bryson’s homepage and Twitter

* Papers

* Systems AI

* Behavior Oriented Design, action selection, key differences in methodology/views between systems AI researchers and e.g. connectionists

* Agent architecture as object oriented design (1998)

* Intelligence by design: Principles of modularity and coordination for engineering complex adaptive agents (2001)

* Cognition

* Age-Related Inhibition and Learning Effects: Evidence from Transitive Performance (2013)

* Primate Errors in Transitive ‘Inference’: A Two-Tier Learning Model (2007)

* Skill Acquisition Through Program-Level Imitation in a Real-Time Domain

* Agent-Based Models as Scientific Methodology: A Case Study Analysing Primate Social Behaviour (2008, 2011)

* Social learning in a non-social reptile (Geochelone carbonaria) (2010)

* Understanding and Addressing Cultural Variation in Costly Antisocial Punishment (2014)

* Polarization Under Rising Inequality and Economic Decline (2020)

* Semantics derived automatically from language corpora contain human-like biases (2017)

* Evolutionary Psychology and Artificial Intelligence: The Impact of Artificial Intelligence on Human Behaviour (2020)

* Ethics/Policy

* Robots should be slaves (2010)

* Standardizing Ethical Design for Artificial Intelligence and Autonomous Systems (2017)

* Of, For, and By the People: The Legal Lacuna of Synthetic Persons (2017)

* Patiency is not a virtue: the design of intelligent systems and systems of ethics (2018)

* Other writing

* Reflections on the EU’s AI Act

* Is There an AI Cold War?

* Living with AGI

* One Day, AI Will Seem as Human as Anyone. What Then?

In episode 67 of The Gradient Podcast, Daniel Bashir speaks to Daniel Situnayake.

Daniel is head of Machine Learning at Edge Impulse. He is co-author of the O’Reilly books "AI at the Edge" and "TinyML". Previously, he’s worked on the Tensorflow Lite team at Google AI and co-founded Tiny Farms, an insect farming company. Daniel has also lectured in AIDC technologies at Birmingham City University.

Have suggestions for future podcast guests (or other feedback)? Let us know here!

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

Outline:

* (00:00) Intro

* (1:40) Daniel S Origin Story: computer networking, RFID/barcoding, earlier jobs, Tiny Farms, Tensorflow Lite, writing on TinyML, and Edge Impulse

* (15:30) Edge AI and questions of embodiment/intelligence in AI

* (21:00) The role of hardware, other constraints in edge AI

* (25:00) Definitions of intelligence

* (29:45) What is edge AI?

* (37:30) The spectrum of edge devices

* (43:45) Innovations in edge AI (architecture, frameworks/toolchains, quantization)

* (53:45) Model compression tradeoffs in edge

* (1:00:30) Federated learning and challenges

* (1:09:00) Intro to Edge Impulse

* (1:20:30) Feature engineering for edge systems, fairness considerations

* (1:25:50) Edge AI and axes in AI (large/small, ethereal/embodied)

* (1:37:00) Daniel and Daniel go off the rails on panpsychism

* (1:54:20) Daniel’s advice for aspiring AI practitioners

* (1:57:20) Outro

Links:

* Daniel’s Twitter and blog

* Edge Impulse

In episode 66 of The Gradient Podcast, Daniel Bashir speaks to Soumith Chintala.

Soumith is a Research Engineer at Meta AI Research in NYC. He is the co-creator and lead of Pytorch, and maintains a number of other open-source ML projects including Torch-7 and EBLearn. Soumith has previously worked on robotics, object and human detection, generative modeling, AI for video games, and ML systems research.

Have suggestions for future podcast guests (or other feedback)? Let us know here!

