Machine Learning Street Talk (MLST)

Support us! https://www.patreon.com/mlst

Professor Murray Shanahan is a renowned researcher on sophisticated cognition and its implications for artificial intelligence. His 2016 article ‘Conscious Exotica’ explores the Space of Possible Minds, a concept first proposed by philosopher Aaron Sloman in 1984, which includes all the different forms of minds from those of other animals to those of artificial intelligence. Shanahan rejects the idea of an impenetrable realm of subjective experience and argues that the majority of the space of possible minds may be occupied by non-natural variants, such as the ‘conscious exotica’ of which he speaks.  In his paper ‘Talking About Large Language Models’, Shanahan discusses the capabilities and limitations of large language models (LLMs). He argues that prompt engineering is a key element for advanced AI systems, as it involves exploiting prompt prefixes to adjust LLMs to various tasks. However, Shanahan cautions against ascribing human-like characteristics to these systems, as they are fundamentally different and lack a shared comprehension with humans. Even though LLMs can be integrated into embodied systems, it does not mean that they possess human-like language abilities. Ultimately, Shanahan concludes that although LLMs are formidable and versatile, we must be wary of over-simplifying their capacities and limitations.

YT version: https://youtu.be/BqkWpP3uMMU

Full references on the YT description. 

[00:00:00] Introduction

[00:08:51] Consciousness and  Consciousness Exotica

[00:34:59] Slightly Consciousness LLMs

[00:38:05] Embodiment

[00:51:32] Symbol Grounding 

[00:54:13] Emergence

[00:57:09] Reasoning

[01:03:16] Intentional Stance

[01:07:06] Digression on Chomsky show and Andrew Lampinen

[01:10:31] Prompt Engineering

Find Murray online:

https://www.doc.ic.ac.uk/~mpsha/

https://twitter.com/mpshanahan?lang=en

https://scholar.google.co.uk/citations?user=00bnGpAAAAAJ&hl=en

MLST Discord: https://discord.gg/aNPkGUQtc5

Support us! https://www.patreon.com/mlst

Sara Hooker is an exceptionally talented and accomplished leader and research scientist in the field of machine learning. She is the founder of Cohere For AI, a non-profit research lab that seeks to solve complex machine learning problems. She is passionate about creating more points of entry into machine learning research and has dedicated her efforts to understanding how progress in this field can be translated into reliable and accessible machine learning in the real-world.

Sara is also the co-founder of the Trustworthy ML Initiative, a forum and seminar series related to Trustworthy ML. She is on the advisory board of Patterns and is an active member of the MLC research group, which has a focus on making participation in machine learning research more accessible.

Before starting Cohere For AI, Sara worked as a research scientist at Google Brain. She has written several influential research papers, including "The Hardware Lottery", "The Low-Resource Double Bind: An Empirical Study of Pruning for Low-Resource Machine Translation", "Moving Beyond “Algorithmic Bias is a Data Problem”" and "Characterizing and Mitigating Bias in Compact Models". 

In addition to her research work, Sara is also the founder of the local Bay Area non-profit Delta Analytics, which works with non-profits and communities all over the world to build technical capacity and empower others to use data. She regularly gives tutorials on machine learning fundamentals, interpretability, model compression and deep neural networks and is dedicated to collaborating with independent researchers around the world.

Sara Hooker is famous for writing a paper introducing the concept of the 'hardware lottery', in which the success of a research idea is determined not by its inherent superiority, but by its compatibility with available software and hardware. She argued that choices about software and hardware have had a substantial impact in deciding the outcomes of early computer science history, and that with the increasing heterogeneity of the hardware landscape, gains from advances in computing may become increasingly disparate. Sara proposed that an interim goal should be to create better feedback mechanisms for researchers to understand how their algorithms interact with the hardware they use. She suggested that domain-specific languages, auto-tuning of algorithmic parameters, and better profiling tools may help to alleviate this issue, as well as provide researchers with more informed opinions about how hardware and software should progress. Ultimately, Sara encouraged researchers to be mindful of the implications of the hardware lottery, as it could mean that progress on some research directions is further obstructed. If you want to learn more about that paper, watch our previous interview with Sara.

