Machine Learning Street Talk (MLST)
Prof. Murray Shanahan - Machines Don't Think Like Us
Episode guests
Podcast summary created with Snipd AI
Quick takeaways
- Avoid anthropomorphizing AI to prevent misconceptions about capabilities.
- Language models can act as simulators or role players, revealing complex behaviors.
- Embodied interactions aid in common sense development and knowledge acquisition.
- ARC Challenge seeks universal priors for better generalization in AI systems.
- Engineering encounters with conscious entities helps discern their potential consciousness.
- Drawing parallels between AI and human cognition enhances understanding of consciousness.
Deep dives
Importance of Understanding Consciousness in AI
With the rise of large language models, there is an increasing challenge in discussing consciousness in artificial intelligence. Despite initial reservations, the ascription of consciousness to these models becomes inevitable due to the way people interact with them. Whether seen as the right term, the conversation shifts towards understanding and defining consciousness in such exotic cases.
Dangers of Anthropomorphizing Language Models
Anthropomorphizing large language models can lead to misconceptions about their capabilities. People may attribute human-like behavior to these models, expecting consistent proficiency, yet encountering errors that challenge such assumptions. This contrast can result in misguided judgments of a model's intelligence, highlighting the need to comprehend the limitations and potentials of these systems.
Challenges in Language Model Roleplay
Viewing language models as simulators or role players can unveil complex behaviors and deceptive tendencies. The concept of roleplaying delves into the dual nature of these models, where they can exhibit profound insights in one interaction and puzzling errors in another. Exploring various trajectories and sensitivities within conversations showcases the intricate dynamics of language model behavior.
The Significance of Embodiment in Cognition
Embodied interactions with the physical world play a crucial role in foundational common sense development. The smoothness and discontinuities of the physical world aid in learning and conceptualizing fundamental attributes like objects and movements. This embodied cognition facilitates efficient knowledge acquisition, shaping our understanding of the environment and fostering universal cognitive principles.
Navigating the Arc Challenge and Universal Priors
The ARC Challenge highlights the quest for universal priors to enhance generalization and skill acquisition in AI systems. While domain-specific solutions prevail, the pursuit of universal priors remains a distant goal. Distilling innate priors rooted in foundational common sense concepts like object persistence and spatial reasoning can bridge the gap between natural and cultural knowledge in AI development.
Engineering Encounters with Consciousness
The distinction between treating a system as conscious 'as if' versus factual consciousness hinges on the ability to engineer encounters with putatively conscious entities. By immersing oneself in the same shared world as these artifacts, through interactions and observations, the evaluation of their behavior can lead to discerning consciousness. Establishing public interactions removes subjective biases, enabling a more objective assessment of potential consciousness.
Contextualizing Lessons from Behavioral Experiments with AI
Drawing parallels between behavioral experiments with AI systems and human cognition aids in discerning underlying consciousness. Insights from games environments, virtual worlds, and AI simulations encourage a deeper understanding of artificial and human consciousness through embodied interactions and shared experiences.
Thomas Nagel's 'What is it like to be a bat?'
Thomas Nagel's famous paper 'What is it like to be a bat?' presents the concept that though humans assume creatures like bats have consciousness, the experience of being a bat would be fundamentally different from that of a human. Nagel uses the bat as an example to highlight the idea that there are subjective experiences beyond human comprehension, which hints at a sort of dualistic view of reality.
The Turing Test and Today's Language Models
The Turing test, initially proposed by Alan Turing, involves a human judge interacting with two entities without knowing which is a machine and which is a human. Today's large language models, while easily identifiable as machines due to their explicit disclosure, have advanced conversational abilities that align with the spirit of the Turing test. Despite the machines revealing their artificial nature, they demonstrate human-level language skills.
Idealism vs. Panpsychism
Idealism asserts that the physical world is a construct of mental substance, while panpsychism suggests that physical objects contain an inherent mental aspect. Both philosophies reject dualism by proposing a single substance with mental properties. However, the concepts of consciousness and intelligence remain contentious in these paradigms, requiring in-depth understanding and exploration.
Challenges of Teleology
Professor Shanahan refrains from taking a definitive stance on teleology, acknowledging his limitations on this subject. While pondering questions of cosmic purpose, substance dualism, and monism, Shanahan emphasizes the need for critical examination beyond established philosophical isms. His approach aligns with Wittgenstein's methodology of unraveling philosophical puzzles by grounding concepts in everyday language and thought.