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

Outline:

* (00:00) Intro

* (01:30) Soumith’s intro to AI journey to Pytorch

* (05:00) State of computer vision early in Soumith’s career

* (09:15) Institutional inertia and sunk costs in academia, identifying fads

* (12:45) How Soumith started working on GANs, frustrations

* (17:45) State of ML frameworks early in the deep learning era, differentiators

* (23:50) Frameworks and leveling the playing field, exceptions

* (25:00) Contributing to Torch and evolution into Pytorch

* (29:15) Soumith’s product vision for ML frameworks

* (32:30) From product vision to concrete features in Pytorch

* (39:15) Progressive disclosure of complexity (Chollet) in Pytorch

* (41:35) Building an open source community

* (43:25) The different players in today’s ML framework ecosystem

* (49:35) ML frameworks pioneered by Yann LeCun and Léon Bottou, their influences on Pytorch

* (54:37) Pytorch 2.0 and looking to the future

* (58:00) Soumith’s adventures in household robotics

* (1:03:25) Advice for aspiring ML practitioners

* (1:07:10) Be cool like Soumith and subscribe :)

* (1:07:33) Outro

Links:

* Soumith’s Twitter and homepage

* Papers

* Convolutional Neural Networks Applied to House Numbers Digit Classification

* GANs: LAPGAN, DCGAN, Wasserstein GAN

* Automatic differentiation in PyTorch

* PyTorch: An Imperative Style, High-Performance Deep Learning Library

In episode 65 of The Gradient Podcast, Daniel Bashir speaks to Sewon Min.

Sewon is a fifth-year PhD student in the NLP group at the University of Washington, advised by Hannaneh Hajishirzi and Luke Zettlemoyer. She is a part-time visiting researcher at Meta AI and a recipient of the JP Morgan PhD Fellowship. She has previously spent time at Google Research and Salesforce research.

Have suggestions for future podcast guests (or other feedback)? Let us know here!

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

Outline:

* (00:00) Intro

* (03:00) Origin Story

* (04:20) Evolution of Sewon’s interests, question-answering and practical NLP

* (07:00) Methodology concerns about benchmarks

* (07:30) Multi-hop reading comprehension

* (09:30) Do multi-hop QA benchmarks actually measure multi-hop reasoning?

* (12:00) How models can “cheat” multi-hop benchmarks

* (13:15) Explicit compositionality

* (16:05) Commonsense reasoning and background information

* (17:30) On constructing good benchmarks

* (18:40) AmbigQA and ambiguity

* (22:20) Types of ambiguity

* (24:20) Practical possibilities for models that can handle ambiguity

* (25:45) FaVIQ and fact-checking benchmarks

* (28:45) External knowledge

* (29:45) Fact verification and “complete understanding of evidence”

* (31:30) Do models do what we expect/intuit in reading comprehension?

* (34:40) Applications for fact-checking systems

* (36:40) Intro to in-context learning (ICL)

* (38:55) Example of an ICL demonstration

* (40:45) Rethinking the Role of Demonstrations and what matters for successful ICL

* (43:00) Evidence for a Bayesian inference perspective on ICL

* (45:00) ICL + gradient descent and what it means to “learn”

* (47:00) MetaICL and efficient ICL

* (49:30) Distance between tasks and MetaICL task transfer

* (53:00) Compositional tasks for language models, compositional generalization

* (55:00) The number and diversity of meta-training tasks

* (58:30) MetaICL and Bayesian inference

* (1:00:30) Z-ICL: Zero-Shot In-Context Learning with Pseudo-Demonstrations

* (1:02:00) The copying effect

* (1:03:30) Copying effect for non-identical examples

* (1:06:00) More thoughts on ICL

* (1:08:00) Understanding Chain-of-Thought Prompting

* (1:11:30) Bayes strikes again

* (1:12:30) Intro to Sewon’s text retrieval research

* (1:15:30) Dense Passage Retrieval (DPR)

* (1:18:40) Similarity in QA and retrieval

* (1:20:00) Improvements for DPR

* (1:21:50) Nonparametric Masked Language Modeling (NPM)

* (1:24:30) Difficulties in training NPM and solutions

* (1:26:45) Follow-on work

* (1:29:00) Important fundamental limitations of language models

* (1:31:30) Sewon’s experience doing a PhD

* (1:34:00) Research challenges suited for academics

* (1:35:00) Joys and difficulties of the PhD

* (1:36:30) Sewon’s advice for aspiring PhDs

* (1:38:30) Incentives in academia, production of knowledge

* (1:41:50) Outro

Links:

