MIT AGI: Artificial General Intelligence

Course Mission and Philosophical Stance

• Core Objective: The course (6S099) at MIT adopts an engineering perspective to "engineer intelligence," prioritizing the scientific construction of intelligent systems over speculative philosophical debates.
• Methodology: The approach is defined by "mind in hand," grounding the exploration of fundamental intelligence concepts in the immediate creation of systems that function in the present day.
• Contrast in Viewpoints: The curriculum explicitly opposes the "black box reasoning" view that focuses solely on the societal impact of a hypothetical human-level AGI (e.g., robot takeovers or utopia) without understanding the underlying mechanics.
• Current Assessment of Capability: Instructors assert that current methods are far from achieving human-level intelligence, requiring a major leap, breakthrough, or paradigm shift.
• Risk Assessment: While the course emphasizes engineering limits, it acknowledges the risk that exponential technological growth could lead to a "singularity" or sudden breakthrough that changes the landscape overnight.
• Societal Impact Context: The course argues that safety, ethics, and future impact questions cannot be addressed constructively without a deep understanding of the actual methods and limitations of current AI systems.
• Media vs. Engineering: A key educational goal is to avoid "futurism thinking" and hype detached from scientific understanding, contrasting media narratives (e.g., "AI created its own baby") with the engineering reality (e.g., a simple parameter grid search).

Key Challenges and Open Questions

• The Fundamental Question: The central unresolved inquiry for the field is "How hard is it to create a human-level artificial intelligence system?"
• Engineering vs. Biological Networks: Significant gaps exist between human brains (100 billion neurons, 100 trillion synapses, asynchronous, energy-efficient, unknown learning algorithms) and artificial networks (fewer synapses, limited topology, backpropagation-based, energy-intensive, trivial supervised learning).
• Data Dependency: Current deep learning relies heavily on massive amounts of supervised data; the inability to learn efficiently without human annotation is a primary limitation.
• Generalization Issues: Systems struggle with edge cases and transferring knowledge between domains, which is critical for real-world applications like autonomous vehicles operating in unpredictable environments.
• Loss Function Flaws: Reward and loss functions are defined by humans and are often flawed when applied to real-world scenarios lacking a "ground truth" testing set (e.g., prioritizing speed over pedestrian safety).
• End-to-End Learning: A major open challenge is determining if the entire AI stack—from raw sensory data to reasoning, knowledge formation, and physical action—can be learned end-to-end without distinct human-defined stages.

Course Projects

• DreamVision: A competition utilizing Google Deep Dream concepts to generate video visualizations by mixing dream states with reality via neural networks; winners are judged on visual beauty via Mechanical Turk.
• ANGEL (Artificial Neural Generator of Emotion and Language): An A/B testing project where agents generate emotion using a face with 26 muscles controlled by an LSTM network; the goal is to impress a viewer to choose the agent over a human or another agent.
• Ethical Car: A deep reinforcement learning simulation extending the "Deep Traffic" competition by introducing stochastic, unpredictable pedestrians; participants must balance the goal of reaching a destination quickly against a loss function that penalizes harm to pedestrians (a Pareto front trade-off).
• Vote AI: An aggregator platform where users submit links to AI articles and vote on their quality and sentiment (positive/negative) to explore arguments for and against AGI.

Scheduled Speakers and Key Themes

• Josh Tenenbaum (MIT): Will discuss computational cognitive science, focusing on building "intuitive physics" and systems that can learn common sense rapidly (few-shot learning) via probabilistic generative models and symbol processing.
• Ray Kurzweil (Google): Scheduled for Wednesday to discuss the exponential growth of intelligence, the definition of natural intelligence, and the future trajectory toward AGI.
• Lisa Feldman Barrett (Northeastern University): Thursday speaker discussing the theory that emotions are "created" rather than innate, arguing that emotional intelligence is a learned societal norm that machines can be taught via machine learning.
• Nate Derbinski: Will present on cognitive modeling architectures to systematically model the complexity of human cognition.
• Andrej Karpathy (Tesla): Will address the limitations and possibilities of deep learning, specifically the power of representational learning (hierarchical feature extraction) and the contrast between biological and artificial neural network efficiency.
• Stephen Wolfram: Monday evening session on knowledge-based programming (Wolfram Alpha), the use of cellular automata to show how simple rules generate complex patterns, and his experience analyzing alien language for the film Arrival.
• Richard Moyes (Article 36): Will cover the legal and policy aspects of banning autonomous weapon systems and the safety hazards of machines making decisions to kill.
• Mark Riber (Boston Dynamics): CEO of Boston Dynamics will discuss the engineering challenges of humanoid robotics and the societal perception of embodied intelligence.
• Ilya Sutskever (OpenAI): Co-founder of OpenAI will discuss deep reinforcement learning in game playing (e.g., AlphaGo Zero self-play) and the feasibility of learning the entire AI stack from raw sensory data to action.

Future Course Direction

• Expansion: The course will continue throughout 2018 with additional speakers and projects beyond the initial two-week intensive.
• Bias and Ethics: Upcoming content will focus on creating AI systems that do not discriminate or inherit human social biases.
• Creativity: Further exploration of using machine learning for art and music generation is planned.
• Neuroscience Integration: A future gap to be filled is the inclusion of computational neuroscience and brain simulation perspectives.
• Natural Language Processing (NLP): A dedicated speaker and project on NLP and the Turing Test are scheduled for March.
• Community Engagement: Registered students are invited to join the "deep-MIT" Slack channel, submit five links, vote on ten articles, and participate in the three main competitions.