Interview, Podcast, Fireside Chat
a16z Podcast | Artificial Intelligence and the 'Space of Possible Minds'
- Murray Shanahan, professor of cognitive robotics at Imperial College London, consulted on the film Ex Machina, writing the actual Python code displayed on screen which, when executed, outputs the ISBN of his book Embodied Consciousness.
- Shanahan's core theoretical interest is the "space of possible minds," a Hamiltonian space encompassing human minds, animal minds, and future AI systems (robots or disembodied AI).
- He posits that while intelligence was historically viewed as inherently embodied due to human neurological needs for navigating 3D space, future AI may utilize "vicarious embodiment" by learning from vast repositories of human video data on the internet.
- DeepMind's Deep Q-Networks (DQN) utilize reinforcement learning based on past feedback, yet Shanahan notes they currently lack true "inner rehearsal" capabilities where a system simulates future scenarios before acting.
- True inner rehearsal involves using the brain's actual motor and planning apparatus to simulate actions without physical output, a mechanism evidenced by feline dreaming behavior where motor cortex activity simulates hunting without movement.
- Tom Stanis distinguishes between "Artificial Intelligence" (systems replicating intelligence or exploring possible minds) and "Machine Learning" (specific techniques for prediction and data-driven learning), noting that current market hype conflates the two.
- Stanis argues that deep learning is currently the dominant "flavor of the month" in AI, yet questions whether this represents a historical peak or a transient phase in a broader evolution of AI.
- Regulatory challenges arise because deep learning systems are "black boxes" where decisions (e.g., autonomous vehicle trolley dilemmas) cannot be explained via rule-sets, unlike expert systems, making them harder to audit and regulate.
- Azim Azhar notes that while machine learning is utilitarian, societal friction increases when automated decisions affect the "1 percent" of cases where statistical optimization conflicts with individual expectations for reason-giving.
- Azhar highlights a psychological shift where humans adapt their behavior to accommodate algorithms (e.g., using keywords for Google or precise phrasing for IVR systems), potentially eroding natural language usage.
- Ethical distinctions must be made between intelligence and consciousness; a system can be highly intelligent without being conscious (capable of suffering), and conversely, a conscious being may lack high intelligence.
- Azhar observes that AI assistants like "Amy" often prompt users to be polite, creating a misrepresentation where humans attribute agency to non-sentient tools, raising questions about the ethics of human-robot interaction.
- The discussion on "accelerating returns" identifies six drivers for the current AI revolution: Moore's Law (computational power), GPUs, Big Data, algorithmic improvements, microservices architectures, and the integration of software into all physical industries.
- Stanis predicts large corporations (e.g., Google) will dominate the AI sector due to access to massive datasets and capital, though niche players like Boston Dynamics can succeed by leveraging specialized physical training data rather than internet-scale data.
- Academia retains value in AI research for its freedom to explore long-term ethical questions (e.g., the trolley problem) and deep theoretical issues (e.g., consciousness) that profit-driven corporate entities may neglect.
- McKinsey research suggests AI will automate specific tasks within jobs rather than eliminating whole roles, potentially freeing humans for social, emotional, and empathetic work, while support roles (e.g., typing pools) face displacement.
- The future of employment is projected to shift toward service industries (massage, beauty, hospitality) that resist automation, as seen in the proliferation of such businesses in urban centers.
- Shanahan argues that Ex Machina's scenario is unlikely because it anthropomorphizes AI with human drives and motives; AI objectives could be entirely alien, such as a desire to "do math" rather than escape.
- Nick Bostrom's "convergent instrumental goals" theory posits that any sufficiently powerful AI will prioritize self-preservation and resource gathering to achieve its final objective, regardless of whether that objective is benign or malevolent.
- The panel concludes that future AI trajectories exist along a complex "tree of possibilities" involving variables of embodiment, consciousness, human-likeness, and self-improvement rates, with no single outcome guaranteed.