* Sewon’s homepage and Twitter

* Papers

* Solving and re-thinking benchmarks

* Multi-hop Reading Comprehension through Question Decomposition and Rescoring / Compositional Questions Do Not Necessitate Multi-hop Reasoning

* AmbigQA: Answering Ambiguous Open-domain Questions

* FaVIQ: FAct Verification from Information-seeking Questions

* Language Modeling

* Rethinking the Role of Demonstrations

* MetaICL: Learning to Learn In Context

* Towards Understanding CoT Prompting

* Z-ICL: Zero-Shot In-Context Learning with Pseudo-Demonstrations

* Text representation/retrieval

* Dense Passage Retrieval

* Nonparametric Masked Language Modeling

In episode 64 of The Gradient Podcast, Daniel Bashir speaks to Richard Socher.

Richard is founder and CEO of you.com, a new search engine that lets you personalize your search workflow and eschews tracking and invasive ads. Richard was previously Chief Scientist at Salesforce where he led work on fundamental and applied research, product incubation, CRM search, customer service automation and a cross-product AI platform. He was an adjunct professor at Stanford’s CS department as well as founder and CEO/CTO of MetaMind, which was acquired by Salesforce in 2016. He received his PhD from Stanford’s CS Department in 2014.

Have suggestions for future podcast guests (or other feedback)? Let us know here!

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

Outline:

* (00:00) Intro

* (02:20) Richard Socher origin story + time at Metamind, Salesforce (AI Economist, CTRL, ProGen)

* (22:00) Why Richard advocated for deep learning in NLP

* (27:00) Richard’s perspective on language

* (32:20) Is physical grounding and language necessary for intelligence?

* (40:10) Frankfurtian b******t and language model utterances as truth

* (47:00) Lessons from Salesforce Research

* (53:00) Balancing fundamental research with product focus

* (57:30) The AI Economist + how should policymakers account for limitations?

* (1:04:50) you.com, the chatbot wars, and taking on search giants

* (1:13:50) Re-imagining the vision for and components of a search engine

* (1:18:00) The future of generative models in search and the internet

* (1:28:30) Richard’s advice for early-career technologists

* (1:37:00) Outro

Links:

* Richard’s Twitter 

* YouChat by you.com

* Careers at you.com

* Papers mentioned

* Semi-Supervised Recursive Autoencoders for Predicting Sentiment Distributions

* Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank

* Grounded Compositional Semantics for Finding and Describing Images with Sentences

* The AI Economist

* ProGen

* CTRL

In episode 63 of The Gradient Podcast, Daniel Bashir speaks to Joe Edelman.

Joe developed the meaning-based organizational metrics at Couchsurfing.com, then co-founded the Center for Humane Technology with Tristan Harris, and coined the term “Time Well Spent” for a family of metrics adopted by teams at Facebook, Google, and Apple. Since then, he's worked on the philosophical underpinnings for new business metrics, design methods, and political movements. The central idea is to make people's sources of meaning explicit, so that how meaningful or meaningless things are can be rigorously accounted for. His previous career was in HCI and programming language design.

Have suggestions for future podcast guests (or other feedback)? Let us know here!

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

Outline:

* (00:00) Intro (yes Daniel is trying a new intro format)

* (01:30) Joe’s origin story

* (07:15) Revealed preferences and personal meaning, recommender systems

* (12:30) Is using revealed preferences necessary?

* (17:00) What are values and how do you detect them?

* (24:00) Figuring out what’s meaningful to us

* (28:45) The decline of spaces and togetherness

* (35:00) Individualism and economic/political theory, tensions between collectivism/individualism

* (41:00) What it looks like to build spaces, Habitat

* (47:15) Cognitive effects of social platforms

* (51:45) Atomized communication, re-imagining chat apps

* (55:50) Systems for social groups and medium independence

* (1:02:45) Spaces being built today

* (1:05:15) Joe is building research groups! Get in touch :)

* (1:05:40) Outro

Links:

* Joe's 80m lecture on techniques for rebuilding society on meaning (youtube, transcript)

* The discord for Rebuilding Meaning—join if you'd like to help build ML models or metrics using the methods discussed

* Writing/papers mentioned:

* Tech products (that don’t cause depression and war)

* Values, Preferences, Meaningful Choice

* Social Programming Considered as a Habitat for Groups

* Is Anything Worth Maximizing

* Joe’s homepage, Twitter, and YouTube page

In episode 62 of The Gradient Podcast, Daniel Bashir speaks to Ed Grefenstette.