YT version: https://youtu.be/7oJui4eSCoY

MLST Discord: https://discord.gg/aNPkGUQtc5

TOC:

[00:00:00] Intro

[00:02:53] Interpretability / Fairness

[00:35:29] LLMs

Find Sara:

https://www.sarahooker.me/

https://twitter.com/sarahookr

Support us! https://www.patreon.com/mlst

Hattie Zhou, a PhD student at Université de Montréal and Mila, has set out to understand and explain the performance of modern neural networks, believing it a key factor in building better, more trusted models. Having previously worked as a data scientist at Uber, a private equity analyst at Radar Capital, and an economic consultant at Cornerstone Research, she has recently released a paper in collaboration with the Google Brain team, titled ‘Teaching Algorithmic Reasoning via In-context Learning’. In this work, Hattie identifies and examines four key stages for successfully teaching algorithmic reasoning to large language models (LLMs): formulating algorithms as skills, teaching multiple skills simultaneously, teaching how to combine skills, and teaching how to use skills as tools. Through the application of algorithmic prompting, Hattie has achieved remarkable results, with an order of magnitude error reduction on some tasks compared to the best available baselines. This breakthrough demonstrates algorithmic prompting’s viability as an approach for teaching algorithmic reasoning to LLMs, and may have implications for other tasks requiring similar reasoning capabilities.

TOC

[00:00:00] Hattie Zhou

[00:19:49] Markus Rabe [Google Brain]

Hattie's Twitter - https://twitter.com/oh_that_hat

Website - http://hattiezhou.com/

Teaching Algorithmic Reasoning via In-context Learning [Hattie Zhou, Azade Nova, Hugo Larochelle, Aaron Courville, Behnam Neyshabur, and Hanie Sedghi]

https://arxiv.org/pdf/2211.09066.pdf

Markus Rabe [Google Brain]:

https://twitter.com/markusnrabe

https://research.google/people/106335/

https://www.linkedin.com/in/markusnrabe

Autoformalization with Large Language Models [Albert Jiang Charles Edgar Staats Christian Szegedy Markus Rabe Mateja Jamnik Wenda Li Yuhuai Tony Wu]

https://research.google/pubs/pub51691/

Discord: https://discord.gg/aNPkGUQtc5

YT: https://youtu.be/80i6D2TJdQ4

Support us! https://www.patreon.com/mlst

(On the main version we released; the music was a tiny bit too loud in places, and some pieces had percussion which was a bit distracting -- here is a version with all music removed so you have the option! )

David Chalmers is a professor of philosophy and neural science at New York University, and an honorary professor of philosophy at the Australian National University. He is the co-director of the Center for Mind, Brain, and Consciousness, as well as the PhilPapers Foundation. His research focuses on the philosophy of mind, especially consciousness, and its connection to fields such as cognitive science, physics, and technology. He also investigates areas such as the philosophy of language, metaphysics, and epistemology. With his impressive breadth of knowledge and experience, David Chalmers is a leader in the philosophical community.

The central challenge for consciousness studies is to explain how something immaterial, subjective, and personal can arise out of something material, objective, and impersonal. This is illustrated by the example of a bat, whose sensory experience is much different from ours, making it difficult to imagine what it's like to be one. Thomas Nagel's "inconceivability argument" has its advantages and disadvantages, but ultimately it is impossible to solve the mind-body problem due to the subjective nature of experience. This is further explored by examining the concept of philosophical zombies, which are physically and behaviorally indistinguishable from conscious humans yet lack conscious experience. This has implications for the Hard Problem of Consciousness, which is the attempt to explain how mental states are linked to neurophysiological activity. The Chinese Room Argument is used as a thought experiment to explain why physicality may be insufficient to be the source of the subjective, coherent experience we call consciousness. Despite much debate, the Hard Problem of Consciousness remains unsolved. Chalmers has been working on a functional approach to decide whether large language models are, or could be conscious. 

Filmed at #neurips22

Discord: https://discord.gg/aNPkGUQtc5

Pod: https://anchor.fm/machinelearningstreettalk/episodes/90---Prof--DAVID-CHALMERS---Slightly-Conscious-LLMs-e1sej50

TOC;

[00:00:00] Introduction

[00:00:40] LLMs consciousness pitch

[00:06:33] Philosophical Zombies

[00:09:26] The hard problem of consciousness

[00:11:40] Nagal's bat and intelligibility 

[00:21:04] LLM intro clip from NeurIPS

[00:22:55] Connor Leahy on self-awareness in LLMs

[00:23:30] Sneak peek from unreleased show - could consciousness be a submodule?

