Objective-Driven AI // Cognitive architectures
Current AI is just a statistics, popular tech like LLMs are useful but very limited in terms of logical reasoning.
Here are the key points from the lecture:
- Professor Yan LeCun gave the 16th annual Dean W Lial endowed lecture at the University of Washington. The lecture series honors Dean Lial, a longtime professor and leader in electrical engineering and computer science at UW.
- LeCun discussed the limitations of current AI systems like large language models. While impressive in certain narrow tasks, they lack common sense, an understanding of physics, and the ability to plan and reason like humans and animals can.
- He proposed an “objective-driven AI” approach, with modules for perception, world models, planning, etc. Key ideas are training world models with video prediction, and having the system optimize actions to satisfy objectives. This is different from today’s “auto-regressive” LLMs.
- LeCun argued for the need for open source AI platforms, to enable customization for different cultures and values. He sees smart assistants as future “common infrastructure”, like Linux is for computing today.
- There are many open challenges to reach human-level AI, including hierarchical planning, handling uncertainty, and learning accurate world models. LeCun believes progress will take longer than many think, but machines will eventually surpass human intelligence in all domains.
In summary, LeCun critiqued current AI’s limitations, proposed a cognitive architecture approach to work towards more capable systems, and argued for open and diverse AI platforms to serve humanity responsibly. But he cautioned that human-level AI is still far off, requiring innovations in learning, reasoning, and planning.