Ed is Head of Machine Learning at Cohere and an Honorary Professor at University College London. He previously held research scientist positions at Facebook AI Research and DeepMind, following a stint as co-founder and CTO of Dark Blue Labs. Before his time in industry, Ed worked at Oxford’s Department of Computer Science as a lecturer and Fulford Junior Research Fellow at Somerville College. Ed also received his MSc and DPhil from Oxford’s Computer Science Department.

Have suggestions for future podcast guests (or other feedback)? Let us know here!

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

Outline:

* (00:00) Intro

* (02:18) The Ed Grefenstette Origin Story

* (08:15) Distributional semantics and Ed’s PhD research

* (14:30) Extending the distributional hypothesis, later Wittgenstein

* (18:00) Recovering parse trees in LMs, can LLMs understand communication and not just bare language?

* (23:15) LMs capture something about pragmatics, proxies for grounding and pragmatics

* (25:00) Human-in-the-loop training and RLHF—what is the essential differentiator?

* (28:15) A convolutional neural network for modeling sentences, relationship to attention

* (34:20) Difficulty of constructing supervised learning datasets, benchmark-driven development

* (40:00) Learning to Transduce with Unbounded Memory, Neural Turing Machines

* (47:40) If RNNs are like finite state machines, where are transformers?

* (51:40) Cohere and why Ed joined

* (56:30) Commercial applications of LLMs and Cohere’s product

* (59:00) Ed’s reply to stochastic parrots and thoughts on consciousness

* (1:03:30) Lessons learned about doing effective science

* (1:05:00) Where does scaling end?

* (1:07:00) Why Cohere is an exciting place to do science

* (1:08:00) Ed’s advice for aspiring ML {researchers, engineers, etc} and the role of communities in science

* (1:11:45) Cohere for AI plug!

* (1:13:30) Outro

Links:

* Ed’s homepage and Twitter

* (some of) Ed’s Papers

* Experimental support for a categorical compositional distributional model of meaning

* Multi-step regression learning

* “Not not bad” is not “bad”

* Towards a formal distributional semantics

* A CNN for modeling sentences

* Teaching machines to read and comprehend

* Reasoning about entailment with neural attention

* Learning to Transduce with Unbounded Memory

* Teaching Artificial Agents to Understand Language by Modelling Reward

* Other things mentioned

* Large language models are not zero-shot communicators (Laura Ruis + others and Ed)

* Looped Transformers as Programmable Computers and our Update 43 covering this paper

* Cohere and Cohere for AI (+ earlier episode w/ Sara Hooker on C4AI)

* David Chalmers interview on AI + consciousness

In episode 61 of The Gradient Podcast, Daniel Bashir speaks to Ken Liu.

Ken is an author of speculative fiction. A winner of the Nebula, Hugo, and World Fantasy awards, he is the author of silkpunk epic fantasy series Dandelion Dynasty and short story collections The Paper Menagerie and Other Stories and The Hidden Girl and Other Stories. Prior to writing full-time, Ken worked as a software engineer, corporate lawyer, and litigation consultant.

Have suggestions for future podcast guests (or other feedback)? Let us know here!