[00:33:44] SeppH

[00:36:15] Tim interviews David at NeurIPS (functionalism / panpsychism / Searle)

[00:45:20] Peter Hase interviews Chalmers (focus on interpretability/safety)

Panel:

Dr. Tim Scarfe

Dr. Keith Duggar

Contact David;

https://mobile.twitter.com/davidchalmers42

https://consc.net/

References;

Could a Large Language Model Be Conscious? [Chalmers NeurIPS22 talk]

https://nips.cc/media/neurips-2022/Slides/55867.pdf

What Is It Like to Be a Bat? [Nagel]

https://warwick.ac.uk/fac/cross_fac/iatl/study/ugmodules/humananimalstudies/lectures/32/nagel_bat.pdf

Zombies

https://plato.stanford.edu/entries/zombies/

zombies on the web [Chalmers]

https://consc.net/zombies-on-the-web/

The hard problem of consciousness [Chalmers]

https://psycnet.apa.org/record/2007-00485-017

David Chalmers, "Are Large Language Models Sentient?" [NYU talk, same as at NeurIPS]

https://www.youtube.com/watch?v=-BcuCmf00_Y

Support us! https://www.patreon.com/mlst

David Chalmers is a professor of philosophy and neural science at New York University, and an honorary professor of philosophy at the Australian National University. He is the co-director of the Center for Mind, Brain, and Consciousness, as well as the PhilPapers Foundation. His research focuses on the philosophy of mind, especially consciousness, and its connection to fields such as cognitive science, physics, and technology. He also investigates areas such as the philosophy of language, metaphysics, and epistemology. With his impressive breadth of knowledge and experience, David Chalmers is a leader in the philosophical community.

The central challenge for consciousness studies is to explain how something immaterial, subjective, and personal can arise out of something material, objective, and impersonal. This is illustrated by the example of a bat, whose sensory experience is much different from ours, making it difficult to imagine what it's like to be one. Thomas Nagel's "inconceivability argument" has its advantages and disadvantages, but ultimately it is impossible to solve the mind-body problem due to the subjective nature of experience. This is further explored by examining the concept of philosophical zombies, which are physically and behaviorally indistinguishable from conscious humans yet lack conscious experience. This has implications for the Hard Problem of Consciousness, which is the attempt to explain how mental states are linked to neurophysiological activity. The Chinese Room Argument is used as a thought experiment to explain why physicality may be insufficient to be the source of the subjective, coherent experience we call consciousness. Despite much debate, the Hard Problem of Consciousness remains unsolved. Chalmers has been working on a functional approach to decide whether large language models are, or could be conscious. 

Filmed at #neurips22

Discord: https://discord.gg/aNPkGUQtc5

YT: https://youtu.be/T7aIxncLuWk

TOC;

[00:00:00] Introduction

[00:00:40] LLMs consciousness pitch

[00:06:33] Philosophical Zombies

[00:09:26] The hard problem of consciousness

[00:11:40] Nagal's bat and intelligibility

[00:21:04] LLM intro clip from NeurIPS

[00:22:55] Connor Leahy on self-awareness in LLMs

[00:23:30] Sneak peek from unreleased show - could consciousness be a submodule?

[00:33:44] SeppH

[00:36:15] Tim interviews David at NeurIPS (functionalism / panpsychism / Searle)

[00:45:20] Peter Hase interviews Chalmers (focus on interpretability/safety)

Panel:

Dr. Tim Scarfe

Dr. Keith Duggar

Contact David;

https://mobile.twitter.com/davidchalmers42

https://consc.net/

References;

Could a Large Language Model Be Conscious? [Chalmers NeurIPS22 talk] 

https://nips.cc/media/neurips-2022/Slides/55867.pdf

What Is It Like to Be a Bat? [Nagel]

https://warwick.ac.uk/fac/cross_fac/iatl/study/ugmodules/humananimalstudies/lectures/32/nagel_bat.pdf

Zombies

https://plato.stanford.edu/entries/zombies/

zombies on the web [Chalmers]

https://consc.net/zombies-on-the-web/

The hard problem of consciousness [Chalmers]

https://psycnet.apa.org/record/2007-00485-017

David Chalmers, "Are Large Language Models Sentient?" [NYU talk, same as at NeurIPS]

https://www.youtube.com/watch?v=-BcuCmf00_Y

Support us! https://www.patreon.com/mlst

Dr. Walid Saba recently reviewed the book Machines Will Never Rule The World, which argues that strong AI is impossible. He acknowledges the complexity of modeling mental processes and language, as well as interactive dialogues, and questions the authors' use of "never." Despite his skepticism, he is impressed with recent developments in large language models, though he questions the extent of their success.

