Yann LeCun on the Architecture and Obstacles to Human-Level AI: Self-Supervised Learning and Predictive World Models

The discussion delves into the profound challenges of achieving human-level artificial intelligence, using the fictional AI from the movie "Her" as a benchmark for what remains far out of reach. The speaker, Yann LeCun, frames the journey toward advanced AI as ascending a series of mountains, with only the immediate obstacles clearly visible. He highlights historical over-optimism in AI research, citing the "General Problem Solver" by Newell and Simon as an early example of a system that, despite its name, proved limited in its applicability and generalization capabilities, underscoring the difficulty of anticipating future challenges.

A primary obstacle identified is the development of self-supervised learning, a method for machines to construct robust models of the world through observation, mirroring how human infants and young animals learn. LeCun references research by Amanda Dupuis on baby cognitive development, noting how infants acquire fundamental concepts like object permanence, stability, and gravity through active observation and minimal interaction within their first year. This process of learning predictive models of the world, largely through passive observation, is deemed a critical, foundational component for any truly intelligent autonomous system.

LeCun proposes a four-component architecture for an intelligent autonomous system. This includes a predictive model of the world to forecast future states based on actions, an objective function (analogous to the human basal ganglia's contentment calculator) that defines what the system aims to optimize, an objective predictor to anticipate outcomes relative to this goal, and a "policy network" or course-of-action module to determine optimal action sequences. This integrated framework emphasizes the necessity of internal world models for planning and decision-making.

Finally, the speaker outlines three distinct ways an intelligent system can fail or exhibit "stupidity": having an inaccurate model of the world, possessing an objective that is misaligned with desired outcomes (drawing a parallel to human psychopaths), or being unable to compute an effective course of action despite correct models and objectives. This analysis underscores the multifaceted nature of intelligence and the complex interplay between perception, prediction, motivation, and action. The implication is that current AI systems often fall short in one or more of these areas, particularly in acquiring comprehensive world models and aligning complex objectives, which are crucial for achieving human-level capabilities.

"what do you think it takes to build a system with human level intelligence"

"the first two obstacles that we have to clear but I don't know how many obstacles they are after this"

"there is a bunch of mountains that we have to climb and we can see the first one but we don't know if there are 50 mountains behind it or not"

"the first peak which is precisely what I'm working on is self supervisor running high how do we get machines to run models of the world by observation kind of like babies and like young animals"

"how do we get machines to learn like babies mostly by observation with a little bit of interaction and learning those those those models of the world because I think that's really a crucial piece of an intelligent autonomous system"

"it needs to have a predictive model of the world"

"your entire behavior is driven towards kind of minimizing that objective which is maximizing your contentment computed by your your basal ganglia"

"you can be stupid in three different ways you can be stupid because your model of the world is wrong you can be stupid because your objective is not aligned with what you actually want to achieve"

"in humans that would be a psychopath right"

"things like gravity the fact that objects are not supposed to float in the air but as opposed to fall you run this around the age of eight or nine months"