257: AI: How Far We Haven’t Actually Come

Episode Summary

Executive Summary: Melanie Mitchell discusses complexity and AI, arguing that today’s neural networks are powerful but still far from human-like intelligence because they lack common sense, robust understanding, and rich feedback/embodiment. The conversation covers complex systems, network science, explainable AI, adversarial vulnerability, DARPA’s role, military risks, and likely near-term progress in cybersecurity and unsupervised learning.

Main Topics: Complex systems and network science: Mitchell explains complexity as the study of systems whose behavior emerges from many interacting parts, using networks to compare brains, the internet, social systems, and economies. Limits of modern AI and neural networks: She contrasts current deep learning with human intelligence, emphasizing that today’s systems are narrow, brittle, and missing key ingredients like feedback, common sense, and context. Explainable AI vs performance: The discussion explores the trade-off between highly effective black-box models and systems that can justify decisions in a human-meaningful way. Bias, hallucination, and human-AI differences: Mitchell argues humans and AI both have biases, but AI systems can hallucinate features from priors and may need some biases to achieve general intelligence. Research vs applied AI: She distinguishes open-ended basic research, where being wrong is normal, from applied work, which has clearer goals and immediate utility. Safety, adversarial attacks, and military use: The conversation addresses fragility in deployed AI, adversarial examples, and concerns about autonomous weapons and the security implications of AI. Future directions: common sense and unsupervised learning: Mitchell highlights common sense, embodied cognition, cybersecurity, and unsupervised learning as major next challenges and likely areas of progress.

Key Arguments: Complex systems are defined by emergent behavior: the whole is more than the sum of its parts, as seen in brains, ant colonies, economies, and the internet. Network structure matters more than raw size; interconnections and dynamics explain intelligence and resilience better than simply adding more neurons or layers. Modern neural networks are inspired by brains but differ fundamentally, especially because biological cognition relies heavily on feedback, prior knowledge, and chemical/embodied processes. Deep learning excels at narrow tasks but lacks general intelligence, transferable common sense, and robust understanding across contexts. Human-like biases and expectations may be inseparable from general intelligence; perfect objectivity may not be compatible with flexible, general cognition. AI systems can appear competent yet remain brittle and vulnerable to adversarial manipulation, which is dangerous for high-stakes uses like security and face recognition. Explainability is partly social and contextual: a useful explanation depends on the audience’s knowledge and the situation, not just on exposing model parameters. Basic research is exploratory and frequently wrong by design, whereas applied AI typically starts with a defined problem and clearer success criteria. Future progress is likely to come from better unsupervised learning and stronger AI security rather than just scaling model size. Embodied cognition suggests that intelligence may require a body and interactions with the physical world, not just disembodied computation.

Data Points: Number of books by Melanie Mitchell: 6 books, with a 7th coming out in September 2019 - Her author background and upcoming AI book PhD students in her research lab: About 6 - Team size described during discussion of her lab Brain neuron count: Hundreds of billions / about 100 billion neurons - Used to compare human brains with neural networks Feedback connections in visual system: About 10 times as many feedback connections as feed-forward connections - Evidence for why biological perception differs from feed-forward neural nets Complexity Explorer course availability: Free online course with no prerequisites - Her Intro to Complexity course on Santa Fe Institute platform Year of Complexity book award: 2010 - Complexity: A Guided Tour won the Phi Beta Kappa Science Book Award in 2010 Year of book recognition by Amazon: 2009 - Named one of Amazon’s 10 best science books of 2009 Shown as data benchmark performance: Machines can surpass humans on specific benchmark datasets - Used to caution that benchmark success is not general intelligence

Pivotal Quotes: "Complexity is a very broad area that deals with what are called complex systems, which are systems that you can say they're sort of more than the sum of their parts." — Melanie Mitchell: Defining complexity early in the interview "I think that there's going to be a trade-off between general intelligence and being able to be kind of unbiased in this sense." — Melanie Mitchell: On cognition, bias, and the limits of super-intelligent AI "The real danger is it's actually too stupid and it's already taken over the world." — Pedro Domingos (quoted by Melanie Mitchell): Referenced in the discussion of AI risks and overreliance on fragile systems

Implications: Listeners should expect progress in narrow AI, but not assume human-like intelligence is imminent. The biggest near-term risks are fragility, misuse, and poor security. The biggest opportunities are common sense, embodied AI, unsupervised learning, and explainable, robust systems.