Taming AI - Matt Jones

The lecture provides a thoughtful and accessible overview of AI safety and governance, using vivid metaphors and real-world examples to illustrate key concepts. Jones’s central argument—that AI should be treated as a power to be tamed rather than a presence to be anthropomorphized—is compelling and well-supported by historical analogies such as fire and the automobile. The distinction between ‘fireplace AI’ and ’elephant AI’ effectively captures the spectrum of controllability in AI systems, from highly regulated medical devices to unpredictable large language models. The discussion of the alignment problem is particularly strong, with concrete examples (Tetris, racing game) that clarify the gap between literal instructions and human intent. The lecture also addresses important issues like legibility, explainability, and the dangers of automation bias, citing real-world failures such as COMPAS and the Boeing 737 MAX. However, the lecture is primarily an opinion piece rather than a rigorous scientific review; it lacks empirical data and detailed technical depth. Some claims are simplified for a general audience, and the proposed solutions, such as ‘constitutional AI’ and ‘human in the loop,’ are presented without critical examination of their limitations. The title accurately reflects the content, and the lecture is well-structured, but it could benefit from more concrete policy recommendations and a deeper exploration of trade-offs. Overall, it is a valuable introduction to AI safety for a non-specialist audience, but it does not break new ground for experts.

The title 'Taming AI' accurately reflects the lecture's focus on methods to control and govern AI systems, using metaphors like fire and elephants.

8/10

The lecture is delivered by a computer scientist with extensive experience in human-centred design, and it draws on established concepts (alignment, explainability, regulation) and historical analogies. However, it is an opinion-based lecture without peer-reviewed citations or empirical data, and some claims are simplified for a general audience.

• 00:00// Introduction: Welcoming Professor Matt Jones
• 00:32// Jurassic Park: When AI Breaks Free
• 01:21// How to Train Your Dragon: A Gentler Approach to Taming
• 03:06// Why Anthropomorphizing AI Fails (The Tiger Who Came to Tea)
• 04:09// Why a Human-AI Brain Merger Won't Work
• 05:40// AI Is a Power, Not a Presence
• 06:08// Taming Fire: Lessons for AI Safety
• 08:11// The Car: Skills, Roads and Licensing
• 12:42// Fireplace AI: Trusted Medical Diagnosis Systems
• 14:33// Elephant AI: Containing Unpredictable Power
• 17:21// The AI Alignment Problem
• 19:48// Why AI Needs Regulation and Governance
• 23:01// Black Box AI: The COMPAS and Boeing 737 Disasters
• 26:02// Legibility: How to Read an AI System
• 30:45// Explainable AI: The Wolves, Dogs and Snow Problem
• 35:16// Shaping AI Behaviour and the Sycophancy Trap
• 39:28// AI Guardrails: Lessons from Chernobyl
• 43:49// Asimov's Laws and Constitutional AI
• 47:14// Human in the Loop and Automation Bias
• 49:27// Conclusion: Participate in AI's Future

• Gresham College — Institution hosting the lecture series.
• Support Gresham College — Mentioned in the description to support the college.
• Lecture page: AI Taming — Official page for this lecture.
• Q&A Session — Follow-up Q&A session for this lecture.

• AI Alignment — Supports the discussion on alignment problem.
• Explainable AI — Relates to the lecture's emphasis on legibility and explainability.

• COMPAS Recidivism Algorithm