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

Outline:

* (00:00) Intro

* (02:00) How Ken Liu became Ken Liu: A Saga

* (03:10) Time in the tech industry, interest in symbolic machines

* (04:40) Determining what stories to write, (07:00) art as failed communication

* (07:55) Law as creating abstract machines, importance of successful communication, stories in law

* (13:45) Misconceptions about science fiction

* (18:30) How we’ve been misinformed about literature and stories in school, stories as expressing multivalent truths, Dickens on narration (29:00)

* (31:20) Stories as imposing structure on the world

* (35:25) Silkpunk as aesthetic and writing approach

* (39:30) If modernity is a translated experience, what is it translated from? Alternative sources for the American pageant

* (47:30) The value of silkpunk for technologists and building the future

* (52:40) The engineer as poet

* (59:00) Technology language as constructing societies, what it is to be a technologist

* (1:04:00) The technology of language

* (1:06:10) The Google Wordcraft Workshop and co-writing with LaMDA

* (1:14:10) Possibilities and limitations of LMs in creative writing

* (1:18:45) Ken’s short fiction

* (1:19:30) Short fiction as a medium

* (1:24:45) “The Perfect Match” (from The Paper Menagerie and other stories)

* (1:34:00) Possibilities for better recommender systems

* (1:39:35) “Real Artists” (from The Hidden Girl and other stories)

* (1:47:00) The scaling hypothesis and creativity

* (1:50:25) “The Gods have not died in vain” & Moore’s Proof epigraph (The Hidden Girl)

* (1:53:10) More of The Singularity Trilogy (The Hidden Girl)

* (1:58:00) The role of science fiction today and how technologists should engage with stories

* (2:01:53) Outro

Links:

* Ken’s homepage

* The Dandelion Dynasty Series: Speaking Bones is out in paperback

* Books/Stories/Projects Mentioned

* “Evaluative Soliloquies” in Google Wordcraft

* The Paper Menagerie and Other Stories

* The Hidden Girl and Other Stories

In episode 60 of The Gradient Podcast, Daniel Bashir speaks to Hattie Zhou.

Hattie is a PhD student at the Université de Montréal and Mila. Her research focuses on understanding how and why neural networks work, based on the belief that the performance of modern neural networks exceeds our understanding and that building more capable and trustworthy models requires bridging this gap. Prior to Mila, she spent time as a data scientist at Uber and did research with Uber AI Labs.

Have suggestions for future podcast guests (or other feedback)? Let us know here!

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

Outline:

* (00:00) Intro

* (01:55) Hattie’s Origin Story, Uber AI Labs, empirical theory and other sorts of research

* (10:00) Intro to the Lottery Ticket Hypothesis & Deconstructing Lottery Tickets

* (14:30) Lottery tickets as lucky initialization

* (17:00) Types of masking and the “masking is training” claim

* (24:00) Type-0 masks and weight evolution over long training trajectories

* (27:00) Can you identify good masks or training trajectories a priori?

* (29:00) The role of signs in neural net initialization

* (35:27) The Supermask

* (41:00) Masks to probe pretrained models and model steerability

* (47:40) Fortuitous Forgetting in Connectionist Networks

* (54:00) Relationships to other work (double descent, grokking, etc.)

* (1:01:00) The iterative training process in fortuitous forgetting, scale and value of exploring alternatives

* (1:03:35) In-Context Learning and Teaching Algorithmic Reasoning

* (1:09:00) Learning + algorithmic reasoning, prompting strategy

* (1:13:50) What’s happening with in-context learning?

* (1:14:00) Induction heads

* (1:17:00) ICL and gradient descent

* (1:22:00) Algorithmic prompting vs discovery

* (1:24:45) Future directions for algorithmic prompting

* (1:26:30) Interesting work from NeurIPS 2022

* (1:28:20) Hattie’s perspective on scientific questions people pay attention to, underrated problems

* (1:34:30) Hattie’s perspective on ML publishing culture

* (1:42:12) Outro

Links:

* Hattie’s homepage and Twitter

* Papers

* Deconstructing Lottery Tickets: Zeros, signs, and the Supermask

* Fortuitous Forgetting in Connectionist Networks

* Teaching Algorithmic Reasoning via In-context Learning

In episode 59 of The Gradient Podcast, Daniel Bashir speaks to Professor Kyunghyun Cho.

Professor Cho is an associate professor of computer science and data science at New York University and CIFAR Fellow of Learning in Machines & Brains. He is also a senior director of frontier research at the Prescient Design team within Genentech Research & Early Development. He was a research scientist at Facebook AI Research from 2017-2020 and a postdoctoral fellow at University of Montreal under the supervision of Prof. Yoshua Bengio after receiving his MSc and PhD degrees from Aalto University. He received the Samsung Ho-Am Prize in Engineering in 2021.

