Max Tegmark on AI, Physics, and the Quest for Intelligible Intelligence

This episode features Max Tegmark, a physicist and AI researcher, discussing the critical intersection of artificial intelligence and physics, particularly focusing on the breakthroughs in AI and the urgent need for its deeper understanding. Tegmark highlights the shift in his own research towards machine learning and the establishment of the AI Institute for Artificial Intelligence and Fundamental Interactions, an NSF-funded center dedicated to leveraging AI for physics and vice-versa. He emphasizes that while AI can significantly advance physics, the physics mindset—rooted in fundamental understanding and provable principles—is crucial for developing trustworthy and safe AI systems, contrasting it with the current engineering-driven approach that often results in powerful but opaque 'black box' models.\n\nTegmark reviews recent AI achievements like AlphaFold 2, GPT-3, and MuZero, acknowledging their power but underscoring the lack of full comprehension of their internal workings. He argues that the real dangers of AI stem not from machines turning 'evil' as depicted in Hollywood, but from human over-trust in systems we don't adequately understand, citing examples like the Boeing 737 MAX and Knight Capital's trading system. This leads to a call for greater humility in science and AI development, advocating for 'intelligible intelligence' where systems are understood at a deep, fundamental level, much like Newton's laws allow for reliable rocket science. He posits that the power of neural networks comes from their differentiability, enabling optimization, rather than any inherent inscrutability.\n\nA key distinction Tegmark makes is between two paths to Artificial General Intelligence (AGI). The first, which he finds alarming, involves simply scaling up neural networks with more data and hardware, leading to ever more powerful but still opaque systems. He warns of a significant risk of unintended catastrophic consequences or misuse by malicious actors if this path is pursued without sufficient understanding. The second, more encouraging path, is his 'intelligible intelligence' approach, exemplified by his 'AI Feynman' project. This method involves using neural networks for initial approximation and intuition, then employing additional AI techniques to dissect and transform these black boxes into understandable, symbolically verifiable models, akin to how Galileo or Newton distilled observations into fundamental equations.\n\nUltimately, Tegmark advocates for combining the strengths of 'good old-fashioned AI' (logic-based, symbolic reasoning) with modern neural networks to achieve AGI that is not only intelligent but also transparent and provably reliable. He expresses optimism that, just as humanity uncovered the laws governing the physical world, we can develop a deep, scientific understanding of intelligence itself. This approach is not just about scientific curiosity but is presented as a critical safety measure, ensuring that advanced AI serves humanity's intended goals rather than leading to unforeseen and potentially disastrous outcomes, thereby defining a safer trajectory for the interplay of technology and human civilization." "concepts": [ "Artificial General Intelligence (AGI)

I believe that already the algorithms that drive our interaction on social media have an intelligence and power that far outstrip the intelligence and power of any one human being.

My vision is that in the future all machine learning systems that actually have impact on on people's lives will be understood at a really really deep level.

The actual bad things that have happened with automation have not been machines turning evil they've been caused by over trust in things we didn't understand as well as we thought we did.

Step one if you want to be a scientist is don't ever fool yourself into thinking you understand things when you actually don't.

My mit research is entirely focused on demystifying this black box intelligible intelligence is my slogan.

I think the real power of neural networks comes not from inscrutability but from differentiability.

If we actually succeed as a species to build artificial general intelligence then we still have no clue how it works I think at least 50 chance we're going to be extinct before too long.

It's much easier to verify that a proof is correct than to come up with a proof in the first place.