We then discussed the successes of cognitive science. Walid believes that something has been achieved which many cognitive scientists would never accept, namely the ability to learn from data empirically. Keith agrees that this is a huge step, but notes that there is still much work to be done to get to the "other 5%" of accuracy. They both agree that the current models are too brittle and require much more data and parameters to get to the desired level of accuracy.

Walid then expresses admiration for deep learning systems' ability to learn non-trivial aspects of language from ingesting text only. He argues that this is an "existential proof" of language competency and that it would be impossible for a group of luminaries such as Montague, Marvin Minsky, John McCarthy, and a thousand other bright engineers to replicate the same level of competency as we have now with LLMs. He then discusses the problem of semantics and pragmatics, as well as symbol grounding, and expresses skepticism about grounded meaning and embodiment. He believes that artificial intelligence should be used to solve real-world problems which require human intelligence but not believe that robots should be built to understand love or other subjective feelings.

We discussed the unique properties of natural human language. Walid believes that the core unique property is the ability to do abductive reasoning, which is the process of reasoning to the best explanation or understanding. Keith adds that there are two types of abduction - one for generating hypotheses and one for justifying them. In both cases, abductive reasoning involves choosing from a set of plausible possibilities.

Finally, we discussed the book "Machines Will Never Rule The World" and its argument that the current mathematics and technology is not enough to model complex systems. Walid agrees with the book's argument but is still optimistic that a new mathematics can be discovered. Keith suggests the possibility of an AGI discovering the mathematics to create itself. They also discussed how the book could serve as a reminder to temper the hype surrounding AI and to focus on exploration, creativity, and daring ideas. Walid ended by stressing the importance of science, noting that engineers should play within the Venn diagrams drawn by scientists, rather than trying to hack their way through it.

Transcript: https://share.descript.com/view/BFQb5iaegJC

Discord: https://discord.gg/aNPkGUQtc5

YT: https://youtu.be/IMnWAuoucjo

TOC:

[00:00:00] Intro

[00:06:52] Walid's change of heart on DL/LLMs and on the skeptics like Gary Marcus

[00:22:52] Symbol Grounding

[00:32:26] On Montague

[00:40:41] On Abduction

[00:50:54] Language of thought

[00:56:08] Why machines will never rule the world book review

[01:20:06] Engineers should play in the scientists Venn Diagram!

Panel;

Dr. Tim Scarfe

Dr. Keith Duggar

Mark Mcguill

Yann LeCun is a French computer scientist known for his pioneering work on convolutional neural networks, optical character recognition and computer vision. He is a Silver Professor at New York University and Vice President, Chief AI Scientist at Meta. Along with Yoshua Bengio and Geoffrey Hinton, he was awarded the 2018 Turing Award for their work on deep learning, earning them the nickname of the "Godfathers of Deep Learning".

Dr. Randall Balestriero has been researching learnable signal processing since 2013, with a focus on learnable parametrized wavelets and deep wavelet transforms. His research has been used by NASA, leading to applications such as Marsquake detection. During his PhD at Rice University, Randall explored deep networks from a theoretical perspective and improved state-of-the-art methods such as batch-normalization and generative networks. Later, when joining Meta AI Research (FAIR) as a postdoc with Prof. Yann LeCun, Randall further broadened his research interests to include self-supervised learning and the biases emerging from data-augmentation and regularization, resulting in numerous publications.

Episode recorded live at NeurIPS. 

YT: https://youtu.be/9dLd6n9yT8U (references are there)

Support us! https://www.patreon.com/mlst 

Host: Dr. Tim Scarfe

TOC:

[00:00:00] LeCun interview

[00:18:25] Randall Balestriero interview (mostly on spectral SSL paper, first ref)

Dr. Petar Veličković  is a Staff Research Scientist at DeepMind, he has firmly established himself as one of the most significant up and coming researchers in the deep learning space. He invented Graph Attention Networks in 2017 and has been a leading light in the field ever since pioneering research in Graph Neural Networks, Geometric Deep Learning and also Neural Algorithmic reasoning. If you haven’t already, you should check out our video on the Geometric Deep learning blueprint, featuring Petar. I caught up with him last week at NeurIPS. In this show, from NeurIPS 2022 we discussed his recent work on category theory and graph neural networks.

https://petar-v.com/

https://twitter.com/PetarV_93/

TOC:

Categories  (Cats for AI) [00:00:00]

Reasoning [00:14:44]

Extrapolation [00:19:09]

Ishan Misra Skit [00:27:50]

Graphs (Expander Graph Propagation) [00:29:18]

YT: https://youtu.be/1lkdWduuN14

MLST Discord: https://discord.gg/V25vQeFwhS

Support us! https://www.patreon.com/mlst

References on YT description, lots of them! 

