Reflections on AI

Lex Fridman interview with Yann LeCun

Apr 30, 2023 Source

This interview took place over a year ago, in January 2022. I sought it out because I kept noticing Yann LeCun’s name in exchanges with other prominent figures like Elon Musk, Max Tegmark, and Eliezer Yudkowsky; and LeCun appeared to be, as far as I could tell, the highest-profile person arguing against the viewpoint that continued AI research is extremely dangerous. Even Sam Altman, despite embracing the “Let it out” approach to AI safety, seems significantly more concerned about AI safety than LeCun does.

Related: here’s a transcript of a conversation between LeCun and Yudkowsky.

LeCun spoke at length about self-supervised learning, which is a technique for enabling AI systems to develop generalized knowledge of the world, also known as common sense, which he called the “dark matter” of intelligence.

One casual observation LeCun made that stuck with me was that a human child can see just a few pictures of an elephant, for example in a children’s book, and from just those few pictures develop a fairly accurate internal model of what an elephant is. In contrast, machine learning models (at least as of when the interview took place) require data sets that are orders of magnitude larger. LeCun believes that the key reason for this is that the human child utilizes extensive background knowledge that the model doesn’t have. He considers the acquisition and role of this background knowledge to be poorly understood by the AI research community, hence the “dark matter” analogy.

Fridman and LeCun also talked for a bit about Meta, Facebook, the history and culture of FAIR (at the time, “Facebook AI Research”, but now rebranded to “Fundamental AI Research”), and the global impact of social media as a whole. Fridman observed that perhaps social media has increased division, but that one should look at all of its effects, not just one in isolation. Adding to this, LeCun made an interesting point—that the printing press increased division, in fact in a huge way, as it gave rise to Protestantism. I myself am not especially sympathetic towards Meta, but I found that to be a thought-provoking point.

Towards the final third of the conversation, I became much more engaged as they started discussing some truly fascinating topics. One in particular that I really enjoyed was the mystery of complexity. They were talking about cellular automata and the phenomenon of emergence. Fridman asked LeCun if he understood or could explain emergence; he responded that no, it is a mystery to physicists, chemists, biologists, all scientists.

LeCun noted that there is no rigorous mathematical way of quantifying complexity. And because we cannot measure it, we cannot validate methods for reducing it. Complexity, LeCun said, is in the eye of the beholder. He gave the example of an image of the MNIST digits: if you took this image and applied a fixed random permutation to every pixel, the result would appear chaotic and random. But then you could imagine wearing a special pair of glasses that reversed the permutation, and what had appeared complicated a moment ago would suddenly seem simple.

He pointed out that if we were to meet an alien race who viewed the world differently, as if they were wearing such permutation-reversing glasses, it could be that what we find simple, they find complex, and vice versa.

It struck me that this is potentially a significant opportunity for someone with the right skills and insight to have a huge impact on science in general, and artificial intelligence in particular. The fact that we seem to be seeing emergence in modern AI systems, yet we lack a robust theoretical or mathematical understanding of emergence, feels important.