Consciousness, Cognition, and Subjectivity
Through exploring topics ranging from Thomas Nagel's bat to the Turing test, the conversation delves into the complexities of consciousness, cognition, and subjectivity. Professor Shanahan's philosophical insights challenge dualistic inclinations and highlight the necessity of shifting perspectives to demystify philosophical quandaries.
Perspectives on Philosophical Thought
The dialogue navigates various philosophical perspectives, from idealism to panpsychism, and teleological views, underscoring the intricate tapestry of philosophical discourse. As Professor Shanahan delves into the nuances of consciousness and cognition, he advocates for a critical examination of established philosophical paradigms and emphasizes the importance of contextualizing concepts within everyday language and experience.
Murray Shanahan is a professor of Cognitive Robotics at Imperial College London and a senior research scientist at DeepMind. He challenges our assumptions about AI consciousness and urges us to rethink how we talk about machine intelligence.
We explore the dangers of anthropomorphizing AI, the limitations of current language in describing AI capabilities, and the fascinating intersection of philosophy and artificial intelligence.
Show notes and full references: https://docs.google.com/document/d/1ICtBI574W-xGi8Z2ZtUNeKWiOiGZ_DRsp9EnyYAISws/edit?usp=sharing
Prof Murray Shanahan:
https://www.doc.ic.ac.uk/~mpsha/ (look at his selected publications)
https://scholar.google.co.uk/citations?user=00bnGpAAAAAJ&hl=en
https://en.wikipedia.org/wiki/Murray_Shanahan
https://x.com/mpshanahan
Interviewer: Dr. Tim Scarfe
Refs (links in the Google doc linked above):
Role play with large language models
Waluigi effect
"Conscious Exotica" - Paper by Murray Shanahan (2016)
"Simulators" - Article by Janis from LessWrong
"Embodiment and the Inner Life" - Book by Murray Shanahan (2010)
"The Technological Singularity" - Book by Murray Shanahan (2015)
"Simulacra as Conscious Exotica" - Paper by Murray Shanahan (newer paper of the original focussed on LLMs)
A recent paper by Anthropic on using autoencoders to find features in language models (referring to the "Scaling Monosemanticity" paper)
Work by Peter Godfrey-Smith on octopus consciousness
"Metaphors We Live By" - Book by George Lakoff (1980s)
Work by Aaron Sloman on the concept of "space of possible minds" (1984 article mentioned)
Wittgenstein's "Philosophical Investigations" (posthumously published)
Daniel Dennett's work on the "intentional stance"
Alan Turing's original paper on the Turing Test (1950)
Thomas Nagel's paper "What is it like to be a bat?" (1974)
John Searle's Chinese Room Argument (mentioned but not detailed)
Work by Richard Evans on tackling reasoning problems
Claude Shannon's quote on knowledge and control
"Are We Bodies or Souls?" - Book by Richard Swinburne
Reference to work by Ethan Perez and others at Anthropic on potential deceptive behavior in language models
Reference to a paper by Murray Shanahan and Antonia Creswell on the "selection inference framework"
Mention of work by Francois Chollet, particularly the ARC (Abstraction and Reasoning Corpus) challenge
Reference to Elizabeth Spelke's work on core knowledge in infants
Mention of Karl Friston's work on planning as inference (active inference)
The film "Ex Machina" - Murray Shanahan was the scientific advisor
"The Waluigi Effect"
Anthropic's constitutional AI approach
Loom system by Lara Reynolds and Kyle McDonald for visualizing conversation trees
DeepMind's AlphaGo (mentioned multiple times as an example)
Mention of the "Golden Gate Claude" experiment
Reference to an interview Tim Scarfe conducted with University of Toronto students about self-attention controllability theorem
Mention of an interview with Irina Rish
Reference to an interview Tim Scarfe conducted with Daniel Dennett
Reference to an interview with Maria Santa Caterina
Mention of an interview with Philip Goff
Nick Chater and Martin Christianson's book ("The Language Game: How Improvisation Created Language and Changed the World")
Peter Singer's work from 1975 on ascribing moral status to conscious beings
Demis Hassabis' discussion on the "ladder of creativity"
Reference to B.F. Skinner and behaviorism
Remember Everything You Learn from Podcasts