Host: Dr. Tim Scarfe

In this NeurIPSs interview, we speak with Laura Ruis about her research on the ability of language models to interpret language in context. She has designed a simple task to evaluate the performance of widely used state-of-the-art language models and has found that they struggle to make pragmatic inferences (implicatures). Tune in to learn more about her findings and what they mean for the future of conversational AI.

Laura Ruis

https://www.lauraruis.com/

https://twitter.com/LauraRuis

BLOOM

https://bigscience.huggingface.co/blog/bloom

Large language models are not zero-shot communicators [Laura Ruis, Akbir Khan, Stella Biderman, Sara Hooker, Tim Rocktäschel, Edward Grefenstette]

https://arxiv.org/abs/2210.14986

[Zhang et al] OPT: Open Pre-trained Transformer Language Models

https://arxiv.org/pdf/2205.01068.pdf

[Lampinen] Can language models handle recursively nested grammatical structures? A case study on comparing models and humans

https://arxiv.org/pdf/2210.15303.pdf

[Gary Marcus] Horse rides astronaut

https://garymarcus.substack.com/p/horse-rides-astronaut

[Gary Marcus] GPT-3, Bloviator: OpenAI’s language generator has no idea what it’s talking about

https://www.technologyreview.com/2020/08/22/1007539/gpt3-openai-language-generator-artificial-intelligence-ai-opinion/

[Bender et al] On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?

https://dl.acm.org/doi/10.1145/3442188.3445922 

[janus] Simulators (Less Wrong)

https://www.lesswrong.com/posts/vJFdjigzmcXMhNTsx/simulators

First in our unplugged series live from #NeurIPS2022

We discuss natural language understanding, symbol meaning and grounding and Chomsky with Dr. Andrew Lampinen from DeepMind. 

We recorded a LOT of material from NeurIPS, keep an eye out for the uploads. 

YT version: https://youtu.be/46A-BcBbMnA

References

[Paul Cisek] Beyond the computer metaphor: Behaviour as interaction

https://philpapers.org/rec/CISBTC

Linguistic Competence (Chomsky reference)

https://en.wikipedia.org/wiki/Linguistic_competence

[Andrew Lampinen] Can language models handle recursively nested grammatical structures? A case study on comparing models and humans

https://arxiv.org/abs/2210.15303

[Fodor et al] Connectionism and Cognitive Architecture: A Critical Analysis

https://ruccs.rutgers.edu/images/personal-zenon-pylyshyn/proseminars/Proseminar13/ConnectionistArchitecture.pdf

[Melanie Mitchell et al] The Debate Over Understanding in AI's Large Language Models

https://arxiv.org/abs/2210.13966

[Gary Marcus] GPT-3, Bloviator: OpenAI’s language generator has no idea what it’s talking about

https://www.technologyreview.com/2020/08/22/1007539/gpt3-openai-language-generator-artificial-intelligence-ai-opinion/

[Bender et al] On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?

https://dl.acm.org/doi/10.1145/3442188.3445922

[Adam Santoro, Andrew Lampinen et al] Symbolic Behaviour in Artificial Intelligence

https://arxiv.org/abs/2102.03406

[Ishita Dasgupta, Lampinen et al] Language models show human-like content effects on reasoning

https://arxiv.org/abs/2207.07051

REACT - Synergizing Reasoning and Acting in Language Models

https://arxiv.org/pdf/2210.03629.pdf

https://ai.googleblog.com/2022/11/react-synergizing-reasoning-and-acting.html

[Fabian Paischer] HELM - History Compression via Language Models in Reinforcement Learning

https://ml-jku.github.io/blog/2022/helm/

https://arxiv.org/abs/2205.12258

[Laura Ruis] Large language models are not zero-shot communicators

https://arxiv.org/pdf/2210.14986.pdf

[Kumar] Using natural language and program abstractions to instill human inductive biases in machines

https://arxiv.org/pdf/2205.11558.pdf

Juho Kim

https://juhokim.com/