Shifting the AI Paradigm: Beyond the Limits of Learning

Machine Learning is failing us. But our ambitions do not. Here’s how to revolutionize AI.

Imagine AI that mimics human intellect. That’s the future Yann LeCun envisions. His latest lecture at Harvard University decodes the journey towards autonomous machine intelligence—an AI that not only learns but also reasons, plans, and operates with common sense, while being safely under our control.

In this lecture, LeCun walks us through the fact that "machine learning sucks". In other words, the current models make so many blatant mistakes we cannot really this "intelligence", artificial or not.

Intelligent AI Assistants

If we really want to build "intelligent" AI assistant, with human-like capabilities, we need novel ways to approach deep learning algorithms. He particularly mentions the following features:

• AI with Intuition: The goal is to build systems that learn about the world just as humans and animals do. Through sensory inputs like vision, these systems would develop an intuitive understanding of physics and reality.
• Persistent Memory and Planning: An advanced machine intelligence must possess a persistent memory, similar to how we remember past experiences, and the ability to plan actions based on objectives.
• Controllability and Safety: Crucially, these systems must be inherently safe and controllable. The safety mechanisms must be built into their architecture, not tacked on as an afterthought.

LeCun criticizes current large language models (LLMs) for their lack of understanding of the world and their tendency to make silly mistakes. These models lack common sense and an accurate world model, which are crucial for achieving human-level AI.

He offers an alternative: Objective-Driven AI Systems. These systems use self-supervised learning and are designed with capabilities to learn, reason, and plan, inspired by the hierarchical planning seen in human cognition.

The idea is to shift from generative models to joint embedding architectures, which represent a fusion of the physical and digital worlds that AI can navigate and understand.

Vision for Universal Virtual Assistants

LeCun predicts that in the future, all of our interactions with the digital world will be mediated by AI assistants. They will not only act as repositories of human knowledge and culture but also serve as a shared infrastructure akin to today’s internet. And to ensure that these platforms remain inclusive and not controlled by a handful, LeCun argues for the necessity of open-source AI platforms (see my previous article on this topic).

Questions To Ponder

This brilliant demonstration and overview of the future comes with a lot of unanswered questions. However, since technology is evolving fast, the objective-driven AIs that LeCun describes may become a reality sooner that we may expect. So we need to get prepared, have discussions within our communities and at large, and influence our organisations to ensure they're ready and make their employees ready for the next waves of innovations. Among those questions, I have three in mind that are top of my lists:

• How can we foster the development of AI that is beneficial and safe for society?
• What steps can we take to ensure the democratization of AI technology?
• In what ways can we prepare current and future generations for a world intertwined with advanced AI systems?

You will find Yann LeCun's lecture here. I really encourage you to take an hour of your time to listen to it. Yes, some concepts and maths may fly high above your head, and you may not have all the maths foundation to get all of his presentation, but you'll get the crux of it, that could be summarized in the following way:

• Current models are highly limited and cannot really get better.
• We need new models that are steerable and controllable, and learn how the world works (not to generate better tokens).
• There are many obstacles along the way and research is much needed.

"There is no question that machine will eventually surpass human intelligence in all domains where humans are intelligent (and more)". But, you can still sleep at night, we are decades away from this!

This article has been crafted with the precious help of ChatGPT 4.