Ex Machina's Scientific Advisor - Murray Shanahan

Academic Trajectory and the "Frame Problem"

• PhD Thesis Focus (1980s): Investigated logic programming (Prolog) and query speed-up techniques by caching relationships between established facts and theorems to avoid redundant computations.
• The Frame Problem: A recurring theme throughout the speaker's career, defined as the challenge of determining relevance versus irrelevance in cognitive processes to avoid being overwhelmed by trivial non-effects of actions.
• Career Pivot (c. 2000): Abandoned classical "Good Old-Fashioned AI" (symbolic logic) due to disillusionment with its inability to achieve Artificial General Intelligence (AGI).
• Neuroscience Phase: Shifted focus to computational neuroscience to study the human brain as a model for intelligence, consciousness, and attention mechanisms.
• Return to AI: Recently returned to AI research via machine learning, specifically seeking to rehabilitate concepts from classical symbolic AI to address the frame problem in modern systems.
• Current Relevance: The speaker notes that the frame problem persists in contemporary machine learning, specifically regarding the ability of systems to filter irrelevant data (e.g., distinguishing relevant game mechanics in Atari games like Space Invaders from irrelevant visual details like color or shape).

Industry Context: Kasparov and AlphaGo

• Kasparov's Chess Analysis: Gary Kasparov highlighted that modern grandmasters rely heavily on computer analysis, reshaping their cognitive approaches to the game; any modern iPhone chess player is likely superior to Deep Blue (1997).
• Human-Machine Partnership: Top Go players, such as Lee Sedol and Ke Jie, have adopted a positive view of AlphaGo, using it to explore strategic territories previously considered impossible, effectively creating new game theory.
• Move 37 Controversy: In the AlphaGo vs. Lee Sedol match, Move 37 was initially deemed a mistake by nine-dan masters but was later recognized as a revolutionary tactical innovation.
• Moravec's Law: Cited by Kasparov, this law suggests that high-level cognitive tasks are computationally hard for AI, while low-level sensorimotor skills are easy, leading to the concept of "Mind Children" (artifacts with independent lives as envisioned in Hans Moravec's Mind Children).

Ex Machina Consultation and Production Details

• Consultation Role: Served as a scientific advisor to director Alex Garland for Ex Machina, primarily validating the script's philosophical alignment with the speaker's work, particularly the concept of consciousness arising from interaction rather than explicit testing.
• The "Garland Test": A specific scene where the character Nathan states the Turing test is obsolete, arguing the goal is to determine if the AI is conscious despite being known as a robot. The speaker praised this line for its philosophical accuracy.
• Embodiment Philosophy: The speaker emphasized that human intelligence is rooted in physical embodiment (manipulating 3D space, locomotion), suggesting that while films often depict human-like robots, real AGI may not require a humanoid form.
• Deleted Ending Scene: The original script included a scene visualizing Ava's alien perception (waveforms, facial recognition vectors) as she escapes. This was cut during editing, as the team felt it made her too explicitly alien; its absence preserves ambiguity regarding her consciousness.
• The Smile Ambiguity: The speaker advised against Ava's final private smile, arguing it was too human, but the director retained it to suggest Ava possesses human-like consciousness.
• The Easter Egg: The speaker inserted a Python code snippet into a prop computer screen in the film. When run, it generates a message containing the ISBN of his book The Inner Life of Artificial Intelligence.
  • Technical Note: The speaker later admitted the code contained an inefficient loop terminating condition, resulting in a "bug" that fans have analyzed.

Future of AI and Technical Challenges

• Human-Level AGI Timeline: While computing power equivalent to the human brain may be available by 2020–2022, the speaker states there are "unknown conceptual breakthroughs" required to utilize that power for intelligence.
• Specialized vs. General AI: Current AI excels at specialized tasks (image recognition, speech-to-text), but genuine semantic understanding remains distant.
• Deep Reinforcement Learning (DRL): The speaker identifies DRL (pioneered by DeepMind's DQN) as a significant step toward general intelligence but notes it is extremely data-inefficient compared to human learning speeds.
• Synthesis of Approaches: Current research aims to integrate symbolic AI concepts (logic, abstraction) into deep learning systems to improve efficiency and relevance filtering.
• Asimov's Laws: The speaker dismisses Asimov's laws as irrelevant to current robotics, as they are narrative devices rather than technical constraints; machines lack the comprehension to interpret or enforce such moral rules.
• Hollywood vs. Reality: AI in film often favors dystopian or humanoid villains for dramatic effect, whereas real-world AI advancement (e.g., autonomous vehicles) may not be embodied in human forms.

Other Creative Collaborations

• Play Elegy: Collaborated with Nick Payne on a play about a couple where one partner suffers a dementia-like condition cured by memory loss, exploring the neuroscience of memory and identity.
• Random International: Consulted on art installations such as Rain Room (a room of rain that avoids the visitor via motion sensors) and 15 Points (a kinetic sculpture using point-light displays to simulate human motion).