Human-Level AI // hype again or something real?
There is no human-level AI, no one knows what is that “human-level”?)
Do we need human-level AI for Assistant? // does human-level AI need us for any reason?)
Yann LeCun, Chief AI Scientist at Meta and Director of AI Research at Facebook, NYU professor, and recipient of the ACM Turing Award, delivered a keynote on the journey towards human-level AI. Here’s a summary of his talk (by llama3.1):
Current AI systems are not capable of understanding the world, remembering things, or having common sense. They are narrow and lack the ability to reason and plan like humans.
To achieve human-level AI, we need systems that can:
1. Model the world
2. Have persistent memory
3. Plan complex action sequences
4. Are controllable and safe
LeCun proposes an architecture called “Objective-Driven AI” that uses optimization to make predictions and decisions. This approach is different from traditional feedforward neural networks.
Key components of Objective-Driven AI:
1. Perception system
2. World model (predicts the effect of actions on the world)
3. Objective functions (measure the extent to which goals are achieved)
4. Actor (optimizes action sequences to achieve goals)
Hierarchical planning is essential for intelligent behavior, but it’s a major challenge for AI systems. LeCun argues that we don’t know how to do hierarchical planning with AI, and it’s still an unsolved problem.
Learning world models is also a significant challenge. LeCun proposes using joint embedding predictive architectures, which learn an abstract representation of the world and predict in that space, rather than trying to predict pixels.
Training these systems requires new methods, such as maximizing information content or using distillation-style methods.
If we succeed in developing human-level AI, it will be an infrastructure, like the internet, and must be open source to ensure cultural diversity and preservation of democracy.
Timeline: LeCun estimates it may take years or decades to reach human-level AI, and it’s likely harder than we think.
Risk: LeCun acknowledges there is risk associated with developing superintelligence but believes that objective-driven AI, where machines are given goals and work to achieve them, will allow for control and mitigate the risk of domination or takeover.