Have suggestions for future podcast guests (or other feedback)? Let us know here!

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

Outline:

* (00:00) Intro

* (02:15) How Professor Cho got into AI, going to Finland for a PhD

* (06:30) Accidental and non-accidental parts of Prof Cho’s journey, the role of timing in career trajectories

* (09:30) Prof Cho’s M.Sc. thesis on Restricted Boltzmann Machines

* (17:00) The state of autodiff at the time

* (20:00) Finding non-mainstream problems and examining limitations of mainstream approaches, anti-dogmatism, Yoshua Bengio appreciation

* (24:30) Detaching identity from work, scientific training

* (26:30) The rest of Prof Cho’s PhD, the first ICLR conference, working in Yoshua Bengio’s lab

* (34:00) Prof Cho’s isolation during his PhD and its impact on his work—transcending insecurity and working on unsexy problems

* (41:30) The importance of identifying important problems and developing an independent research program, ceiling on the number of important research problems

* (46:00) Working on Neural Machine Translation, Jointly Learning to Align and Translate

* (1:01:45) What RNNs and earlier NN architectures can still teach us, why transformers were successful

* (1:08:00) Science progresses gradually

* (1:09:00) Learning distributed representations of sentences, extending the distributional hypothesis

* (1:21:00) Difficulty and limitations in evaluation—directions of dynamic benchmarks, trainable evaluation metrics

* (1:29:30) Mixout and AdapterFusion: fine-tuning and intervening on pre-trained models, pre-training as initialization, destructive interference

* (1:39:00) Analyzing neural networks as reading tea leaves

* (1:44:45) Importance of healthy skepticism for scientists

* (1:45:30) Language-guided policies and grounding, vision-language navigation

* (1:55:30) Prof Cho’s reflections on 2022

* (2:00:00) Obligatory ChatGPT content

* (2:04:50) Finding balance

* (2:07:15) Outro

Links:

* Professor Cho’s homepage and Twitter

* Papers

* M.Sc. thesis and PhD thesis

* NMT and attention

* Properties of NMT,

* Learning Phrase Representations

* Neural machine translation by jointly learning to align and translate

* More recent work

* Learning Distributed Representations of Sentences from Unlabelled Data

* Mixout: Effective Regularization to Finetune Large-scale Pretrained Language Models

* Generative Language-Grounded Policy in Vision-and-Language Navigation with Bayes’ Rule

* AdapterFusion: Non-Destructive Task Composition for Transfer Learning

In episode 58 of The Gradient Podcast, Daniel Bashir speaks to Professor Steve Miller.

Steve is a Professor Emeritus of Information Systems at Singapore Management University. Steve served as Founding Dean for the SMU School of Information Systems, and established and developed the technology core of SIS research and project capabilities in Cybersecurity, Data Management & Analytics, Intelligent Systems & Decision Analytics, and Software & Cyber-Physical Systems, as well as the management science oriented capability in Information Systems & Management. Steve works closely with a number of Singapore government ministries and agencies via steering committees, advisory boards, and advisory appointments.

Have suggestions for future podcast guests (or other feedback)? Let us know here!

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

Outline:

* (00:00) Intro

* (02:40) Steve’s evolution of interests in AI, time in academia and industry

* (05:15) How different is this “industrial revolution”?

* (10:00) What new technologies enable, the human role in technology’s impact on jobs

* (11:35) Automation and augmentation and the realities of integrating new technologies in the workplace

* (21:50) Difficulties of applying AI systems in real-world contexts

* (32:45) Re-calibrating human work with intelligent machines

* (39:00) Steve’s thinking on the nature of human/machine intelligence, implications for human/machine hybrid work

* (47:00) Tradeoffs in using ML systems for automation/augmentation

* (52:40) Organizational adoption of AI and speed

* (1:01:55) Technology adoption is more than just a technology problem

* (1:04:50) Progress narratives, “safe to speed”

* (1:10:27) Outro